A large-span pc continuous rigid frame bridge construction risk causal level modeling method based on visual perception

Abstract

The application discloses a large-span PC continuous rigid frame bridge construction risk causal level modeling method based on visual perception, and relates to the technical field of bridge construction risk control. Risk elements, frequency, improved YOLOv8 structure, model training and inference output, and risk quantitative mapping of visual results are provided. Index matrix construction, range standardization, and CRITIC weight calculation are provided. Direct influence definition and direct influence matrix normalization are provided. Total relationship matrix construction, influence degree and affected degree calculation, significant relationship screening, and adjacent matrix construction are provided. Reachable matrix acquisition and level division are provided. Comprehensive priority calculation and path risk degree calculation are provided. Through the combination of the YOLOv8 model and dynamic convolution and multi-branch attention mechanism, the detection accuracy of different risk elements in a dynamic construction scene is improved, and through accurate calculation and adaptive weight adjustment, more accurate data support is provided for risk assessment.

Description

Technical Field

[0001] This invention relates to the field of bridge construction risk management technology, specifically a causal hierarchical modeling method for construction risks of long-span PC continuous rigid frame bridges that integrates visual perception. Background Technology

[0002] As infrastructure construction expands, construction risk management becomes increasingly important, especially in the construction of long-span PC continuous rigid frame bridges. Traditional risk assessment methods rely on expert experience and historical data, which, while providing some reference for decision-making, also have the following problems: First, over-reliance on expert judgment makes them susceptible to the influence of personal experience and emotions, leading to biased assessment results; second, risk assessment is usually completed during the construction planning stage, making it difficult to address risks that change in real time during construction; and finally, with the growth of data volume, traditional methods struggle to handle massive datasets, resulting in the failure to detect potential risks in a timely manner.

[0003] To address these issues, the DEMATEL-ISM method has been introduced into construction risk assessment in recent years. This method identifies and quantifies the interactions between risk factors by constructing a causal matrix and conducting expert evaluation. However, this method still has limitations, relying on expert evaluation and lacking real-time capability when dealing with dynamic changes in the construction environment.

[0004] With the development of computer vision and artificial intelligence, risk assessment based on image data is gradually becoming an emerging technology. The YOLOv8 model performs excellently in real-time object detection and can be used for risk element detection at construction sites (such as unprotected heights and overloaded hoisting). While this method can provide real-time risk identification, it also faces the challenge of image recognition accuracy being affected by environmental changes. More importantly, image recognition cannot comprehensively cover all risk factors at construction sites, such as environmental changes and equipment conditions.

[0005] Therefore, current risk assessment methods either rely too much on expert experience or focus on image recognition, lack adaptive fusion of subjective and objective weights, and fail to make full use of multiple data sources. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention provides a causal hierarchical modeling method for construction risks of long-span PC continuous rigid frame bridges that integrates visual perception. By combining the YOLOv8 model with dynamic convolution and multi-branch attention mechanisms, it improves the detection accuracy of different risk elements in dynamic construction scenarios. Through precise calculation and adaptive weight adjustment, it provides more accurate data support for risk assessment.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a causal hierarchy modeling method for construction risks of long-span PC continuous rigid frame bridges that integrates visual perception, comprising the following steps:

[0008] Step 1: Visual perception and quantification of risk elements;

[0009] ①Establish a risk element set Risk element , Risk factors in the predetermined actual construction scenarios are acquired through visual perception, and their frequency of occurrence is statistically analyzed and calculated within the observation period. , Indicates the number of frames representing the risk element. Indicates the total number of frames;

[0010] ② Introduce dynamic convolution and three-branch attention in the backbone and detector head of YOLOv8;

[0011] ③ During the training phase, the model uses the standard YOLOv8 combined loss function; during the inference phase, YOLOv8 outputs the class confidence score for each candidate box. and pixel area ratio ;

[0012] ④ Class confidence Mapped to probability of occurrence The risk intensity is defined by combining frequency and scope of impact. , The slope parameter;

[0013] Step Two: Indicator Matrix and CRITIC Objective Weights;

[0014] ①Construction OK Column index matrix Matrix elements Indicates the first Indicators of risk elements in each sample, including risk intensity. Frequency and pixel area ratio , For sample index;

[0015] ② For the index matrix Perform range standardization to obtain matrix elements. Standardized value , ;

