Method for processing multi-dimensional supervision data through adaptive cross-modal attention mechanism

Abstract

The invention discloses a method for processing multi-dimensional supervision data through an adaptive cross-modal attention mechanism. Comprising the following steps: S1, acquiring first data representing the uncertainty of a project macroscopic process; s2, acquiring second data representing a project micrological relationship; s3, based on the first data and the second data, processing through a preset fusion model to generate an indication signal representing the attribution probability of the risk source; s4, generating a scheduling strategy containing a control instruction based on the indication signal; s5, executing a scheduling strategy to perform adaptive adjustment on project resources; according to the method, the risk entropy flow field representing the uncertainty of the macroscopic process and the engineering knowledge graph representing the microscopic logic relation are deeply fused, and the high-confidence accurate diagnosis of the risk root is realized by using the bidirectional causal reasoning model; a prospective closed-loop control system is driven, and predictive interception and active intervention of risks are realized by adaptively generating a disposal plan and optimally scheduling resources.

Description

Technical Field

[0001] This invention belongs to the field of engineering supervision technology, and more specifically, it relates to an adaptive cross-modal attention mechanism for processing multi-dimensional supervision data. Background Technology

[0002] In the fields of modern industrial automation and large-scale systems engineering management, digital transformation is driving the evolution of decision-making models towards data-driven intelligence. Complex systems, such as large-scale manufacturing processes, transportation networks, or energy systems, typically generate massive amounts of multimodal monitoring data. One type is quantitative time-series data characterizing the macroscopic operating state of the system, and the other is unstructured text data containing domain expert knowledge and logical rules. How to effectively integrate these two heterogeneous modalities of data to gain insights into the deep-seated causes of system anomalies and, based on this, implement precise and forward-looking closed-loop control is a key technological challenge in improving the intelligence and autonomy of complex systems.

[0003] In the specific application field of engineering supervision, existing technologies have the following main areas for improvement in utilizing data to enhance risk management capabilities:

[0004] The "semantic gap" between macro-level status data and micro-level domain knowledge: Existing project management systems can effectively collect and analyze macro-level quantitative data such as construction progress to identify anomalies like "which stage has experienced delays." Meanwhile, a wealth of micro-level knowledge bases exist, including numerous construction specifications and supervision logs. However, these two data systems are often independent. Current technology lacks an effective means to automatically and accurately link a quantified, semantically ambiguous macro-level anomaly signal (e.g., increased schedule deviation) to the specific micro-level technical cause (e.g., non-compliance with a particular specification). This semantic gap means that risk warnings often remain at the level of "knowing what happened, but not why," generating alerts that highlight risks but lacking sufficient guidance for subsequent precise handling.

[0005] Limitations of "Passive Response" and "Suboptimal Decision-Making" in Risk Response Measures: After a risk is identified, the existing system's response measures are usually triggered passively. The handling process often relies on manual judgment or personnel scheduling based on static, fixed rules, such as assigning the nearest or available supervisor. This model has two limitations: First, it is a "post-event response" rather than "pre-event intervention," making it difficult to effectively intercept risks before they spread and escalate. Second, when allocating resources, it rarely considers multiple dimensions such as the urgency of the risk, the specific professional skills required for the handling plan, and the potential value of proactive intervention. This makes scheduling decisions often localized or suboptimal, making it difficult to achieve the most efficient allocation of overall project resources.

[0006] To address these challenges, it is necessary to establish a fusion bridge that can bridge the semantic gap between macroscopic states and microscopic knowledge, and based on this, construct an intelligent closed loop from accurate diagnosis to proactive action. Summary of the Invention

[0007] To address the problems existing in the prior art, the purpose of this invention is to provide an adaptive cross-modal attention mechanism for processing multi-dimensional supervision data, which deeply couples macro-level state quantification indicators with micro-level domain knowledge models. Through an original knowledge graph activation and transmission model under the risk entropy flow field, it realizes the automatic and accurate tracing of semantically meaningless macro-risk signals to specific micro-level technical roots, and based on this, triggers a dual linkage control of countermeasure generation and personnel pre-set, thereby constructing an intelligent engineering management system with proactive risk immunity and self-healing capabilities.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data, comprising the following steps:

[0009] S1: Obtain the first data characterizing the uncertainty of the project's macro-process. The first data includes the actual completion time series of each construction task.

[0010] S2: Obtain the second data representing the micro-logical relationships of the project. The second data includes an engineering knowledge graph of engineering entities and the relationships between entities.

[0011] S3: Based on the first and second data, the data is processed through a preset fusion model to generate an indication signal that represents the probability of risk root cause attribution.

[0012] S4: Generate a scheduling strategy containing control commands based on the indication signal;

[0013] S5: Execute scheduling strategies to adaptively adjust project resources.

[0014] Furthermore, in S1, based on the actual completion time series of each construction task, the construction cycle entropy, which characterizes the uncertainty of the project's macro process, is calculated using the information entropy model; and based on the preset task dependencies, a risk entropy flow field describing the intensity of risk transmission between tasks is constructed.

[0015] Furthermore, the second data in S2 automatically extracts engineering entities and relationships between entities from unstructured supervision documents using natural language processing technology to construct and update the engineering knowledge graph; and labels each node in the knowledge graph with its preset mapping position in the project time and space.

[0016] Furthermore, the steps for calculating and obtaining the construction cycle entropy and risk entropy flow field are as follows:

[0017] Construction cycle entropy: Based on the deviation between the actual completion time and the planned completion time of the task, and combined with the task resource constraint degree that characterizes the scarcity of task resources, the entropy of resource-constrained tasks is calculated.

