A highway construction risk prediction method and system based on big data

Abstract

The present application relates to the technical field of risk prediction, in particular to a highway construction risk prediction method and system based on big data, comprising the following steps: obtaining construction safety regulations and equipment fault tree, constructing management-fault coupling logic tree, combining digital twin state vector and regulation defects to determine initial risk probability, then analyzing real-time instructions to update digital twin state, finally using Bayesian network to update basic event instantaneous occurrence probability at discrete time steps to generate risk probability evolution sequence.In the present application, by constructing the logical association of management and failure, the failure induced by regulation defects is revealed, the risk traceability path is formed, the isolated risk assessment is abandoned, the equipment state is quantified by using real-time data, the expert scoring is replaced, the objective evaluation is realized, the risk is dynamically adjusted in real time by analyzing the instructions, the probability evolution sequence is generated based on the Bayesian network, and the trend prediction and prospective intervention are supported.

Description

Technical Field

[0001] This invention relates to the field of risk prediction technology, and in particular to a method and system for predicting highway construction risks based on big data. Background Technology

[0002] Risk prediction technology is an interdisciplinary field that utilizes the principles of information theory and systems theory to identify potential hazards in a system by collecting and analyzing relevant information. It then uses qualitative or quantitative methods to predict the probability and consequences of future risk events. The core of this field lies in building predictive models that leverage historical and real-time monitoring data to make probabilistic assessments and inferences about uncertain future events. Traditional highway construction risk prediction methods primarily rely on expert experience and qualitative analysis to identify and assess risks. These methods typically employ expert scoring, where experts subjectively score various risk indicators based on a pre-defined risk list, such as personnel qualifications, equipment status, geological conditions, and weather factors. The weights of each risk factor are then calculated using the analytic hierarchy process (AHP), ultimately leading to a comprehensive risk level.

[0003] Traditional methods for predicting risks in highway construction mainly rely on subjective scoring by experts on a pre-set list. This approach leads to significant human uncertainty in the assessment results and fails to reveal the deep coupling between management behavior and equipment status. Risk factors are viewed in isolation. In practice, because the assessment process is not continuous, the conclusions often lag behind dynamic changes on site. For example, when a piece of equipment is rated as low-risk in a static assessment, on-site violations or erroneous operations may cause its operating parameters to exceed safety thresholds. Such emerging risks generated by dynamic interactions cannot be identified in a timely manner, and may ultimately lead to on-site safety accidents due to a lack of early warning. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for predicting highway construction risks based on big data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting highway construction risks based on big data, comprising the following steps:

[0006] S1: Obtain the construction safety operation procedure and equipment fault tree, parse the construction safety operation procedure to construct the management process logic tree, associate the management process logic tree with the equipment fault tree, and calculate the management-fault coupling logic tree;

[0007] S2: Construct a digital twin state vector consisting of motor output power, hydraulic oil viscosity, and structural fatigue damage accumulation coefficient; analyze the defects in the construction safety operation procedure to determine the failure rate improvement value; update the initial probability of the basic event node of the management-fault coupling logic tree; and calculate the initial risk probability.

[0008] S3: Parse the real-time management instructions to extract the instruction parameters, calculate the dynamic adjustment rate to update the digital twin state vector, and generate the adjusted digital twin state vector;

[0009] S4: Based on the management-fault coupling logic tree, the initial risk probability, and the adjusted digital twin state vector, update the instantaneous occurrence probability of the basic events of the management-fault coupling logic tree using a Bayesian network at discrete time steps, connect the top-level event probabilities, and generate a risk probability evolution sequence.

[0010] As a further embodiment of the present invention, the management-fault coupling logic tree includes management failure event nodes, physical failure event nodes and cross-domain causal association paths, the initial risk probability includes risk benchmark value, static defect influence coefficient and top-level event identifier of the logic tree, the adjusted digital twin state vector includes instantaneous motor power prediction value, hydraulic oil viscosity after dynamic decay and structural fatigue damage coefficient after accelerated accumulation, and the risk probability evolution sequence includes discrete time step nodes and instantaneous failure probability value.

[0011] As a further aspect of the present invention, the calculation steps of the management-fault coupling logic tree are specifically as follows:

[0012] S101: Obtain the construction safety operation procedure, perform structured parsing on the text content of the construction safety operation procedure, identify management events, triggering conditions and execution order, convert the triggering conditions into logical AND gates, convert parallel execution relationships into logical OR gates, connect all the management events, and establish a management process logic tree;

[0013] S102: Obtain the equipment fault tree and call the management process logic tree. Extract the text descriptions of multiple leaf node events in the management process logic tree and multiple underlying fault events in the equipment fault tree. Calculate the semantic vector cosine similarity between the leaf node events and the underlying fault events. Associate event pairs whose cosine similarity exceeds a preset matching threshold value to generate a management event-fault event mapping set.

[0014] S103: Based on the management event-fault event mapping set, traverse the device fault tree. When the multiple underlying fault events are found to exist in the management event-fault event mapping set, call the management process logic tree or subtree structure associated with the multiple underlying fault events, replace the corresponding multiple underlying fault event nodes in the device fault tree, complete the tree structure coupling, and obtain the management-fault coupling logic tree.

[0015] As a further aspect of the present invention, the calculation steps for the initial risk probability are specifically as follows:

[0016] S201: Obtain the real-time output power of the equipment motor, the real-time viscosity of the hydraulic oil, and the cumulative coefficient of structural fatigue damage. Use the collected three parameter values ​​as components to construct a three-dimensional digital twin state vector. Simultaneously analyze all clauses of the construction safety operation procedure, identify two types of procedure defects: instruction logic conflict and missing prerequisite operation conditions. Based on the preset defect severity level mapping table, convert the identified procedure defects into quantified failure rate improvement values.

[0017] S202: Invoke the management-fault coupling logic tree, and according to the failure rate boosting value, retrieve the basic event nodes directly associated with and transformed by the procedure defect in the management-fault coupling logic tree, and perform a summation operation on the failure rate boosting value and the initial probability of the retrieved basic event nodes to establish an updated basic event probability set;

[0018] S203: Based on the overall topology of the management-fault coupling logic tree, the updated basic event probability set is called as the input probability of all the basic event nodes. According to the logic AND gate multiplication and logic OR gate addition operation rules defined in the management-fault coupling logic tree, the probability of occurrence of multiple intermediate events is passed from bottom to top and calculated layer by layer until the probability of occurrence of the top event is solved, and the initial risk probability is obtained.

[0019] As a further aspect of the present invention, the step of obtaining the adjusted digital twin state vector specifically comprises:

[0020] S301: Monitor the real-time management instruction data stream, perform text serialization parsing on the received real-time management instructions, identify instruction type keywords, core operation object identifiers and quantized values, and perform structured aggregation of the instruction type keywords, the core operation object identifiers and the quantized values ​​to generate an instruction parameter set;

[0021] S302: Extract the core operation object identifier and the quantized value from the instruction parameter set, and match the corresponding weight coefficient for each core operation object identifier according to the preset parameter influence weight lookup table. After multiplying all the quantized values ​​and the corresponding weight coefficients, perform an accumulation and summation operation to obtain the dynamic adjustment rate of the state vector.

[0022] S303: Call the three-dimensional digital twin state vector, and for the dynamic adjustment rate of the state vector, perform element-wise multiplication of the three components of the three-dimensional digital twin state vector—motor output power, hydraulic oil viscosity, and structural fatigue damage accumulation coefficient—with the dynamic adjustment rate of the state vector to obtain multi-component adjustment increments. Then, perform matrix summation of the multi-component adjustment increments with the original corresponding component values ​​of the three-dimensional digital twin state vector to generate the adjusted digital twin state vector.