[0016] ③ Calculate the contrast intensity of CRITIC , Indicates the first The standard deviation of each risk element indicator Other risk metrics used to represent correlation comparisons The correlation coefficient is represented, and then normalized to obtain the weights of the risk element indicators. , And the sum of the weights of all risk element indicators equals 1;

[0017] Step 3: Construct the direct influence matrix;

[0018] ① In the time window Inside, if risk element After it happened, There is another risk element in memory. If it is more likely to occur, then the risk element is considered to be... For another risk element It had a direct impact, and the statistics of each risk element were analyzed. After it happened Has at least one risk element appeared inside the window? , For the first Sub-risk element The time of occurrence is represented as follows:

[0019]

[0020] In the formula, Indicates risk element Starting from the moment of occurrence, examine the subsequent length. The window, if at least one risk element occurs. Then count 1 and apply it to all risk elements. Summation occurs. This indicates an indicator function; the value is 1 if the condition within the parentheses is true, and 0 otherwise. Indicates risk element The set of times of occurrence, Indicates the time of occurrence;

[0021] This leads to the conditional probability of the unbiased frequency estimate. And define the elements that directly affect the matrix. , Risk element The intensity of the risk, For all risk elements The highest risk intensity value, This is the scaling factor;

[0022] ② Normalize the matrix by the maximum row sum to obtain the normalized direct influence matrix. ;

[0023] Step 4: Causal analysis and screening of significant relationships;

[0024] ① Based on the normalized direct influence matrix Calculate the overall relation matrix , It is the identity matrix;

[0025] ② Through the elements of the overall relation matrix Calculate the impact of each risk element and degree of impact , and The first Line number Column elements and the first Line number Column elements;

[0026] ③ Use evaluation thresholds To filter out significant causal relationships and obtain the elements of the filtered overall relationship matrix. , Total relation matrix The mean, To filter parameters, Total relation matrix standard deviation For indicator functions;

[0027] ④ Obtain the adjacency matrix based on the filtering results. ;

[0028] Step 5: Hierarchical decomposition;

[0029] ① For the adjacency matrix Perform a transitive closure operation to obtain the reachability matrix. ;

[0030] ② Through the reachability matrix The elements in determine the risk element In the hierarchical partitioning, the level at which the set of reachable sets is defined. With the Preview , and Represent the reachability matrix respectively The first in Line number The elements of the column and the first Line number The elements of the column are stripped to obtain the first... Risk element set at the causal level Repeat the hierarchical decomposition until all risk elements are layered.

[0031] Step Six: Prioritizing Key Risks and Identifying Propagation Paths;

[0032] ① Define the overall priority , The larger the value, the earlier the position in the same causal hierarchy;

[0033] ② Define path risk level , For a directed path, For path length, For elements The maximum value.

[0034] Furthermore, in step 1, ②, the dynamic convolution output is represented as:

[0035]

[0036]

[0037] In the formula, This represents the output feature map. Indicates a gating network The dynamic coefficients are generated and normalized by Softmax. Indicates the input feature map, Indicates the number of parallel cores. Represents convolution. Indicates the first One candidate convolutional kernel;

[0038] The output of the three-branch attention function is represented as follows:

[0039]

[0040] In the formula, Indicates the features after fusion. , and These represent the outputs of the channel attention, spatial attention, and context pyramid attention branches, respectively. , and Let each represent a learnable or posteriorly calibrated branch fusion coefficient, satisfying... .

[0041] Furthermore, in step 1, ③, the combined loss function includes classification loss, bounding box loss, and targetability loss, expressed as:

[0042]

[0043] In the formula, Represents the combination loss function. For classifying losses, For bounding box loss, For targeted losses, and All of them are hyperparameters that are greater than zero.

[0044] Furthermore, in step 1, ③, the category confidence... This indicates that the candidate box belongs to the risk element. The probability of pixel area ratio It is the ratio of the candidate frame area to the total frame area.

[0045] Furthermore, in step 1, ④, the slope parameter Values ​​are assigned based on the type of construction site environment: Urban construction environment Take 10, mountain road construction environment Take 8, rural construction environment Take 5, bridge / tunnel construction Take 10, highway / railway construction Take 5.

[0046] Furthermore, in step 3, ①, the scaling factor Values ​​are: beam bridge Take 1, arch bridge Take 2.5, suspension bridge Take 7, cable-stayed bridge Take 6.5.