[0018] The task resource constraint is calculated as follows: obtain the concurrent demand and total availability of each type of resource associated with the task, calculate the resource saturation, and perform a weighted sum of the saturation of all associated resources. Finally, the result is obtained by mapping through a preset normalization function.

[0019] Constructing a risk entropy flow field: Based on the real-time construction cycle entropy of upstream tasks, dynamically adjust the risk transmission coefficient connecting upstream and downstream tasks;

[0020] The dynamic adjustment steps are as follows: take the construction cycle entropy of the upstream task as input, calculate the dynamic enhancement factor through the exponential function, and multiply the dynamic enhancement factor by the basic transmission weight to obtain the updated risk transmission coefficient.

[0021] Furthermore, the preset fusion model in S3 is a knowledge graph activation and transmission model under the risk entropy flow field; the processing flow of the knowledge graph activation and transmission model is as follows:

[0022] When the local construction cycle entropy in the risk entropy flow field exceeds the preset activation threshold, knowledge graph nodes that have a preset mapping relationship with the local area are activated; and according to the risk transmission path defined in the risk entropy flow field, activation energy is transmitted in the knowledge graph to calculate the attribution probability of each knowledge graph node as the root cause of risk, thereby generating an indication signal.

[0023] Furthermore, when the local construction cycle entropy in the risk entropy flow field exceeds the threshold, the associated knowledge graph nodes are activated and energy transfer is performed to calculate the attribution probability.

[0024] During energy transfer, the forward propagation process and the backward reasoning process are executed in parallel;

[0025] The forward propagation process is as follows: based on the risk transmission coefficient defined in the risk entropy flow field, the forward propagation confidence of each node is iteratively updated in the knowledge graph;

[0026] The backward reasoning process is as follows: Based on the pre-defined logical relationship type weights in the knowledge graph, the backward reasoning confidence of each node in the knowledge graph is iteratively updated;

[0027] The attribution probability is calculated as follows: For each knowledge graph node, the final forward propagation confidence and the final backward reasoning confidence of the knowledge graph node are multiplied and fused to generate the attribution score;

[0028] The attribution probability is calculated by normalizing the attribution scores of all nodes using a flexible maximum function to obtain the final attribution probability.

[0029] Furthermore, the generation process of the scheduling policy in S4 is as follows:

[0030] The knowledge graph node with the highest attribution probability in the indication signal is used as the query index to retrieve the associated solution node in the knowledge graph, so as to automatically generate a high-risk task handling plan.

[0031] Furthermore, the scheduling policies implemented in S5 include:

[0032] The indicator signal is used as a dynamic weight to adjust the parameters of the preset personnel scheduling algorithm; and based on the adjusted algorithm, supervisory personnel whose professional fields match those required by the high-risk task handling plan are pre-scheduled to future high-risk areas predicted by the risk transmission path.

[0033] Furthermore, the scheduling strategy is generated based on indication signals:

[0034] The scheduling strategy is to use the knowledge graph node with the highest attribution probability in the indication signal as the query index, retrieve the associated solution in the knowledge graph, and generate a response plan.

[0035] The contingency plan includes analyzing the risk entropy flow field to determine the downstream key areas of risk transmission, and dynamically enhancing the retrieved solutions based on the downstream key areas.

[0036] Furthermore, the execution of the scheduling strategy involves using indication signals as dynamic weights to adjust the parameters of the personnel scheduling algorithm in order to pre-schedule matching personnel.

[0037] Adjusting the parameters of the personnel scheduling algorithm includes: constructing a multi-objective scheduling utility function that includes team capability matching degree, risk handling urgency, travel time cost, and proactive intervention value;

[0038] The value of proactive intervention is calculated based on the probability of risk propagation to downstream interception points and the importance coefficient of the engineering components associated with the interception points;

[0039] Pre-scheduling involves selecting candidate teams that maximize the multi-objective scheduling utility function value and scheduling them to future high-risk areas predicted by the risk entropy flow field.

[0040] The technical effects and advantages of this invention are as follows:

[0041] This invention transforms discrete, semantically meaningless schedule deviation data into a continuously changing, quantitative physical field that dynamically describes the stability of the project's macro-process, enabling macro-level and quantitative perception of risk. Furthermore, by constructing a dynamically updated engineering knowledge graph, it transforms static, unstructured document knowledge into a micro-logical network precisely registered with the project's spatiotemporal context. By fusing these two heterogeneous modalities, the invention effectively breaks down the data barriers between macro-level conditions and micro-level knowledge, allowing for a deeper understanding of the underlying implications and laying a solid data foundation for in-depth risk tracing.

[0042] This invention employs a fusion model of bidirectional causal reasoning; instead of unidirectional cause-to-effect reasoning, it performs forward propagation risk evolution simulation and backward reasoning logical root cause exploration in parallel; only when a node simultaneously satisfies the two conditions of being a reasonable risk evolution path and being a logically consistent risk root cause will it be assigned a high attribution probability; the dual evidence chain verification mechanism greatly suppresses false positives and false negatives, and can accurately locate the specific technical node causing the problem with high confidence from complex relationships;

[0043] This invention uses high-confidence diagnostic signals as the driving core to construct a closed-loop control system that tightly couples suggestion and execution. It can adaptively generate context-aware contingency plans that include key downstream interception points based on the root causes of risks. Furthermore, by solving a multi-objective utility function that includes the value of proactive intervention, it can synthesize and schedule the optimal response team and pre-deploy them to key locations where risks are about to erupt. This enhances the resilience of projects against risk disturbances and the efficiency of resource utilization, and endows the entire management system with intelligent self-healing capabilities.