[0023] As a further aspect of the present invention, the step of obtaining the risk probability evolution sequence specifically includes:

[0024] S401: Obtain the discrete time step, based on the management-fault coupling logic tree, the initial risk probability and the adjusted digital twin state vector, take multiple components of the adjusted digital twin state vector as Bayesian network evidence nodes, set the probability values ​​of associated basic events in the initial risk probability as prior probabilities, perform conditional probability inference, update the posterior probability of each basic event in the management-fault coupling logic tree, and obtain the instantaneous basic event probability set;

[0025] S402: Invoke the topology of the management-fault coupling logic tree, and according to the instantaneous basic event probability set, perform probability multiplication operation on the logical AND gate in the management-fault coupling logic tree, perform probability addition operation on the logical OR gate, and iteratively calculate the occurrence probability of all intermediate events from bottom to top until the top event, and obtain the single-step probability of the top event;

[0026] S403: For a preset number of consecutive discrete time steps, repeatedly execute the aforementioned process of updating the posterior probability and solving the probability of the top event occurrence, and arrange and combine the single-step top event probabilities generated for each discrete time step in chronological order to establish the risk probability evolution sequence.

[0027] As a further aspect of the present invention, the calculation steps for the dynamic adjustment rate of the state vector are as follows:

[0028] Obtain the instruction parameter set, parse the instruction parameter set to extract all the core operation object identifiers, the quantized values ​​and the corresponding instruction timestamps;

[0029] The preset parameter influence weight lookup table is invoked to obtain the basic weight value based on the identifier of each core operation object, and the time decay coefficient is calculated based on the time difference between the instruction timestamp and the current prediction time point.

[0030] Multiply the base weight value by the time decay coefficient to obtain the dynamic weight coefficient, and calculate the dynamic adjustment rate of the state vector according to the following formula:

[0031] ;

[0032] in, This represents the dynamic adjustment rate of the state vector. This represents the total number of instructions included in the instruction parameter set. From 1 to An integer used to identify the instruction. Representing the The quantized value corresponding to each instruction Representing the The basic weight value is obtained by querying the core operation object identifier of the instruction. The timestamp representing the current prediction time. Representing the The timestamp of the instruction, This represents the preset time decay constant used to control the decay rate;

[0033] Substitute the quantized values ​​of all the instructions, the basic weight values, the timestamp of the current prediction time, the timestamp of each instruction, and the time decay constant into the formula, perform cumulative summation and averaging processing, and generate the dynamic adjustment rate of the state vector.

[0034] As a further aspect of the present invention, the calculation steps for the quantified failure rate improvement value are specifically as follows:

[0035] The entire construction safety operation procedure is invoked, and the logical structure and dependencies of each clause are analyzed using a natural language processing model to identify all logical conflicts in the instructions and defects in the missing prerequisite operation conditions.

[0036] Assign a defect type coefficient and a propagation impact factor to each identified procedure defect, wherein the defect type coefficient for the instruction logic conflict is higher than the defect type coefficient for the missing precondition.

[0037] The quantified failure rate improvement value is calculated according to the following formula:

[0038] ;

[0039] in, Representing the The quantized failure rate improvement value associated with each basic event node. Representative and the The baseline failure rate, derived from historical data statistics and associated with each basic event node. Representative and the The total number of procedural defects directly associated with each of the aforementioned basic event nodes. From 1 to Integers used to identify procedural defects. Representing the The procedure defect described herein corresponds to a quantified severity score in the preset defect severity level mapping table. Representing the The defect type coefficient of the procedure defect described in the item. Representing the The propagation impact factor is determined by the location of the procedural defect in the management process logic tree;

[0040] Iterate through all the procedural defects associated with a single basic event node, obtain the quantified severity level score, the defect type coefficient, and the propagation impact factor, perform product and summation operations to obtain the quantified failure rate improvement value.

[0041] As a further aspect of the present invention, the calculation steps for the semantic vector cosine similarity are specifically as follows:

[0042] Obtain the text descriptions of the leaf node events of the management process logic tree and the underlying fault events of the device fault tree;

[0043] The pre-trained bidirectional encoder representation model is invoked, and the text descriptions of the leaf node events and the underlying fault events are respectively input into the pre-trained bidirectional encoder representation model to perform deep semantic feature extraction and obtain high-dimensional semantic vectors corresponding to multiple text descriptions.

[0044] For each event pair, extract the two corresponding high-dimensional semantic vectors, and calculate the standard cosine similarity between the two high-dimensional semantic vectors by vector dot product and vector magnitude product.

[0045] Simultaneously, the hierarchical depth of the leaf node events in the management process logic tree and the critical path contribution of the bottom-level fault events in the equipment fault tree are analyzed, and the hierarchical importance weight and path criticality weight are calculated respectively.

[0046] Multiply the hierarchy importance weight by the path criticality weight to obtain a comprehensive weight coefficient, and multiply the standard cosine similarity by the comprehensive weight coefficient to obtain the semantic vector cosine similarity.

[0047] A highway construction risk prediction system based on big data, wherein the system is used to implement the aforementioned highway construction risk prediction method based on big data, and the system includes:

[0048] The coupling logic tree construction module is used to obtain the construction safety operation procedures and equipment fault tree, parse the construction safety operation procedures to construct the management process logic tree, associate the management process logic tree with the equipment fault tree, calculate the management-fault coupling logic tree, and pass it to the initial risk quantification module.

[0049] The initial risk quantification module is used to construct a digital twin state vector consisting of motor output power, hydraulic oil viscosity, and structural fatigue damage accumulation coefficient; analyze the defects of the construction safety operation procedure to determine the failure rate improvement value; update the initial probability of the basic event node of the management-fault coupling logic tree; calculate the initial risk probability; and pass the digital twin state vector to the twin state dynamic update module.

[0050] The digital twin state dynamic update module is used to parse real-time management instructions to extract instruction parameters, calculate the dynamic adjustment rate to update the digital twin state vector, generate the adjusted digital twin state vector, and transmit it to the risk evolution sequence generation module.

[0051] The risk evolution sequence generation module is used to update the instantaneous occurrence probability of the basic events of the management-fault coupling logic tree using a Bayesian network at discrete time steps through the management-fault coupling logic tree, the initial risk probability, and the adjusted digital twin state vector, and connect the top event probability to generate a risk probability evolution sequence.