[0047] Furthermore, in step 4, ③, the parameters are filtered. The value ranges from 0.5 to 1.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. Improve the accuracy of construction risk identification: This invention effectively improves the detection accuracy of risk elements in complex construction scenarios by combining the YOLOv8 model with dynamic convolution and multi-branch attention mechanisms. In particular, the accuracy is significantly improved compared with traditional methods in complex environments such as occlusion, backlighting, and scale changes.

[0050] 2. Risk assessment combining objective and subjective methods: This invention combines YOLOv8 to detect risk elements at bridge construction sites and introduces real-time image data, avoiding the limitations of traditional methods that rely too heavily on expert judgment, and enhancing the objectivity and accuracy of risk assessment results.

[0051] 3. Dynamic risk warning and control: By introducing dynamic convolution and time-varying adaptive branch fusion coefficients, this invention can respond to changes in the construction environment in real time and achieve accurate dynamic risk warning. Compared with traditional static risk assessment methods, it can provide more real-time and forward-looking construction safety assurance.

[0052] 4. Improve construction safety management efficiency: This invention adopts automated image detection and intelligent risk assessment, which can reduce the workload of manual inspection and assess construction risks in real time through intelligent means, significantly improving the safety management efficiency of the construction site and reducing errors and delays caused by human factors. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0054] A causal hierarchy modeling method for construction risks of long-span PC continuous rigid frame bridges that integrates visual perception includes the following steps:

[0055] Step 1: Visual perception and quantification of risk elements;

[0056] ① Risk elements and frequency:

[0057] Establish risk element set Risk elements in the set , Risk factors in the predetermined actual construction scenario are acquired through visual perception. Within the observation period, the frequency of occurrence of each risk element is statistically analyzed and calculated, as shown below:

[0058]

[0059] In the formula, This indicates the number of frames in which risk elements were detected within the observation period. This represents the total number of frames in the observation period. This indicates the frequency of occurrence, which is directly statistically analyzed from the detection logs.

[0060] ② Improved YOLOv8 structure:

[0061] Dynamic convolution and three-branch attention are introduced into the backbone and detection head of YOLOv8, where:

[0062] The output of dynamic convolution is represented as:

[0063]

[0064]

[0065] In the formula, This represents the output feature map. Indicates a gating network The dynamic coefficients are generated and normalized by Softmax. Indicates the input feature map, Indicates the number of parallel cores. Represents convolution. Indicates the first There are 10 candidate convolution kernels.

[0066] Simultaneously, the three-branch attention of the parallel construction channel, spatial, and context pyramids is fused into the output, represented as:

[0067]

[0068] In the formula, Indicates the features after fusion. , and These represent the outputs of the channel attention, spatial attention, and context pyramid attention branches, respectively. , and Let each represent a learnable or posteriorly calibrated branch fusion coefficient, satisfying... .

[0069] ③ Model training and inference output:

[0070] During the training phase, the model employs the standard YOLOv8 combined loss function, including classification loss, bounding box loss, and object detection loss, to ensure the reliability of the detection results in class recognition, location accuracy, and object presence detection, expressed as:

[0071]

[0072] In the formula, Represents the combination loss function. For classifying losses, For bounding box loss, For targeted losses, and All of them are hyperparameters that are greater than zero.

[0073] During the inference phase, YOLOv8 outputs a category confidence score for each candidate box. and pixel area ratio Category confidence This indicates that the candidate box belongs to the risk element. The probability of pixel area ratio It is the ratio of the candidate frame area to the total frame area.

[0074] ④ Risk quantification mapping of visual results:

[0075] Category confidence Mapped to probability of occurrence The risk intensity is defined by combining frequency and scope of impact. , represented as:

[0076]

[0077]

[0078] In the formula, The slope parameter should be selected based on the type of construction site environment. A recommended value is: for urban construction environments. Take 10, mountain road construction environment Take 8, rural construction environment Take 5, bridge / tunnel construction Take 10, highway / railway construction Take 5.

[0079] Step Two: Indicator Matrix and CRITIC Objective Weights;

[0080] ①Indicator matrix construction:

[0081] Build OK Column index matrix , represented as:

[0082]

[0083] Among them, matrix elements Indicates the first Risk element in each sample Indicators, including risk intensity Frequency and pixel area ratio , This is the sample index.

[0084] ② Range standardization:

[0085] To ensure numerical comparison of each indicator matrix and to prevent certain indicator matrices from being affected by large numerical ranges in the calculation, the indicator matrices were... Standardizing the range, it is expressed as:

[0086]

[0087] In the formula, and They represent the first The minimum and maximum values ​​of each risk element across all samples. The standardized value. .