[0044] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the steps provided by the present invention;

[0046] Figure 2 This is a schematic diagram of the calculation and acquisition steps of the construction cycle entropy and risk entropy flow field provided by the present invention;

[0047] Figure 3 This is a schematic diagram of the calculation steps for attribution probability provided by the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0049] like Figures 1 to 3 As shown in the figure, an adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data provided by an embodiment of the present invention includes the following steps:

[0050] S1: Obtain the first data characterizing the uncertainty of the project's macro-process. The first data includes the actual completion time series of each construction task.

[0051] Based on the actual completion time series of each construction task, S1 calculates the construction cycle entropy, which characterizes the uncertainty of the project's macro process, using the information entropy model; and based on the preset task dependencies, it constructs a risk entropy flow field that describes the intensity of risk transmission between tasks.

[0052] The actual start and end times of each construction task are obtained. In the entropy field engine module of the information entropy model calculation layer, the time series data is input into the preset information entropy model (in this embodiment, by calculating the Körbek-Leibler divergence between the actual time distribution and the planned time distribution). The output is a quantified and continuously changing construction cycle entropy value. Combined with the pre- and post-task logical relationships extracted from the BIM model, the risk transmission coefficient between each task node is calculated by a graph theory algorithm (in this embodiment, a PageRank-like algorithm). Thus, a dynamic "risk entropy flow field" is constructed in the four-dimensional spatiotemporal (three-dimensional space + time) of the project.

[0053] The first data is concretized into construction cycle entropy and risk entropy flow fields with clear physical meaning; it has the ability to conduct macroscopic, quantitative and continuous monitoring of the overall health of the project.

[0054] S2: Obtain the second data representing the micro-logical relationships of the project. The second data includes an engineering knowledge graph of engineering entities and the relationships between entities.

[0055] The second data in S2 automatically extracts engineering entities and relationships between entities from unstructured supervision documents using natural language processing technology to construct and update the engineering knowledge graph; and marks each node in the knowledge graph with its preset mapping position in the project time and space.

[0056] In the knowledge graph engine of the model computation layer, a natural language processing pipeline is deployed. The natural language processing pipeline continuously processes the input supervision documents, including construction plans and acceptance specifications in this embodiment. It extracts engineering entities such as "concrete pouring process" and "reinforcement binding specifications" through named entity recognition technology, and then identifies the logical relationships such as preconditions and constraints between them through relation extraction technology, thereby constructing or dynamically updating an engineering knowledge graph. During the construction process, the context information in the document or the association with the BIM model is used to mark each graph node with its corresponding component ID, spatial region and planned timestamp, etc., to complete the registration of logical space and physical space-time.

[0057] The second data is concretized into a dynamic and automatically constructed engineering knowledge graph; this solves the pain point of the difficulty in updating traditional expert knowledge bases, enabling continuous learning of the latest project knowledge and standards. The introduction of spatiotemporal tags lays an indispensable technical foundation for the subsequent fusion core to accurately map and align macroscopic physical field signals with microscopic logical graph nodes.

[0058] The steps for calculating and obtaining the flow field of construction cycle entropy and risk entropy are as follows:

[0059] Construction cycle entropy: Based on the deviation between the actual completion time and the planned completion time of the task, and combined with the task resource constraint degree that characterizes the scarcity of task resources, the entropy of resource-constrained tasks is calculated.

[0060] The task resource constraint is calculated as follows: obtain the concurrent demand and total availability of each type of resource associated with the task, calculate the resource saturation, and perform a weighted sum of the saturation of all associated resources. Finally, the result is obtained by mapping through a preset normalization function.

[0061] Constructing a risk entropy flow field: Based on the real-time construction cycle entropy of upstream tasks, dynamically adjust the risk transmission coefficient connecting upstream and downstream tasks;

[0062] The dynamic adjustment steps are as follows: take the construction cycle entropy of the upstream task as input, calculate the dynamic enhancement factor through the exponential function, and multiply the dynamic enhancement factor by the basic transmission weight to obtain the updated risk transmission coefficient.

[0063] In this embodiment, the key parameters involved are defined as follows:

[0064] The parameter symbol for the planned task cycle is: This refers to the index marked as For a single construction task, the standard completion time, measured in hours, is preset in the project baseline plan; the planned task cycle. It is obtained directly from the baseline plan database of the project management software (in this embodiment, Primavera P6 or Microsoft Project is used);

[0065] The parameter symbol for the actual task cycle is: This refers to the index marked as The actual time, expressed in hours, consumed from the start to the completion of a single construction task; the actual task cycle. The data is collected in real time through the work hour reporting system on the mobile terminal application of on-site construction personnel, or through the data interface with IoT sensors (in this embodiment, sensors associated with the on / off status of specific equipment).

[0066] The parameter symbol for task resource constraint is: A dimensionless parameter whose value ranges within the interval [0,1], representing the index label. The intensity of resource competition or resource scarcity faced by a single construction task during execution; the resource constraint degree of the task. The higher the value, the stronger the task's dependence on scarce resources, and the greater the potential impact of its delay.