[0052] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0053] In this invention, by constructing the inherent logical connection between management processes and equipment failures, it is possible to reveal how deficiencies in operating procedures can directly induce equipment malfunctions, thereby forming a more comprehensive risk tracing path. This changes the previous assessment method where risk factors were independent of each other. By using real-time monitoring data such as motor output power to construct a state representation of construction equipment, it replaces qualitative scoring that relies on expert experience, achieving objective quantification of risk assessment. It also analyzes management instructions in real time to dynamically adjust equipment status, allowing risk probabilities to fluctuate in real time with on-site operations. Based on Bayesian networks, it generates a continuous risk probability evolution sequence, providing data support for risk trend prediction and proactive intervention. Attached Figure Description

[0054] Figure 1 This is a flowchart of the highway construction risk prediction method based on big data according to the present invention;

[0055] Figure 2 This is a flowchart of the management-fault coupling logic tree calculation process of the present invention;

[0056] Figure 3 This is a flowchart of the initial risk probability calculation for this invention;

[0057] Figure 4 This is a flowchart of the adjusted digital twin state vector generation process according to the present invention;

[0058] Figure 5 This is a flowchart of the risk probability evolution sequence generation process of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0060] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0061] Please see Figure 1 and Figure 2 This invention provides a technical solution: a method for predicting highway construction risks based on big data, comprising the following steps:

[0062] S1: Obtain the construction safety operation procedures and equipment fault tree, parse the construction safety operation procedures to construct the management process logic tree, and associate the management process logic tree with the equipment fault tree to calculate the management-fault coupling logic tree;

[0063] The management-fault coupling logic tree includes management failure event nodes, physical failure event nodes, and cross-domain causal relationship paths;

[0064] The specific steps for calculating the management-fault coupling logic tree are as follows:

[0065] S101: Obtain the construction safety operation procedure, perform structured parsing of the text content of the construction safety operation procedure, identify management events, triggering conditions and execution order, convert the triggering conditions into logical AND gates, convert the parallel execution relationship into logical OR gates, connect all management events, and establish a management process logic tree;

[0066] S102: Obtain the equipment fault tree and call the management process logic tree. Extract the text descriptions of multiple leaf node events in the management process logic tree and multiple underlying fault events in the equipment fault tree. Calculate the semantic vector cosine similarity between the leaf node events and the underlying fault events. Associate event pairs whose cosine similarity exceeds the preset matching threshold to generate a management event-fault event mapping set.

[0067] The specific steps for calculating the cosine similarity of semantic vectors are as follows:

[0068] Obtain text descriptions of leaf node events in the management process logic tree and underlying fault events in the device fault tree;

[0069] The pre-trained bidirectional encoder representation model is invoked, and the text descriptions of leaf node events and underlying fault events are input into the pre-trained bidirectional encoder representation model respectively to perform deep semantic feature extraction and obtain high-dimensional semantic vectors corresponding to multiple text descriptions.

[0070] For each event pair, extract its two corresponding high-dimensional semantic vectors, and calculate the standard cosine similarity between the two high-dimensional semantic vectors by vector dot product and vector magnitude product.

[0071] Meanwhile, the hierarchical depth of leaf node events in the management process logic tree and the contribution of bottom-level fault events to the critical path in the equipment fault tree are analyzed, and the hierarchical importance weight and path criticality weight are calculated respectively.

[0072] Multiply the hierarchy importance weight by the path criticality weight to obtain a comprehensive weight coefficient, and multiply the standard cosine similarity by the comprehensive weight coefficient to obtain the semantic vector cosine similarity.

[0073] S103: Based on the management event-fault event mapping set, traverse the device fault tree. When multiple underlying fault events are found to exist in the management event-fault event mapping set, call the management process logic tree or subtree structure associated with the multiple underlying fault events, replace the corresponding multiple underlying fault event nodes in the device fault tree, complete the tree structure coupling, and obtain the management-fault coupling logic tree.

[0074] S101: Taking a certain type of crawler crane used in highway bridge erection as the object, firstly, obtain its electronic document "Safety Operation Procedures for Crawler Cranes"; perform structured parsing of the document content, the specific actions of which are: firstly, process the procedure text into sentences; then perform lexical analysis on each sentence, including Chinese word segmentation and part-of-speech tagging; next, use the deterministic headword method to identify the subject-verb-object core structure of the sentence, identify verbs or verb-object phrases as 'management events', and identify time and conditional adverbs as 'triggering conditions'; finally, determine the 'execution order' between events by analyzing the relationships between sentences (e.g., conjunctions such as 'then', 'simultaneously'); for example, for the procedure clause "Step 5: Before lifting, it is necessary to confirm that the sling and lifting point are firmly connected, and a trial lift shall be performed, with a trial lift height not exceeding 0.5 meters", the program identifies "confirm that the sling and lifting point are firmly connected" and "perform a trial lift" as two serially executed management events, denoted as event M1 and event M2; it also identifies "before lifting" as "…". The trigger condition is defined as follows: For the procedure clause "Step Six: During the lifting process, the main operator and the signalman must maintain continuous communication, and at the same time, the monitor should closely observe the equipment status instruments," the program identifies "maintaining communication between the main operator and the signalman" and "the monitor observing the equipment status instruments" as two parallel management events, denoted as events M3 and M4. All identified management events are chained together according to their execution order in the procedure. Serial relationships, such as event M1 must be followed by event M2, are converted into logical AND gates, i.e., the probabilities of M1 and M2 are multiplied. Parallel relationships, such as events M3 and M4, are converted into logical OR gates, i.e., the probabilities of M3 and M4 are added. In this way, all clauses of the entire operating procedure are parsed and chained together, ultimately constructing a tree-like management process logic tree. The root node is "crane safety operation," and the leaf nodes are the most basic operating actions or inspection items, such as "checking ground flatness" and "confirming that the wind speed is below the permissible value."

[0075] S102: Obtain the equipment fault tree for this model of crawler crane. The top event of the fault tree is "crane boom instability," and the bottom fault events include "hydraulic oil leakage," "wire rope breakage," and "operator misjudgment of load," etc. Call the management process logic tree generated in S101 and extract all its leaf node events, such as leaf node event ME1: "Operator failed to check load weight as required." Simultaneously, extract all bottom fault events of the equipment fault tree, such as bottom fault event FE1: "Operator overloaded the crane." Next, calculate the semantic vector cosine similarity between ME1 and FE1. The first step of this calculation is to obtain the text descriptions of ME1 ("Operator failed to check load weight as required") and FE1 ("Operator overloaded the crane"). Input these two texts into a bidirectional encoder representation model based on the BERT-base-Chinese architecture. This model has been pre-trained on a general Chinese corpus. Through multi-layer attention mechanisms and feedforward network operations within the model, deep semantic features of the text are extracted, resulting in two 768-dimensional high-dimensional semantic vectors. and ;

[0076] For example, =[0.81,-0.23,…,0.45], =[0.79,-0.25,…,0.48]; Calculate the standard cosine similarity between these two vectors, specifically by calculating their vector dot product, for example, the result is 358.6, then calculate the magnitudes of the two vectors and multiply them, for example... =21.5, =22.1, the product of modulus and length is 475.15, then the standard cosine similarity is 358.6 / 475.15=0.7547; subsequently, the comprehensive weight coefficient is calculated; the hierarchical depth of leaf node event ME1 in the management process logic tree is analyzed, assuming its depth is 5 (the hierarchy is counted from the root node 0), and the hierarchy importance weight is determined. Through function Calculation, where For the level depth, then =1+0.1*5=1.5; The basis for setting this function is: the deeper the management event, the more specific its operation, and the more direct its impact on the final security.

[0077] The above function is set based on the fact that the deeper the management event, the more specific its operation and the more direct its impact on the final safety. This means that the form and parameters of the function are obtained by statistically analyzing 50 historical related safety accident reports, establishing the Spearman's rank correlation between the depth of the management event in the process tree and the severity of the accident consequences, and then performing linear regression fitting.

[0078] Further analysis of the contribution of the underlying failure event FE1 to the critical path in the equipment fault tree is performed. This is done by ranking the importance of all minimal cut sets containing FE1 to determine its contribution value. For example, if the failure rate of the path containing FE1 accounts for 25% of the total failure rate of the top event, then the path criticality weight is... The weight is set to 1.25; the basis for this weight setting is that the higher the contribution of a failure event, the greater its impact on systemic risk.