[0088] ③ CRITIC weight calculation:

[0089] First, the CRITIC contrast intensity is calculated and expressed as follows:

[0090]

[0091] In the formula, Indicates the first The standard deviation of each risk element indicator, based on Perform calculations. Indicates the first The risk element indicator and the first The correlation coefficients of individual risk indicators Indicates the current number Other risk indicators for correlation comparison of individual risk indicators. This indicates the intensity of contrast; the higher the value, the stronger the distinction and conflict.

[0092] Then, the weights of the risk element indicators are obtained through normalization, as follows:

[0093]

[0094] In the formula, This represents the weight of the risk element indicator, and the sum of the weights of all risk element indicators equals 1.

[0095] Step 3: Construct the direct influence matrix;

[0096] ① Directly affects the definition

[0097] In the time window Inside, if risk element After it happened, There is another risk element in memory. If it is more likely to occur, then the risk element is considered to be... For another risk element It had a direct impact. Statistics on each risk element... After it happened Has at least one risk element appeared inside the window? , For the first Sub-risk element The time of occurrence is represented as follows:

[0098]

[0099] In the formula, Indicates risk element Starting from the moment of occurrence, examine the subsequent length. The window, if at least one risk element occurs. Then count 1 and apply it to all risk elements. Summation occurs. This indicates an indicator function; the value is 1 if the condition within the parentheses is true, and 0 otherwise. Indicates risk element The set of times of occurrence, Indicates the time of occurrence.

[0100] The conditional probability of the unbiased frequency estimate is then obtained, as follows:

[0101]

[0102] Then define the elements that directly affect the matrix. , means as follows:

[0103]

[0104] In the formula, Risk element The intensity of the risk, For all risk elements The highest risk intensity value is used for normalization. This is a scaling factor used to adjust the dimensions of probability and strength. Recommended values ​​are: [Values ​​for beam bridges]. Take 1, arch bridge Take 2.5, suspension bridge Take 7, cable-stayed bridge Take 6.5.

[0105] ② Directly affects matrix normalization:

[0106] To ensure the convergence of the method, normalization is performed based on the maximum row sum, as shown below:

[0107]

[0108]

[0109] In the formula, This is the unnormalized direct influence matrix. This is the normalized direct influence matrix.

[0110] Step 4: Causal analysis and screening of significant relationships;

[0111] ① Construction of the overall relationship matrix:

[0112] Based on the normalized direct influence matrix Calculate the total relation matrix , means as follows:

[0113]

[0114] In the formula, It is an identity matrix.

[0115] ② Calculate the degree of influence and the degree of impact:

[0116] Through the overall relationship matrix elements Calculate each risk element Influence and degree of impact , means as follows:

[0117] ,

[0118] In the formula, The first in the total relation matrix Line number Column elements, The first in the total relation matrix Line number The elements of the column.

[0119] ③Significant relation screening:

[0120] A threshold combining the mean and standard deviation is used to screen for significant causal relationships, as shown below:

[0121] ,

[0122] In the formula, To evaluate the threshold, Total relation matrix The mean, This is a filtering parameter, typically with a value between 0.5 and 1. Total relation matrix standard deviation For indicator functions, These are the elements of the filtered overall relation matrix, used for subsequent construction of the adjacency matrix.

[0123] ④ Construct the adjacency matrix:

[0124] Based on the results of the significant relation filtering, a thresholded directed adjacency matrix is ​​obtained. Used to represent each risk element The causal relationship between them.

[0125] Step 5: Hierarchical decomposition;

[0126] ① Obtain the reachability matrix:

[0127] For adjacency matrix Perform a transitive closure operation to obtain the reachability matrix. It is expressed as follows:

[0128]

[0129] In the formula, This represents a logical OR operation, ensuring the transitivity of matrices. It is an identity matrix to ensure the validity of transitive closure operations.

[0130] ② Divide into hierarchical levels:

[0131] Through the reachability matrix The elements in determine the risk element In the hierarchical partitioning, the level of the location, the reachable set, and the antecedent set are defined as follows:

[0132]

[0133]

[0134] In the formula, Represents the reachable set, that is, from the risk element The set of elements reachable from the starting point. This represents the antecedent set, i.e., the set that can reach the risk element. The set of elements, defined by the reachability matrix Row / column reading, Represents the reachability matrix The first in Line number Column elements, Represents the reachability matrix The first in Line number The elements of the column.