[0067] The calculation logic is as follows: by combining with the task Each type of key resource associated The concurrent demand is divided by the total available resources to obtain the resource saturation. Then, the saturation of all associated critical resources is weighted and summed, and finally mapped using an S-shaped normalization function. The specific calculation is as follows: first, calculate the resource... In the mission The number of concurrent tasks within the execution time window is denoted as ; Obtain resources The total number of available units, denoted as Resources The saturation level is the ratio of the number of concurrent tasks to the total number of available units, which determines the saturation level of the tasks. The saturation level of all critical resources on which the system relies (in this embodiment, critical resources include a specific model of tower crane and skilled welders) is multiplied by the resource importance weight assigned by experts based on historical data. The summation yields the comprehensive resource pressure index. This index is then input into a preset logistic function, which smoothly maps it to the [0,1] interval to obtain the final task resource constraint. .

[0068] The parameter symbol for the entropy of resource-constrained tasks is: ; A dimensionless parameter whose value ranges within the interval [0,1], representing an index label. The degree of uncertainty or chaos in the process of a single construction task under the influence of resource constraints;

[0069] The calculation logic is as follows: The planned task period... Compared with the actual task cycle These are considered as the expected values ​​of two probability distributions, and two Poisson distributions are constructed based on them to simulate the probability of task completion time. Then, the Körbek-Leibler divergence between these two Poisson distributions is calculated to obtain the basic time deviation. The core improvement lies in combining this basic time deviation with the previously defined task resource constraint. Multiplying these values ​​yields a weighted bias; its physical meaning is that for tasks highly dependent on scarce resources, even slight delays are amplified, resulting in higher entropy values. Mapping the weighted biases of all tasks to the [0,1] interval using a max-min normalization method yields the final resource-constrained task entropy. .

[0070] The parameter sign of the entropy-induced risk transmission coefficient is: A dimensionless parameter with values ​​in the interval [0,1], representing the risk from upstream tasks. Downstream tasks The dynamic probability or intensity of conduction;

[0071] The calculation logic is as follows: Based on the pre- and post-task logical relationships in the BIM model or project management software, a basic task dependency directed graph is constructed, where each edge has an initial basic transmission weight, which is calibrated by experts based on the tightness of the relationship between tasks; the core improvement is that the basic transmission weight is not static; in each calculation iteration, from the task... Pointing to task The weight of the edge will be determined by the upstream task. Current resource-constrained task entropy Dynamic modulation is performed; specifically, the final entropy-induced risk transmission coefficient. It is the basic transmission weight and the resource-constrained task entropy. It is the product of the input exponential function values; the design of the exponential function value ensures that when the entropy value of the upstream task is low, it has little impact on the transmission coefficient; while when the entropy value is close to 1, it will exponentially amplify the transmission coefficient; which means that a task node in a highly chaotic state will become a super-propagator of risk.

[0072] The entropy field engine module and the knowledge graph engine work together to calculate the process, and the specific steps are as follows:

[0073] The system receives three types of data in parallel: real-time start and end times of each construction task from the field terminal; baseline plans, task dependency logic, and resource allocation plans from the project management software; and unstructured supervision documents from the document server. In this embodiment, unstructured supervision documents include construction plans, safety specifications, etc., in PDF and Word formats.

[0074] Knowledge graph construction and spatiotemporal mapping (executed by the knowledge graph engine):

[0075] Document parsing and information extraction: The natural language processing pipeline is started to process the received supervision documents; through the named entity recognition model, engineering entities such as "C50 high-strength concrete pouring" and "deep foundation pit first-level hot work operation" are extracted; through the relation extraction model, the logical relationships between these entities are identified, such as "preceded" and "constraint condition is".

[0076] Graph generation: Extracted entities are used as nodes and relationships are used as edges, and stored in a graph database to form an engineering knowledge graph;

[0077] Spatiotemporal registration: During the information extraction phase, information related to BIM component ID, floor, area code and planned date in the document is parsed simultaneously, and this information is attached as attributes to each node of the knowledge graph to complete the accurate mapping from logical space to project physical spatiotemporal space;

[0078] Dynamic risk entropy flow field construction (executed by the entropy field engine):

[0079] Parameter calculation: for each task currently in progress or already completed. Calculate the resource constraints of each task. and resource-constrained task entropy ;

[0080] Transmission coefficient update: Traverse each edge in the task dependency directed graph, starting from the task... arrive According to the upstream task The latest resource-constrained task entropy Value, real-time updated entropy-induced risk transmission coefficient ;

[0081] Risk value iterative calculation: On the updated dynamic weight graph, a PageRank-like iterative algorithm is executed once; the risk value of each task node is initialized, and then in multiple iterations, its own risk value is continuously updated according to the risk values ​​pointing to all upstream nodes and the corresponding entropy-induced risk transmission coefficient, until the risk values ​​of all nodes converge.

[0082] Field Construction and Output: The risk values ​​of all task nodes after convergence, and the entropy-induced risk transmission coefficients between nodes, together constitute a complete and dynamic risk entropy flow field; the output format of the risk entropy flow field is a structured dataset containing task ID, task risk value, and risk transmission path and intensity.

[0083] Final output: The final output consists of two independent but spatiotemporally registered data structures: a dynamically updated engineering knowledge graph and a real-time calculated risk entropy flow field. These two data structures are passed to the downstream fusion core as input for its risk attribution and linkage control.

[0084] S3: Based on the first and second data, the data is processed through a preset fusion model to generate an indication signal that represents the probability of risk root cause attribution.

[0085] The pre-defined fusion model in S3 is a knowledge graph activation and propagation model under a risk entropy flow field; the processing flow of the knowledge graph activation and propagation model is as follows:

[0086] When the local construction cycle entropy in the risk entropy flow field exceeds the preset activation threshold, knowledge graph nodes that have a preset mapping relationship with the local area are activated; and according to the risk transmission path defined in the risk entropy flow field, activation energy is transmitted in the knowledge graph to calculate the attribution probability of each knowledge graph node as the root cause of risk, thereby generating an indication signal.