[0079] The above weight is set based on the principle that the higher the contribution of a fault event, the greater its impact on systemic risk. This means that the weight value is calculated based on the Fussell-Vesely importance index in fault tree analysis, specifically the normalized coefficient obtained by dividing the FV importance value of the underlying fault event by the average FV importance of all underlying events.

[0080] Multiply the hierarchy importance weight by the path criticality weight to obtain the comprehensive weight coefficient. Finally, the standard cosine similarity is multiplied by the comprehensive weighting coefficient to obtain the comprehensive correlation score. This score is no longer the standard cosine similarity, but a correlation metric that integrates semantic information and structural importance. A preset matching threshold of 0.9 is set. This threshold is determined by performing the above similarity calculation on 1000 pairs of labeled management event-fault event pairs, statistically analyzing the distribution of the calculation results, and selecting a value that achieves an accuracy of 99.0% and a recall of 96.5% for association judgment. Experimental data verifies that 0.9 is a value that achieves a good balance between accuracy and recall in association judgment. Since the calculated correlation score of 1.415 exceeds the threshold of 0.9, events ME1 and FE1 are judged as a strongly correlated event pair and added to the management event-fault event mapping set. This process is repeated by traversing all combinations of leaf node events and underlying fault events to finally generate a complete mapping set.

[0081] S103: Based on the management event-fault event mapping set generated in S102, the management process logic tree generated in S101 is expanded in topology. Specifically, under each leaf node event (such as ME1) of the management process logic tree, a logical AND gate is used to connect all strongly related underlying fault events (such as FE1) in the mapping set. In this way, all mapping relationships are added to the management process logic tree, thus forming a management-fault coupling logic tree with management logic at the top, equipment fault logic at the bottom, and management-fault coupling through mapping relationships in the middle.

[0082] Please see Figure 1 and Figure 3S2: Construct a digital twin state vector consisting of motor output power, hydraulic oil viscosity, and structural fatigue damage accumulation coefficient; analyze the defects in construction safety operation procedures to determine the failure rate improvement value; update the initial probability of the basic event nodes of the management-fault coupling logic tree; and calculate the initial risk probability.

[0083] The initial risk probability includes the risk baseline value, the static defect impact coefficient, and the top-level event identifier of the logic tree;

[0084] The specific steps for calculating the initial risk probability are as follows:

[0085] S201: Acquire the real-time output power of the equipment motor, the real-time viscosity of the hydraulic oil, and the cumulative coefficient of structural fatigue damage. Use the collected three parameter values ​​as components to construct a three-dimensional digital twin state vector. Simultaneously analyze all clauses of the construction safety operation procedure, identify two types of procedure defects: instruction logic conflict and missing pre-operation conditions. Based on the preset defect severity level mapping table, convert the identified procedure defects into quantified failure rate improvement values.

[0086] The specific steps for calculating the quantified failure rate improvement value are as follows:

[0087] The entire construction safety operation procedure is invoked, and the logical structure and dependencies of each clause are analyzed using a natural language processing model to identify all logical conflicts in instructions and defects such as missing prerequisite operating conditions.

[0088] Each identified procedural defect is assigned a defect type coefficient and a propagation impact factor, wherein the defect type coefficient for instruction logic conflict is higher than that for missing prerequisite operating conditions.

[0089] The quantified failure rate improvement value is calculated using the following formula:

[0090] ;

[0091] in, Representing the The quantified failure rate improvement value associated with each basic event node. Representative and the The baseline failure rate, derived from historical data statistics and associated with each basic event node. Representative and the The total number of procedural defects directly associated with each basic event node. From 1 to Integers used to identify procedural defects. Representing the The quantitative severity score corresponding to each procedure defect in the preset defect severity level mapping table. Representing the Defect type coefficient of the procedure defect, Representing the The propagation impact factors are determined by the position of the procedural defect in the management process logic tree;

[0092] Iterate through all procedural defects associated with a single basic event node, obtain the quantified severity level score, defect type coefficient, and propagation impact factor, perform product and summation operations to obtain the quantified failure rate increase value;

[0093] S202: Invoke the management-fault coupling logic tree, and based on the failure rate boost value, retrieve the basic event nodes directly associated with and transformed by the procedure defect in the management-fault coupling logic tree. Add the failure rate boost value to the initial probability of the retrieved basic event nodes one by one to establish the updated basic event probability set.

[0094] S203: Based on the overall topology of the management-fault coupling logic tree, the updated basic event probability set is called as the input probability of all basic event nodes. According to the logic AND gate multiplication and logic OR gate addition operation rules defined in the management-fault coupling logic tree, the probability of occurrence of multiple intermediate events is passed from bottom to top and calculated until the probability of occurrence of the top event is solved, and the initial risk probability is obtained.

[0095] S201: At the construction site of a highway bridge, sensors deployed on a crawler crane acquire real-time equipment operating parameters. Specifically, the real-time output power of the motor is collected as 260kW via the motor controller interface; the real-time viscosity of the hydraulic oil at the current operating temperature is collected as 42cSt via a viscosity sensor installed in the hydraulic circuit; and the cumulative fatigue damage coefficient of the structure is calculated to be 0.28 by combining strain gauge data attached to key stress points of the boom with the material's SN curve model. These three parameter values ​​are used as a component to construct a three-dimensional digital twin state vector. Simultaneously, the system synchronously calls up all clauses of the digitized "Safety Operation Procedures for Tracked Cranes," constructs a rule template library, uses regular expressions to match deontic logical keywords such as 'must,' 'prohibited,' and 'only if…' in the clauses, and combines dependency parsing to analyze the conditions and dependencies between clauses, thereby identifying logical conflicts in instructions and missing preconditions. The core function of this model is to identify the logical relationships between clauses; for example, the procedure contains clause A: "Before lifting a heavy object, the slewing brake should be in the braking position," and clause B: "The slewing brake can only be released after the heavy object has been lifted"; the model analysis finds that if an instruction requires a small-angle slewing during the lifting process, it constitutes a logical conflict with clauses A and B; another example is that the procedure requires "stopping work when the wind force exceeds level six," but this premise is not emphasized again in the "High-altitude Operations" section, so it is identified as a missing precondition; based on a preset defect severity level mapping table, each identified defect in the procedure is quantified;

[0096] Table 1. Severity Level Mapping Table for Procedure Defects:

[0097] ;

[0098] As shown in Table 1, the severity level S1 of the "instruction logic conflict" defect is 5, and the severity level S2 of the "missing prerequisite operation condition" defect is 3.