[0135] Based on the separation of reachability set and prior set:

[0136]

[0137] In the formula, Indicates the first A set of risk elements at the causal level (whose reachable set is no longer expanded by other nodes).

[0138] From the directed adjacency matrix peeling The rows and columns corresponding to the risk elements contained in the matrix are then used to obtain a new directed adjacency matrix. This hierarchical decomposition is repeated until all risk elements are layered. The value starts from 1 and does not exceed 1. .

[0139] Step Six: Prioritizing Key Risks and Identifying Propagation Paths;

[0140] ①Comprehensive priority calculation:

[0141] Define the overall priority It is expressed as follows:

[0142]

[0143] in, The larger the value, the earlier the position in the same causal hierarchy.

[0144] ② Path risk calculation:

[0145] Define path risk It is expressed as follows:

[0146]

[0147] in, For a directed path, For path length, For elements The maximum value.

[0148] Example

[0149] This embodiment takes the construction site of Tongshunhe Bridge as an example. The construction site type is a highway construction environment, the bridge type is a beam bridge, and two high-definition cameras are deployed for image acquisition.

[0150] In this embodiment, the observation period is taken. Frame, time window Frame, Parameter , , .

[0151] Risk element set: Not wearing a helmet No protection at height No isolation zone in the area Open flame operations were carried out without supervision.

[0152] The statistics for each risk element are as follows:

[0153] Number of occurrences: , , , .

[0154] The frequency of occurrence was calculated as follows: , , , .

[0155] After improving the YOLOv8 detection model, we obtained:

[0156] Category confidence level: , , , ;

[0157] Pixel area ratio: , , , .

[0158] Category confidence Mapped to probability of occurrence The risk intensity is defined by combining frequency and scope of impact. ,get:

[0159] Probability of occurrence: , , , ;

[0160] Risk level: , , , .

[0161] Indicator Matrix Build as:

[0162]

[0163] The range standardization yields the standardized matrix as follows:

[0164]

[0165] based on Calculate the standard deviation:

[0166]

[0167] Correlation coefficient:

[0168]

[0169] Contrast intensity:

[0170]

[0171] Weights of risk element indicators:

[0172]

[0173] by ( Taking "not wearing a helmet" as an example: Total number of triggers ,and The counts are respectively , , ,but:

[0174] , ,

[0175] Similarly, substitute the values ​​into the other rows. , , With the corresponding value.

[0176] The conditional probability of unbiased frequency estimation is as follows:

[0177]

[0178] The directly affected matrix is ​​as follows:

[0179]

[0180] The normalized matrix is ​​as follows:

[0181]

[0182] Based on the normalized direct influence matrix Calculate and obtain the total relation matrix as follows:

[0183]

[0184] Through the overall relationship matrix Calculate and obtain each risk element Influence and degree of impact as follows:

[0185]

[0186] The directed adjacency matrix is ​​as follows:

[0187]

[0188] The reachability matrix is ​​obtained by performing a transitive closure operation on the adjacency matrix:

[0189]

[0190] Divide into hierarchical levels:

[0191]

[0192] Overall Priority Represented as:

[0193]

[0194] Path risk Represented as:

[0195]

[0196] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the present invention.