[0087] The knowledge graph activation and transmission model for fusion cores; its algorithm logic is as follows:

[0088] Field-graph coupling: The risk entropy flow field is used as the background field, and the knowledge graph with pre-labeled spatiotemporal tags is used as the embedded network;

[0089] Node activation: Periodically scan the entropy flow field; once the construction cycle entropy of a certain spatial region exceeds the preset activation threshold, immediately query the knowledge graph node set that has a spatiotemporal mapping relationship with the region, and assign these nodes initial activation energy;

[0090] Energy transmission: The activated energy is not limited to the initial node; it will be iteratively transmitted in the network topology of the knowledge graph according to the transmission path defined in the entropy flow field and described by the risk transmission coefficient; the transmission algorithm is based on belief propagation of graph neural network, and the energy will decay according to the relationship type of graph edge (in this embodiment, the relationship type includes "strong constraint" and "weak influence") during the transmission process;

[0091] Attribution calculation: After several iterations, the final energy value of each node in the knowledge graph constitutes the risk attribution probability vector; each component in the risk attribution probability vector represents the probability that the corresponding micro-technical node is the current macro-risk root cause; the risk attribution probability vector is the final generated indicator signal.

[0092] When the local construction cycle entropy in the risk entropy flow field exceeds the threshold, the associated knowledge graph node is activated and energy transfer is performed to calculate the attribution probability.

[0093] During energy transfer, the forward propagation process and the backward reasoning process are executed in parallel;

[0094] The forward propagation process is as follows: based on the risk transmission coefficient defined in the risk entropy flow field, the forward propagation confidence of each node is iteratively updated in the knowledge graph;

[0095] The backward reasoning process is as follows: Based on the pre-defined logical relationship type weights in the knowledge graph, the backward reasoning confidence of each node in the knowledge graph is iteratively updated;

[0096] The attribution probability is calculated as follows: For each knowledge graph node, the final forward propagation confidence and the final backward reasoning confidence of the knowledge graph node are multiplied and fused to generate the attribution score;

[0097] The attribution probability is calculated by normalizing the attribution scores of all nodes using a flexible maximum function to obtain the final attribution probability.

[0098] In this embodiment, the key parameters involved are defined as follows:

[0099] The parameter sign for the weight of relation type in a knowledge graph is: A dimensionless parameter whose value ranges within the interval [0,1], representing the first element in the knowledge graph. The importance or strength of various logical relation types (in this embodiment, relation types include "preceded", "constraint condition", and "constituent part") in causal reasoning;

[0100] The calculation logic is as follows: An expert group composed of multiple senior domain experts (in this embodiment, senior domain experts include the chief engineer and project manager) is organized to evaluate all possible logical relationship types in the project; using the Delphi method or the analytic hierarchy process, experts are asked to compare and score the contribution of different relationships in the cause-effect derivation chain pairwise; in this embodiment, A is a prerequisite for B, and the weight of this strong causal relationship should be significantly higher than that of the weak correlation relationship where A and B belong to the same sub-project; the scoring matrix of all experts is collected, its geometric mean is calculated, and a consistency test is performed to finally obtain a set of convergent weight values ​​that can reflect the domain consensus, and these values ​​are normalized to the [0,1] interval.

[0101] Numerical determination example: Weights of the "preceding" relationship, as determined by experts. The weight of the "constraint condition is" relation is determined to be 0.9. The weight was determined to be 0.7; while the weight of the "constituent part is" relation was... It was determined to be 0.2;

[0102] The parameter symbol for the initial activation energy is: A dimensionless parameter with values ​​in the interval (0,1], representing knowledge graph nodes associated with high-entropy tasks. The intensity at which it is initially ignited;

[0103] The calculation logic is as follows: A global entropy activation threshold is set. This threshold is automatically determined by statistical analysis of historical project data using the maximum inter-class variance method to distinguish between normal fluctuations and abnormal events; when a task... Resource-constrained task entropy When the entropy activation threshold is exceeded, the knowledge graph nodes associated with its spatiotemporal mapping... That is, it is activated; initial activation energy The value is obtained by representing the portion exceeding the entropy activation threshold, i.e., the entropy of resource-constrained tasks. The difference between the value and the entropy activation threshold is linearly mapped to the interval (0,1]; where the entropy activation threshold is 0.7.

[0104] The parameter symbol for the forward propagation confidence is: The dimensionless parameter, which is dynamically updated during the calculation process and has a value range of [0,1], represents the knowledge graph node. Confidence level as part of the risk evolution path;

[0105] The symbol for the backward reasoning confidence score is: The dimensionless parameter, which is dynamically updated during the calculation process and has a value range of [0,1], represents the knowledge graph node. Confidence level as the initial source of risk;

[0106] The parameter symbol for attribution probability is: The dimensionless parameter, whose value ranges from [0,1], is the final output of this embodiment, representing the knowledge graph node. It is the total probability of the root causes of the currently observed macroeconomic risks;

[0107] The specific steps are as follows:

[0108] Data input: Receive real-time risk entropy flow field dataset; engineering knowledge graph with spatiotemporal registration completed provided by knowledge graph engine;

[0109] Node activation: Traverse all task nodes in the risk entropy flow field; for each resource-constrained task entropy... Tasks exceeding the entropy activation threshold Find the corresponding node in the knowledge graph. Calculate the initial activation energy ;

[0110] Confidence initialization: For all nodes in the knowledge graph , forward propagation confidence and backward reasoning confidence All are initialized to the initial activation energy. For nodes that are not activated, the initial value is 0.