[0099] S202: Invoke the management event-failure event mapping set generated in S102. For each procedural defect identified in S201, find the management event corresponding to the procedural clause it belongs to, and further find the underlying failure event associated with the management event through the mapping set. For example, defect 1 "instruction logic conflict" occurs in clauses A and B, which correspond to the management event "turning brake operation". This management event is associated with the underlying failure event "operator misjudgment of load" through the mapping set. Subsequently, based on the failure rate increase value calculation method defined in S201, calculate the increment of the procedural defect on the base probability of the associated underlying failure event. In this embodiment, the original base probability of the underlying failure event "operator misjudgment of load" is... for This incident relates to defect 1 "instruction logic conflict" and defect 2 "missing prerequisite conditions"; the severity level of defect 1 is... Its type coefficient is 5. The propagation impact factor is set at 1.1. The severity level of defect 2 is 0.8. Its type coefficient is 3. Set to 1.0, propagation impact factor. The probability increment is 0.6; the effects of the two defects are summed to calculate the total probability increment. The calculation method is as follows: The calculation results here This is the increment in the base probability caused by procedural defects; adding this increment to the original base probability yields the updated base event probability: The result This is the updated base probability of the "operator misjudgment of load" event after taking into account the impact of procedural defects; this calculation is repeated for all underlying fault events affected by procedural defects to generate an updated base event probability set;

[0100] The above-mentioned factors affecting transmission The value is 0.8. This value is calculated based on the defect's position in the management process logic tree. The closer the defect is to a leaf node, the more direct its impact, and the higher the factor value. The specific calculation method is as follows: the propagation impact factor is calculated using the formula... The calculation yielded the result; among which, Represents the dissemination impact factor. Represents the base of the natural logarithm. This represents the impact attenuation coefficient, calibrated based on historical data; in this embodiment, it is set to 0.2. This represents the path length between the management event node where the defect is located and its most directly related leaf node event, expressed in units of nodes; for defect 1, its path length is 1. Take one decimal place as 0.8.

[0101] S203: Based on the complete topology of the management-fault coupling logic tree, the updated basic event probability set generated in S202 is invoked; each probability value in this probability set is used as the input probability of the corresponding basic event node; subsequently, according to the logic gate operation rules defined in the tree, probability propagation and calculation are performed layer by layer from bottom to top; for example, a logic AND gate connects two basic events A and B, and their updated probabilities are respectively and Then the probability of the output event C of the AND gate occurring is... A logical OR gate connects events D and E, with probabilities respectively. and The probability of the output event F of the OR gate occurring is... This calculation process iterates upwards along the tree structure until the final probability of the top event "crane boom instability" is calculated; assuming the final calculated value is... This value is the initial risk probability; this probability value is compared with the risk benchmark value (e.g., the value specified in industry safety standards). The calculated value is compared with the calculated value. It is a preset risk benchmark value The value is 35.8 times higher than the standard, falling within the 'high-risk' range as defined by safety management regulations (greater than...). This indicates a high initial risk in the current construction plan and equipment status. This initial risk probability, along with the top-level event identifier "crane boom instability" in its logic tree, is recorded.

[0102] Please see Figure 1 and Figure 4 S3: Parse real-time management instructions to extract instruction parameters, calculate dynamic adjustment rate to update digital twin state vector, and generate adjusted digital twin state vector;

[0103] The adjusted digital twin state vector includes the instantaneous motor power prediction, the hydraulic oil viscosity after dynamic decay, and the structural fatigue damage coefficient after accelerated accumulation.

[0104] The specific steps for obtaining the adjusted digital twin state vector are as follows:

[0105] S301: Monitors real-time management command data stream, performs text serialization parsing on received real-time management commands, identifies command type keywords, core operation object identifiers and quantized values, and performs structured aggregation of command type keywords, core operation object identifiers and quantized values ​​to generate command parameter sets;

[0106] S302: Extract the core operation object identifier and quantized value from the instruction parameter set, and match the corresponding weight coefficient for each core operation object identifier according to the preset parameter influence weight lookup table. After multiplying all quantized values ​​and corresponding weight coefficients, perform an accumulation and summation operation to obtain the dynamic adjustment rate of the state vector.

[0107] The specific steps for calculating the dynamic adjustment rate of the state vector are as follows:

[0108] Obtain the instruction parameter set, parse the instruction parameter set to extract all core operation object identifiers, quantized values ​​and corresponding instruction timestamps;

[0109] The preset parameter influence weight lookup table is invoked to obtain the basic weight value based on the identifier of each core operation object, and the time decay coefficient is calculated based on the time difference between the instruction timestamp and the current prediction time.

[0110] Multiply the base weight value by the time decay coefficient to obtain the dynamic weight coefficient, and calculate the dynamic adjustment rate of the state vector according to the following formula:

[0111] ;

[0112] in, Represents the dynamic adjustment rate of the state vector. This represents the total number of instructions included in the instruction parameter set. From 1 to An integer used to identify the instruction. Representing the The quantization value corresponding to each instruction. Representing the The core operation object identifier of the instruction is used to obtain the basic weight value. The timestamp representing the current prediction time point, Representing the The timestamp of the instruction This represents the preset time decay constant used to control the decay rate;

[0113] Substitute the quantized values ​​of all instructions, the basic weight values, the timestamp of the current prediction time, the timestamp of each instruction, and the time decay constant into the formula, perform cumulative summation and averaging, and generate the dynamic adjustment rate of the state vector.

[0114] S303: Call the three-dimensional digital twin state vector, and for the dynamic adjustment rate of the state vector, perform element-wise multiplication of the three components of the three-dimensional digital twin state vector—motor output power, hydraulic oil viscosity, and structural fatigue damage accumulation coefficient—with the dynamic adjustment rate of the state vector to obtain the multi-component adjustment increment. Then, perform matrix summation of the multi-component adjustment increment with the original corresponding component values ​​of the three-dimensional digital twin state vector to generate the adjusted digital twin state vector.

[0115] S301: The system continuously monitors the real-time management command data stream issued by the construction site through the wireless intercom system or dispatch platform; at a certain moment, it receives a voice command: "Hoist beam No. 3, increase the counterweight to 80 tons, increase the boom speed, adjust to 2 degrees per second"; the system performs text serialization parsing on this voice command, converting it into text: "Hoist beam No. 3, increase the counterweight to 80 tons, adjust the boom speed to 2 degrees per second"; then, by loading a predefined 'core operation object' dictionary (containing 'counterweight', 'boom speed', etc.), it uses the maximum forward matching algorithm to identify the core operation object; subsequently, it utilizes... The named entity recognition model based on Conditional Random Field (CRF) is trained using annotated command corpus to identify 'quantified values' associated with the operation object. For example, it identifies command type keywords such as "lifting", "increase", and "adjust", core operation object identifiers such as "beam No. 3", "counterweight", and "lifting speed", and quantified values ​​such as "80" and "2". The identified information is then aggregated in a structured manner to generate a command parameter set containing multiple command parameters, for example: {command 1:{object:'counterweight', value:80}, command 2:{object:'lifting speed', value:2}}.

[0116] S302: Extract the core operation object identifiers ('counterweight', 'arm lifting speed') and quantified values ​​(80, 2) from the instruction parameter set generated by S301; the system calls a preset parameter influence weight lookup table to match the corresponding weight coefficient for each core operation object identifier;

[0117] Table 2. Parameter Influence Weight Lookup Table:

[0118] ;

[0119] As shown in Table 2, the weighting coefficients of 'counterweight' Weighting coefficient for 'lifting speed' The quantified value of each core operation object is compared with its corresponding rated equipment parameter value to calculate the normalized command strength. For example, if the rated counterweight of this crane model is 100 tons and the rated boom speed is 1.5 degrees per second, the command strength is calculated as follows: ;

[0120] The dynamic adjustment rate of the state vector is calculated according to the formula defined in S302, and the normalized command strength is used. As a quantized value in the formula .

[0121] The calculation formula is: ;

[0122] Among them, the total number of instructions Assume both instructions were just issued, i.e., the time difference is... Time decay constant .

[0123] The calculation is as follows: ;

[0124] ;

[0125] ;

[0126] ;

[0127] Therefore, the calculated result of 0.506 is the dynamic adjustment rate of the state vector calculated after combining the two real-time instructions received at the moment.