[0197] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A causal hierarchy modeling method for construction risks of long-span PC continuous rigid frame bridges that integrates visual perception, characterized in that: Includes the following steps: Step 1: Visual perception and quantification of risk elements; ①Establish a risk element set Risk element , Risk factors in the predetermined actual construction scenarios are acquired through visual perception, and their frequency of occurrence is statistically analyzed and calculated within the observation period. , Indicates the number of frames representing the risk element. Indicates the total number of frames; ② Introduce dynamic convolution and three-branch attention in the backbone and detector head of YOLOv8; ③ During the training phase, the model uses the standard YOLOv8 combined loss function; during the inference phase, YOLOv8 outputs the class confidence score for each candidate box. and pixel area ratio ; ④ Class confidence Mapped to probability of occurrence The risk intensity is defined by combining frequency and scope of impact. , The slope parameter; Step Two: Indicator Matrix and CRITIC Objective Weights; ①Construction OK Column index matrix Matrix elements Indicates the first Indicators of risk elements in each sample, including risk intensity. Frequency and pixel area ratio , For sample index; ② For the index matrix Perform range standardization to obtain matrix elements. Standardized value , ; ③ Calculate the contrast intensity of CRITIC , Indicates the first The standard deviation of each risk element indicator Other risk metrics used to represent correlation comparisons The correlation coefficient is represented, and then normalized to obtain the weights of the risk element indicators. , And the sum of the weights of all risk element indicators equals 1; Step 3: Construct the direct influence matrix; ① In the time window Inside, if risk element After it happened, There is another risk element in memory. If it is more likely to occur, then the risk element is considered to be... For another risk element It had a direct impact, and the statistics of each risk element were analyzed. After it happened Has at least one risk element appeared inside the window? , For the first Sub-risk element The time of occurrence is represented as follows: In the formula, Indicates risk element Starting from the moment of occurrence, examine the subsequent length. The window, if at least one risk element occurs. Then count 1 and apply it to all risk elements. Summation occurs. This indicates an indicator function; the value is 1 if the condition within the parentheses is true, and 0 otherwise. Indicates risk element The set of times of occurrence, Indicates the time of occurrence; This leads to the conditional probability of the unbiased frequency estimate. And define the elements that directly affect the matrix. , Risk element The intensity of the risk, For all risk elements The highest risk intensity value, This is the scaling factor; ② Normalize the matrix by the maximum row sum to obtain the normalized direct influence matrix. ; Step 4: Causal analysis and screening of significant relationships; ① Based on the normalized direct influence matrix Calculate the overall relation matrix , It is the identity matrix; ② Through the elements of the overall relation matrix Calculate the impact of each risk element and degree of impact , and The first Line number Column elements and the first Line number Column elements; ③ Use evaluation thresholds To filter out significant causal relationships and obtain the elements of the filtered overall relationship matrix. , Total relation matrix The mean, To filter parameters, Total relation matrix standard deviation For indicator functions; ④ Obtain the adjacency matrix based on the filtering results. ; Step 5: Hierarchical decomposition; ① For the adjacency matrix Perform a transitive closure operation to obtain the reachability matrix. ; ② Through the reachability matrix The elements in determine the risk element In the hierarchical partitioning, the level at which the set exists is defined as the reachable set. With the Preview , and Represent the reachability matrix respectively The first in Line number The elements of the column and the first Line number The elements of the column are stripped to obtain the first... Risk element set at the causal level Repeat the hierarchical decomposition until all risk elements are layered. Step Six: Prioritizing Key Risks and Identifying Propagation Paths; ① Define the overall priority , The larger the value, the earlier the position in the same causal hierarchy; ② Define path risk level , For a directed path, For path length, For elements The maximum value.

2. The method for hierarchical modeling of construction risk causality of a long-span PC continuous rigid frame bridge integrating visual perception as described in claim 1, characterized in that: In step 1, ②, the dynamic convolution output is represented as: In the formula, This represents the output feature map. Indicates a gating network The dynamic coefficients are generated and normalized by Softmax. Indicates the input feature map, Indicates the number of parallel cores. Represents convolution. Indicates the first One candidate convolutional kernel; The output of the three-branch attention function is represented as follows: In the formula, Indicates the features after fusion. , and These represent the outputs of the channel attention, spatial attention, and context pyramid attention branches, respectively. , and Let each represent a learnable or posteriorly calibrated branch fusion coefficient, satisfying... .

3. The method for hierarchical modeling of construction risk causality of a long-span PC continuous rigid frame bridge integrating visual perception as described in claim 1, characterized in that: In step 1, ③, the combined loss function includes classification loss, bounding box loss, and targetability loss, expressed as: In the formula, Represents the combination loss function. For classifying losses, For bounding box loss, For targeted losses, and All of them are hyperparameters that are greater than zero.

4. The method for hierarchical modeling of construction risk of long-span PC continuous rigid frame bridges integrating visual perception as described in claim 1, characterized in that: In step 1, ③, the category confidence level This indicates that the candidate box belongs to the risk element. The probability of pixel area ratio It is the ratio of the candidate frame area to the total frame area.

5. The method for hierarchical modeling of construction risk causality of a long-span PC continuous rigid frame bridge integrating visual perception as described in claim 1, characterized in that: In step 1, ④, the slope parameter Values ​​are assigned based on the type of construction site environment: Urban construction environment Take 10, mountain road construction environment Take 8, rural construction environment Take 5, bridge construction Take 10, highway / railway construction Take 5.

6. The method for hierarchical modeling of construction risk causality of a long-span PC continuous rigid frame bridge integrating visual perception as described in claim 1, characterized in that: In step 4, ③, the parameters are filtered. The value ranges from 0.5 to 1.

Images