[0111] Bidirectional causal reasoning iteration (core loop): Executes a fixed number of iterations; in each iteration, the following two update processes are executed in parallel:

[0112] Forward propagation process: For each node in the graph Update forward propagation confidence The new confidence level is calculated using an aggregation function, whose input includes nodes. The self-confidence of the previous iteration, and all nodes pointing to the upstream task nodes. The confidence level of the previous iteration and the entropy-induced risk transmission coefficient between them The sum of products; this process simulates how risk spreads in physically dependent networks;

[0113] Backward reasoning process: For each node in the graph Updated inference confidence The new confidence level is also calculated using an aggregation function, but its input is the node. The self-confidence of the previous iteration, and all downstream nodes it points to. The confidence score of the previous iteration and the weight of the knowledge graph relationship type between them The sum of the products of these products; this process simulates how to deduce the cause from the result.

[0114] Confidence fusion: After the iteration loop ends, for each node in the knowledge graph... Calculate the final, unnormalized attribution score; the attribution score is calculated by dividing the nodes... Final forward propagation confidence Its final backward reasoning confidence Perform multiplication;

[0115] Normalization and output:

[0116] Final attribution probability calculation: The unnormalized attribution scores of all nodes are input as a vector into the flexible maximum function. The flexible maximum function exponentially normalizes all unnormalized attribution scores, ensuring that the sum of the output values ​​of all nodes is 1, while highlighting high-scoring terms. In the vector output by the function, each node... The corresponding value is the final attribution probability;

[0117] Final output: The final output of the process is a risk attribution probability vector; the dimension of the risk attribution probability vector is equal to the number of nodes in the knowledge graph, and the value of each dimension is the probability that the corresponding node is the root cause of the risk; after being formatted, the risk attribution probability vector is transmitted as an indicator signal to the downstream decision-making and execution layers.

[0118] S4: Generate a scheduling strategy containing control commands based on the indication signal;

[0119] The process of generating the scheduling policy in S4 is as follows:

[0120] The knowledge graph node with the highest attribution probability in the indication signal is used as the query index to retrieve the associated solution node in the knowledge graph, so as to automatically generate a high-risk task handling plan.

[0121] After the fusion core outputs the risk attribution probability vector, the risk attribution probability vector is first sorted. The top N knowledge graph nodes with the highest probability values ​​are selected as high-suspicion root causes. The IDs of these nodes are used as query instructions to perform a graph traversal query in the graph database to find all nodes connected to these nodes through relationship edges such as "solution", "inspection items", and "association specifications". These query results are then structured and combined into a human-readable high-risk task handling plan. In this embodiment, this plan includes a Markdown or PDF document containing inspection steps, reference specifications, and precautions.

[0122] S5: Execute scheduling strategies to adaptively adjust project resources.

[0123] The scheduling strategies implemented in S5 include:

[0124] The indicator signal is used as a dynamic weight to adjust the parameters of the preset personnel scheduling algorithm; and based on the adjusted algorithm, supervisory personnel whose professional fields match those required by the high-risk task handling plan are pre-scheduled to the future high-risk areas predicted by the risk transmission path.

[0125] Upon receiving the risk attribution probability vector, control actions are executed. On one hand, based on the contingency plan generated by S4, the required professional skills for the task are determined. On the other hand, the attribution probability vector is used as a weight for the dynamic task urgency and professional matching degree, and the objective function of its internal personnel scheduling algorithm is modified in real time. The modified algorithm will greatly increase the probability of selecting professionally matched and currently idle supervisors, and push preset instructions containing "contingency plan" and "suggested destination" to the mobile terminal based on the risk transmission path prediction of the next hop location of the risk.

[0126] The scheduling strategy is generated based on indication signals:

[0127] The scheduling strategy is to use the knowledge graph node with the highest attribution probability in the indication signal as the query index, retrieve the associated solution in the knowledge graph, and generate a response plan.

[0128] The contingency plan includes analyzing the risk entropy flow field to determine the downstream key areas of risk transmission, and dynamically enhancing the retrieved solutions based on the downstream key areas.

[0129] The execution of the scheduling strategy involves using indicator signals as dynamic weights to adjust the parameters of the personnel scheduling algorithm in order to pre-schedule matching personnel.

[0130] Adjusting the parameters of the personnel scheduling algorithm includes: constructing a multi-objective scheduling utility function that includes team capability matching degree, risk handling urgency, travel time cost, and proactive intervention value;

[0131] The value of proactive intervention is calculated based on the probability of risk propagation to downstream interception points and the importance coefficient of the engineering components associated with the interception points;

[0132] Pre-scheduling involves selecting candidate teams that maximize the value of the multi-objective scheduling utility function and scheduling them to future high-risk areas predicted by the risk entropy flow field;

[0133] In this embodiment, the key parameters involved are defined as follows:

[0134] The parameter notation of the risk attribution probability vector is: The vector output by the fusion core, each component of which represents the probability that the corresponding node in the knowledge graph is a source of risk;

[0135] The risk entropy flow field, a structured dataset output by the entropy field engine, describes the dynamic transmission intensity and path of risk between tasks.