[0128] S303: Call the 3D digital twin state vector constructed in S201 The dynamic adjustment rate of the state vector calculated by S302 is compared with this vector. Perform calculations to obtain the adjusted digital twin state vector. To achieve differentiated adjustments for different components, a directional adjustment coefficient vector is introduced. ,in Corresponding motor output power, Corresponding to hydraulic oil viscosity, Corresponding structural fatigue damage accumulation coefficient; adjusted digital twin state vector The components are obtained through the formula The calculation integrates the original state, dynamic adjustment rate, and the adjustment direction and weight of each component; the calculation process is as follows: Instantaneous motor power prediction value = Hydraulic oil viscosity after dynamic decay = The coefficient of structural fatigue damage after accelerated accumulation = Therefore, the generated adjusted digital twin state vector is The components of this vector represent the instantaneous predicted motor power, the hydraulic oil viscosity after dynamic decay, and the structural fatigue damage coefficient after accelerated accumulation, respectively.

[0129] Please see Figure 1 and Figure 5 S4: Based on the management-fault coupling logic tree, the initial risk probability and the adjusted digital twin state vector, the instantaneous occurrence probability of the basic events of the management-fault coupling logic tree is updated by Bayesian network at discrete time steps, the top-level event probabilities are connected, and the risk probability evolution sequence is generated.

[0130] The risk probability evolution sequence includes discrete time step nodes and instantaneous failure probability values;

[0131] The specific steps for obtaining the risk probability evolution sequence are as follows:

[0132] S401: Obtain the discrete time step, based on the management-fault coupling logic tree, the initial risk probability and the adjusted digital twin state vector, take multiple components of the adjusted digital twin state vector as Bayesian network evidence nodes, set the probability values ​​of the associated basic events in the initial risk probability as prior probabilities, perform conditional probability inference, update the posterior probability of each basic event in the management-fault coupling logic tree, and obtain the instantaneous basic event probability set;

[0133] S402: Invoke the topology of the management-fault coupling logic tree, and based on the instantaneous basic event probability set, perform probability multiplication operation on the logical AND gates in the management-fault coupling logic tree, perform probability addition operation on the logical OR gates, and iterate from bottom to top to calculate the occurrence probability of all intermediate events until the top event, and obtain the single-step top event probability.

[0134] S403: For a preset number of consecutive discrete time steps, repeat the aforementioned process of updating the posterior probability and solving the top event occurrence probability. Arrange and combine the single-step top event probabilities generated in each discrete time step according to the chronological order to establish a risk probability evolution sequence.

[0135] S401: Set a discrete time step, for example, 10 seconds; based on the management-fault coupling logic tree generated in S1, the initial risk probability calculated in S2 (of which the basic event probability set is), and the adjusted digital twin state vector generated in S3. Bayesian network inference is performed; the three components of the adjusted digital twin state vector are used as evidence nodes of the Bayesian network, that is, evidence is defined. The values ​​are: "Motor power = 391.56kW", "Hydraulic oil viscosity = 31.374cSt", and "Fatigue damage coefficient = 0.4500". The updated base event probability set calculated in S202 (e.g., the probability of "operator misjudging load") is... ) as the prior probability of the corresponding basic event node in the network Update the posterior probability of each basic event by performing conditional probability inference. This reasoning process relies on the network's conditional probability table (CPT), which defines the dependencies between evidence and underlying events.

[0136] The reasoning process described above relies on the network's Conditional Probability Table (CPT), which defines the dependency between evidence and basic events. Specifically, the network's CPT is constructed through statistical learning of the equipment's historical operating data (including sensor readings and fault records). The sensor data is discretized into several intervals (e.g., viscosity is divided into high, medium, and low levels), and then the frequency of each basic fault event is statistically analyzed in different data intervals to obtain the conditional probability distribution.

[0137] For example, for the basic event H "hydraulic pump wear", its CPT defines When the evidence "hydraulic oil viscosity = 20.748 cSt" (belonging to the "low" range) is input, the system calculates based on CPT. Its posterior probability is based on Bayes' theorem. Calculations are performed; taking the basic event H'hydraulic pump wear' as an example, its prior probability is... From the constructed CPT, it was found that under conditions of hydraulic pump wear, the probability of low viscosity is [missing information]. Total probability Historical data shows that the viscosity is lower when the hydraulic pump is not worn. , ;therefore Then the posterior probability Perform post-abortive probability calculations on all basic events associated with evidence nodes to obtain a novel set of instantaneous basic event probabilities that reflects the impact of the current instantaneous state of the device;

[0138] S402: The instantaneous basic event probability set generated in S401 is used as the update input probability of the corresponding basic event node in the management-fault coupling logic tree; again, according to the logic gate operation rules defined in the tree, the probability is passed and calculated layer by layer from bottom to top, finally obtaining a single-step top event probability of the top event "crane boom instability"; assuming the calculated value is... This value represents the instantaneous risk probability at the current time step, taking into account both the device status and the impact of management commands.

[0139] S403: The system repeatedly executes the complete process of S401 and S402 according to a preset series of discrete time steps (e.g., predicting the next 5 minutes, i.e., 30 time steps of 10 seconds each). At the beginning of each time step, the system receives a new real-time management instruction and performs the calculation of S3 accordingly to obtain the adjusted digital twin state vector for that step. Then, this vector is used as new evidence to perform a Bayesian posterior probability update, and the single-step top event probability for that step is calculated through the logic tree topology structure; for example, in the second time step ( Upon receiving the instruction "slewing deceleration", the adjusted digital twin state vector becomes... The calculated probability of a single-step top event may decrease slightly to ; at the third time step ( No new instructions were given, but the continuous operation of the equipment led to fatigue accumulation, and the state vector changed. The calculated probability rose to The probability values ​​of each discrete time step and its corresponding single-step top event are arranged and combined in chronological order to ultimately establish a risk probability evolution sequence; the specific form of this sequence is a time series array. .

[0140] A highway construction risk prediction system based on big data, used to execute the aforementioned highway construction risk prediction method based on big data, the system comprising:

[0141] The coupling logic tree construction module is used to obtain the construction safety operation procedures and equipment fault tree, parse the construction safety operation procedures to construct the management process logic tree, associate the management process logic tree with the equipment fault tree, calculate the management-fault coupling logic tree, and pass it to the initial risk quantification module.

[0142] The initial risk quantification module is used to construct a digital twin state vector consisting of motor output power, hydraulic oil viscosity, and structural fatigue damage accumulation coefficient. It analyzes the defects in construction safety operation procedures to determine the failure rate improvement value, updates the initial probability of the basic event nodes of the management-fault coupling logic tree, calculates the initial risk probability, and passes the digital twin state vector to the twin state dynamic update module.

[0143] The twin state dynamic update module is used to parse real-time management instructions to extract instruction parameters, calculate the dynamic adjustment rate to update the digital twin state vector, generate the adjusted digital twin state vector, and pass it to the risk evolution sequence generation module.

[0144] The risk evolution sequence generation module is used to update the instantaneous occurrence probability of the basic events in the management-fault coupling logic tree using a Bayesian network at discrete time steps, by using the management-fault coupling logic tree, the initial risk probability, and the adjusted digital twin state vector, and then connecting the top event probability to generate a risk probability evolution sequence.

[0145] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the scope of protection defined by the claims of the present invention.

Claims

1. A big data-based road construction risk prediction method, characterized by, The method comprises the following steps: S1: obtaining a construction safety operation regulation and a device fault tree, analyzing the construction safety operation regulation to construct a management process logical tree, associating the management process logical tree with the device fault tree, and calculating a management-fault coupled logical tree; S2: constructing a digital twin state vector composed of motor output power, hydraulic oil viscosity, and structure fatigue damage accumulation coefficient, analyzing defects of the construction safety operation regulation to determine failure rate improvement value, updating initial probability of a basic event node of the management-fault coupled logical tree, and calculating initial risk probability; The calculation step of the initial risk probability is specifically as follows: S201: obtaining real-time motor output power, real-time hydraulic oil viscosity, and structure fatigue damage accumulation coefficient, constructing a three-dimensional digital twin state vector by taking the three parameter values as components, synchronously analyzing all clauses of the construction safety operation regulation, identifying two types of regulation defects of instruction logic conflict and missing pre-operation conditions, converting the identified regulation defects into quantified failure rate improvement value according to a preset defect severity level mapping table; S202: calling the management-fault coupled logical tree, searching for a basic event node directly associated and converted by the regulation defect in the management-fault coupled logical tree according to the failure rate improvement value, performing item-by-item addition operation on the failure rate improvement value and the initial probability of the searched basic event node, and establishing an updated basic event probability set; S203: based on the overall topology structure of the management-fault coupled logical tree, calling the updated basic event probability set as input probability of all the basic event nodes, according to the logic and gate multiplication and the logic or gate addition operation rules defined in the management-fault coupled logical tree, performing layer-by-layer transmission and calculation of a plurality of intermediate event occurrence probabilities from bottom to top until the solution of the top event occurrence probability is completed, and obtaining the initial risk probability; The calculation step of the quantified failure rate improvement value is specifically as follows: calling all clauses of the construction safety operation regulation, analyzing the logical structure and dependency relationship of each clause by using a natural language processing model, and identifying all instruction logic conflicts and pre-operation condition missing defects; allocating a defect type coefficient and a propagation influence factor to each identified regulation defect, wherein the defect type coefficient of the instruction logic conflict is higher than the defect type coefficient of the pre-operation condition missing; calculating the quantified failure rate improvement value according to the following formula: ； wherein, a quantized failure rate increase value associated with the kth basic event node, a reference failure rate associated with the kth basic event node statistically derived from historical data, a total number of procedure defects directly associated with the kth basic event node, j is an integer from 1 to a procedure defect identifier, a quantized severity level score corresponding to the jth procedure defect in the pre-defined defect severity level mapping table, a defect type coefficient of the jth procedure defect, a propagation impact factor determined by the position of the jth procedure defect in the management flow logic tree. traversing all the regulation defects associated with a single basic event node, obtaining the quantified severity level score, the defect type coefficient, and the propagation influence factor, performing product and accumulation sum operation, and obtaining the quantified failure rate improvement value; S3: analyzing real-time management instructions to extract instruction parameters, calculating a dynamic adjustment rate to update the digital twin state vector, and generating an adjusted digital twin state vector; The obtaining step of the adjusted digital twin state vector is specifically as follows: S301: Monitor the real-time management instruction data stream, text sequence analysis on the received real-time management instruction, identify instruction type keywords, core operation object identification and quantitative values, and structure aggregation of the instruction type keywords, the core operation object identification and the quantitative values, to generate an instruction parameter set; S302: Extract the core operation object identification and the quantitative values in the instruction parameter set, and match a corresponding weight coefficient for each core operation object identification according to a preset parameter influence weight lookup table. After the product operation of all the quantitative values and the corresponding weight coefficients, the sum operation is performed to obtain a state vector dynamic adjustment rate; S303: Call the three-dimensional digital twin state vector, and for the state vector dynamic adjustment rate, perform element-by-element multiplication operation on the motor output power, hydraulic oil viscosity and structure fatigue damage accumulation coefficient of the three-dimensional digital twin state vector to obtain a multi-component adjustment increment, and then perform matrix addition on the multi-component adjustment increment and the original corresponding component value of the three-dimensional digital twin state vector to generate an adjusted digital twin state vector; The calculation steps of the state vector dynamic adjustment rate are as follows: Obtain the instruction parameter set, analyze the instruction parameter set to extract all the core operation object identifications, the quantitative values and the corresponding instruction time stamps; Call the preset parameter influence weight lookup table, query the basic weight value according to each core operation object identification, and calculate the time decay coefficient based on the time difference between the instruction time stamp and the current prediction time point; Multiply the basic weight value and the time decay coefficient to obtain a dynamic weight coefficient, and calculate the state vector dynamic adjustment rate according to the following formula: ； wherein, represents a dynamic adjustment rate of the state vector, N represents a total number of instructions included in the instruction parameter set, i is an integer from 1 to N, and is used to identify an instruction, represents the quantized value corresponding to the i th instruction, represents the basic weight value obtained by querying the core operation object identifier of the i th instruction, represents the timestamp of the current prediction time point, represents the timestamp of the i th instruction, represents a preset time attenuation constant used to control the attenuation speed; Substitute all the quantitative values of the instructions, the basic weight values, the time stamps of the current prediction time point, the time stamps of each instruction and the time decay constant into the formula, perform the sum operation and the average processing to generate the state vector dynamic adjustment rate; S4: Based on the management-failure coupled logic tree, the initial risk probability and the adjusted digital twin state vector, update the instantaneous occurrence probability of the management-failure coupled logic tree basic event at discrete time steps using the Bayesian network, connect the top event probability, and generate a risk probability evolution sequence.

2. The big data based highway construction risk prediction method of claim 1, wherein, The management-failure coupled logic tree includes management failure event nodes, physical failure event nodes and cross-domain causal correlation paths. The initial risk probability includes a risk reference value, a static defect influence coefficient and a logic tree top event identification. The adjusted digital twin state vector includes an instantaneous motor power prediction value, a dynamically decayed hydraulic oil viscosity and an accelerated accumulated structure fatigue damage coefficient. The risk probability evolution sequence includes discrete time step nodes and instantaneous failure probability values.

3. The big data based highway construction risk prediction method of claim 2, wherein, The calculation steps of the management-failure coupled logic tree are as follows: S101: Obtain a construction safety operation regulation, perform structured analysis on text content of the construction safety operation regulation, identify management events, trigger conditions and execution sequences, convert the trigger conditions into logical AND gates, convert parallel execution relationships into logical OR gates, concatenate all the management events, and establish a management flow logic tree; S102: Obtain a device fault tree, call the management flow logic tree, extract text descriptions of multiple leaf node events in the management flow logic tree and multiple bottom layer fault event in the device fault tree respectively, calculate semantic vector cosine similarity between the leaf node events and the bottom layer fault events, associate event pairs with cosine similarity exceeding a preset matching threshold, and generate a management event-fault event mapping set; S103: According to the management event-fault event mapping set, traverse the device fault tree, when it is found that the multiple bottom layer fault events exist in the management event-fault event mapping set, call the management flow logic tree or sub-tree structure associated with the multiple bottom layer fault events, replace the nodes corresponding to the multiple bottom layer fault events in the device fault tree, complete tree structure coupling, and obtain a management-fault coupled logic tree.