[0136] The parameter symbols for the personnel professional skill vector are: For indexes marked as For the supervisors, an N-dimensional skill vector is pre-established for them; each dimension of the skill vector corresponds to a professional skill (in this embodiment, it includes "high voltage electrician certification" and "deep foundation pit support experience"), and its value is a value in the range of [0,1], representing the proficiency or qualification level of the personnel in the skill.

[0137] The calculation logic is as follows: The administrator enters and quantifies the personnel's resumes, qualification certificates, historical work evaluations and training records into the preset skill library; in this embodiment, the rating is 0.9 for those holding senior technician certificates, 0.7 for intermediate level, 0.5 for junior level, and 0.2 for those without certificates but with relevant experience.

[0138] The parameter notation of the task skill requirement vector is: For the generated index marked as The high-risk task handling plan is analyzed to generate a demand vector with the same dimension as the personnel skill vector; each dimension of the demand vector represents the degree of dependence of the plan on the corresponding skills.

[0139] The calculation logic is as follows: After the emergency response plan is generated, keyword extraction and semantic analysis are performed on its text content, and it is matched with skill tags in the skill library; the matched skill dimension is assigned a value of 1, and the unmatched dimension is assigned a value of 0; for complex tasks that require multiple skills, higher weight values ​​are further assigned by limiting words in the text (in this embodiment, including "must" and "critical").

[0140] The parameter symbol for team capability matching is: A dimensionless parameter with a value in the range of [0,1], representing the degree of matching between the comprehensive professional skills of the candidate team (or a single person) and the task requirements;

[0141] The calculation logic is as follows: The comprehensive skill vector of the candidate team is obtained by combining the professional skill vectors of all team members. The vectors obtained by bitwise addition are used to calculate the team's overall skill vector and the task skill requirement vector. The cosine similarity between them; the physical meaning of the calculation method is that it not only considers whether the team has the required skills, but also the redundancy and complementarity of skills;

[0142] The symbol for the urgency of risk management is: A dimensionless parameter with a value in the range [0,1], representing the urgency of dealing with the current risk;

[0143] The calculation logic is as follows: The risk attribution probability vector... The highest probability value is mapped through a pre-defined exponential function to amplify the urgency of high-probability risks;

[0144] The parameter symbol for the value of active intervention is: The dimensionless parameter with a value range of [0,1] creatively quantifies the value of preventive behavior and represents the potential losses that can be avoided by sending the team downstream of the risk transmission path to intercept it.

[0145] The calculation logic is as follows: By multiplying the risk transmission coefficients of all nodes along the risk transmission path from the risk root node to the downstream predicted interception point, the probability of unimpeded risk propagation is obtained; this probability is then multiplied by the importance coefficient of the engineering component associated with the downstream interception point (in this embodiment, the importance coefficient is a preset coefficient determined by experts based on its importance in structure, function, and cost); the physical meaning is that intercepting the risk that is about to impact the critical load-bearing column is far more valuable than intercepting the risk that will only affect the non-load-bearing wall, where the importance coefficient is 0.95.

[0146] The specific steps of the control process are as follows:

[0147] Data input: Receive risk attribution probability vector and risk entropy flow field;

[0148] Root cause identification and contingency plan generation:

[0149] High-confidence root cause screening: risk attribution probability vector The knowledge graph nodes with probability values ​​greater than a preset confidence threshold (in this embodiment, the confidence threshold is set to 0.5) are selected as high-confidence risk sources after being sorted in descending order.

[0150] Context-adaptive contingency plan generation: For each high-confidence root source node, perform a graph traversal in the knowledge graph to retrieve all nodes connected to the node through relationships such as solutions and check items, forming a basic handling plan;

[0151] At the same time, query the risk entropy flow field to find the transmission path with the highest risk transmission coefficient starting from the root node, and determine the downstream key task or BIM component that the path points to.

[0152] Information on downstream key tasks or components is dynamically and structurally added to the basic treatment plan to generate high-priority inspection instructions with clear objectives. In this embodiment, "inspect concrete curing records" is enhanced to "immediately conduct rebound hammer testing on beam B-101, as there is a high risk of improper curing at its upstream concrete pouring node." The task skill requirement vector is parsed from the enhanced plan text. ;

[0153] Optimal team scheduling and command issuance:

[0154] Target area identification: The physical construction area where the downstream critical tasks are located, as determined by the analysis;

[0155] Candidate team construction: Obtain a list of all supervisors whose status is "available" and generate all possible combinations of personnel (single or multi-person teams within a preset size) as candidate teams;

[0156] Multi-objective utility evaluation: For each candidate team, calculate the comprehensive scheduling utility function value; the calculation logic of the comprehensive scheduling utility function value is as follows: [Calculate the team capability matching degree]. urgency of risk management Active intervention value Each component is multiplied by its respective strategic weight and then summed, and then the weighted average travel time cost for team members to reach the target area is subtracted; the strategic weights are configured by the project commander-in-chief based on the current project stage (in this embodiment, the time cost weight can be increased during the "rush to meet deadlines" stage);

[0157] Optimal decision: Select the candidate team with the highest comprehensive scheduling utility function value as the final execution team;

[0158] Command generation and issuance: Generate a digital command package containing three elements: the finalized list of the execution team; a context-adaptive contingency plan; and the BIM model location of the target area. The digital command package is then pushed to the mobile terminal APP of team members via a mobile communication network.

[0159] The final output is the executed, closed-loop control action, meaning the optimal team has been dispatched to the most critical location to execute the most suitable contingency plan.