4. The big data based highway construction risk prediction method of claim 1, wherein, The risk probability evolution sequence acquisition step is specifically: S401: Obtain a discrete time step, based on the management-fault coupled logic tree, the initial risk probability and the adjusted digital twin state vector, take multiple components of the adjusted digital twin state vector as Bayesian network evidence nodes, set the probability values of the associated basic events in the initial risk probability as prior probabilities, perform conditional probability reasoning, update the posterior probability of each basic event in the management-fault coupled logic tree, and obtain an instantaneous basic event probability set; S402: Call the topology structure of the management-fault coupled logic tree, and according to the instantaneous basic event probability set, perform probability multiplication operation for the logical AND gates in the management-fault coupled logic tree, perform probability addition operation for the logical OR gates, iteratively calculate all intermediate events from bottom to top until the occurrence probability of the top event, and obtain a single-step top event probability; S403: For a preset plurality of consecutive discrete time steps, repeat the posterior probability updating and top event occurrence probability solving processes, arrange and combine the single-step top event probabilities generated by each discrete time step in chronological order, and establish the risk probability evolution sequence.

5. The big data based highway construction risk prediction method of claim 3, wherein, The semantic vector cosine similarity calculation step is specifically: Obtain the text descriptions of the leaf node events of the management flow logic tree and the bottom layer fault events of the device fault tree; Call a pre-trained bidirectional encoder representation model, input the text descriptions of the leaf node events and the text descriptions of the bottom layer fault events into the pre-trained bidirectional encoder representation model respectively, perform deep semantic feature extraction, and obtain high-dimensional semantic vectors corresponding to the text descriptions; For each of the event pairs, extract the two corresponding high-dimensional semantic vectors, and calculate the standard cosine similarity between the two high-dimensional semantic vectors through vector dot product operation and vector module length multiplication operation; At the same time, analyze the hierarchical depth of the leaf node event in the management process logic tree and the critical path contribution of the underlying fault event in the device fault tree, and calculate the hierarchical importance weight and path criticality weight respectively; Multiply the hierarchical importance weight and the path criticality weight to obtain a comprehensive weight coefficient, and multiply the standard cosine similarity and the comprehensive weight coefficient to obtain the semantic vector cosine similarity. 6.A big data based highway construction risk prediction system, characterized in that, The system comprises: A coupled logic tree construction module is configured to obtain a construction safety operation regulation and a device fault tree, analyze the construction safety operation regulation to construct a management process logic tree, associate the management process logic tree with the device fault tree, calculate a management-fault coupled logic tree, and pass the management-fault coupled logic tree to an initial risk quantification module; An initial risk quantification module is configured to construct a digital twin state vector composed of motor output power, hydraulic oil viscosity, and structural fatigue damage accumulation coefficient, analyze defects in the construction safety operation regulation to determine an increase in failure rate, update the initial probability of a basic event node in the management-fault coupled logic tree, calculate an initial risk probability, and pass the digital twin state vector to a twin state dynamic updating module; The calculation of the initial risk probability comprises the following steps: Obtain real-time motor output power, real-time hydraulic oil viscosity, and structural fatigue damage accumulation coefficient, use the three parameters as components to construct a three-dimensional digital twin state vector, and simultaneously analyze all provisions of the construction safety operation regulation to identify two types of regulation defects, i.e., instruction logic conflicts and missing pre-operation conditions, and convert the identified regulation defects into quantified failure rate increases according to a pre-set defect severity level mapping table; Call the management-fault coupled logic tree, search for a basic event node directly associated with the regulation defects in the management-fault coupled logic tree according to the failure rate increase, perform item-by-item addition operation on the failure rate increase and the initial probability of the searched basic event node to establish an updated basic event probability set, and call the updated basic event probability set as the input probability of all basic event nodes based on the overall topology of the management-fault coupled logic tree, and perform multiplication and addition operations on the logic and gate and the logic or gate defined in the management-fault coupled logic tree from bottom to top layer by layer to calculate a plurality of intermediate event occurrence probabilities until the top event occurrence probability is solved to obtain the initial risk probability. The calculation of the quantified failure rate increase comprises the following steps: Call all provisions of the construction safety operation regulation, analyze the logical structure and dependency of each provision using a natural language processing model, and identify all instruction logic conflicts and pre-operation condition missing defects. ​ assigning a defect type coefficient and a propagation impact factor to each identified procedure defect, wherein the defect type coefficient of the instruction logic conflict is higher than the defect type coefficient of the missing pre-operation condition; calculating the quantified failure rate improvement value according to the following formula: ； wherein, a quantized failure rate increase value associated with the kth basic event node, a baseline failure rate associated with the kth basic event node statistically derived from historical data, a total number of procedure defects directly associated with the kth basic event node, j is an integer from 1 to identifying a procedure defect, a quantized severity level score corresponding to the jth procedure defect in the pre-defined defect severity level mapping table, a defect type coefficient of the jth procedure defect, a propagation impact factor determined by the position of the jth procedure defect in the management flow logic tree. traversing all the procedure defects associated with a single base event node, obtaining the quantified severity level score, the defect type coefficient and the propagation impact factor, performing a product and accumulation summation operation to obtain the quantized failure rate improvement value; a twin state dynamic updating module for parsing real-time management instructions to extract instruction parameters, calculating a dynamic adjustment rate to update the digital twin state vector, generating an adjusted digital twin state vector, and passing it to the risk evolution sequence generation module; the step of obtaining the adjusted digital twin state vector is specifically: monitoring real-time management instruction data streams, performing text serialization analysis on received real-time management instructions, identifying instruction type keywords, core operation object identifiers and quantitative values, and structuring and aggregating the instruction type keywords, core operation object identifiers and quantitative values to generate an instruction parameter set; extracting the core operation object identifiers and quantitative values in the instruction parameter set, and matching corresponding weight coefficients for each core operation object identifier according to a preset parameter influence weight lookup table, performing a product operation on all quantitative values and corresponding weight coefficients, and then performing an accumulation summation operation to obtain a state vector dynamic adjustment rate; S303: calling the three-dimensional digital twin state vector, and performing element-by-element multiplication operation on the motor output power, hydraulic oil viscosity and structure fatigue damage accumulation coefficient of the three components of the three-dimensional digital twin state vector and the state vector dynamic adjustment rate for the state vector dynamic adjustment rate, to obtain a multi-component adjustment increment, and then performing matrix addition on the multi-component adjustment increment and the original corresponding component values of the three-dimensional digital twin state vector to generate an adjusted digital twin state vector; the step of calculating the state vector dynamic adjustment rate is specifically: obtaining the instruction parameter set, parsing the instruction parameter set to extract all core operation object identifiers, quantitative values and corresponding instruction time stamps; calling the preset parameter influence weight lookup table to query and obtain a basic weight value for each core operation object identifier, and calculating a time decay coefficient based on the time difference between the instruction time stamp and the current prediction time point; multiplying the basic weight value and the time decay coefficient to obtain a dynamic weight coefficient, and calculating the state vector dynamic adjustment rate according to the following formula: ； wherein, represents a dynamic adjustment rate of the state vector, N represents a total number of instructions included in the instruction parameter set, i is an integer from 1 to N, and is used to identify an instruction, represents the quantized value corresponding to the i th instruction, represents the basic weight value obtained by querying the core operation object identifier of the i th instruction, represents the timestamp of the current prediction time point, represents the timestamp of the i th instruction, represents a preset time attenuation constant used to control the attenuation speed; substituting all the quantitative values of the instructions, the basic weight values, the time stamps of the current prediction time point, the time stamps of each instruction and the time decay constant into the formula, performing accumulation summation and averaging processing to generate the state vector dynamic adjustment rate; a risk evolution sequence generation module for generating a risk probability evolution sequence by updating the instantaneous occurrence probability of the management-failure coupled logic tree base events with a Bayesian network over discrete time steps through the management-failure coupled logic tree, the initial risk probability, and the adjusted digital twin state vector, concatenating top-level event probabilities.

Images