[0160] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for processing multi-dimensional supervision data using an adaptive cross-modal attention mechanism, characterized in that, Includes the following steps: S1: Obtain the first data characterizing the uncertainty of the project's macro-process. The first data includes the actual completion time series of each construction task. S2: Obtain the second data representing the micro-logical relationships of the project. The second data includes an engineering knowledge graph of engineering entities and the relationships between entities. S3: Based on the first and second data, the data is processed through a preset fusion model to generate an indication signal representing the attribution probability of the root cause of the risk. S4: Generate a scheduling strategy containing control commands based on the indication signal; S5: Execute scheduling strategies to adaptively adjust project resources.

2. The method for processing multi-dimensional supervision data using an adaptive cross-modal attention mechanism according to claim 1, characterized in that: In S1, based on the actual completion time series of each construction task, the construction cycle entropy, which characterizes the uncertainty of the project's macro process, is calculated using the information entropy model; and based on the preset task dependencies, a risk entropy flow field describing the intensity of risk transmission between tasks is constructed.

3. The adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data according to claim 2, characterized in that: The second data in S2 automatically extracts engineering entities and relationships between entities from unstructured supervision documents using natural language processing technology to construct and update the engineering knowledge graph; and marks the preset mapping position of each node in the knowledge graph in the project time and space.

4. The adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data according to claim 3, characterized in that: The steps for calculating and obtaining the flow field of construction cycle entropy and risk entropy are as follows: Construction cycle entropy: Based on the deviation between the actual completion time and the planned completion time of the task, and combined with the task resource constraint degree that characterizes the scarcity of task resources, the entropy of resource-constrained tasks is calculated. The task resource constraint is calculated as follows: obtain the concurrent demand and total availability of each type of resource associated with the task, calculate the resource saturation, and perform a weighted sum of the saturation of all associated resources. Finally, the result is obtained by mapping through a preset normalization function. Constructing a risk entropy flow field: Based on the real-time construction cycle entropy of upstream tasks, dynamically adjust the risk transmission coefficient connecting upstream and downstream tasks; The dynamic adjustment steps are as follows: take the construction cycle entropy of the upstream task as input, calculate the dynamic enhancement factor through the exponential function, and multiply the dynamic enhancement factor by the basic transmission weight to obtain the updated risk transmission coefficient.

5. The adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data according to claim 4, characterized in that: The pre-defined fusion model in S3 is a knowledge graph activation and propagation model under a risk entropy flow field; the processing flow of the knowledge graph activation and propagation model is as follows: When the local construction cycle entropy in the risk entropy flow field exceeds the preset activation threshold, knowledge graph nodes that have a preset mapping relationship with the local area are activated; and according to the risk transmission path defined in the risk entropy flow field, activation energy is transmitted in the knowledge graph to calculate the attribution probability of each knowledge graph node as the root cause of risk, thereby generating an indication signal.

6. The adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data according to claim 5, characterized in that: When the local construction cycle entropy in the risk entropy flow field exceeds the threshold, the associated knowledge graph node is activated and energy transfer is performed to calculate the attribution probability. During energy transfer, the forward propagation process and the backward reasoning process are executed in parallel; The forward propagation process is as follows: based on the risk transmission coefficient defined in the risk entropy flow field, the forward propagation confidence of each node is iteratively updated in the knowledge graph; The backward reasoning process is as follows: Based on the pre-defined logical relationship type weights in the knowledge graph, the backward reasoning confidence of each node in the knowledge graph is iteratively updated; The attribution probability is calculated as follows: For each knowledge graph node, the final forward propagation confidence and the final backward reasoning confidence of the knowledge graph node are multiplied and fused to generate the attribution score; The attribution probability is calculated by normalizing the attribution scores of all nodes using a flexible maximum function to obtain the final attribution probability.

7. The adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data according to claim 6, characterized in that: The process of generating the scheduling policy in S4 is as follows: The knowledge graph node with the highest attribution probability in the indication signal is used as the query index to retrieve the associated solution node in the knowledge graph, so as to automatically generate a high-risk task handling plan.

8. The adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data according to claim 7, characterized in that: The scheduling strategies implemented in S5 include: The indicator signal is used as a dynamic weight to adjust the parameters of the preset personnel scheduling algorithm; and based on the adjusted algorithm, supervisory personnel whose professional fields match those required by the high-risk task handling plan are pre-scheduled to future high-risk areas predicted by the risk transmission path.

9. The adaptive cross-modal attention mechanism method for processing multi-dimensional supervision data according to claim 8, characterized in that: The scheduling strategy is generated based on indication signals: The scheduling strategy is to use the knowledge graph node with the highest attribution probability in the indication signal as the query index, retrieve the associated solution in the knowledge graph, and generate a response plan. The contingency plan includes analyzing the risk entropy flow field to determine the downstream key areas of risk transmission, and dynamically enhancing the retrieved solutions based on the downstream key areas.

10. The method for processing multi-dimensional supervision data using an adaptive cross-modal attention mechanism according to claim 9, characterized in that: The execution of the scheduling strategy involves using indicator signals as dynamic weights to adjust the parameters of the personnel scheduling algorithm in order to pre-schedule matching personnel. Adjusting the parameters of the personnel scheduling algorithm includes: constructing a multi-objective scheduling utility function that includes team capability matching degree, risk handling urgency, travel time cost, and proactive intervention value; The value of proactive intervention is calculated based on the probability of risk propagation to downstream interception points and the importance coefficient of the engineering components associated with the interception points; Pre-scheduling involves selecting candidate teams that maximize the multi-objective scheduling utility function value and scheduling them to future high-risk areas predicted by the risk entropy flow field.

Images