An engineering safety assessment method and system based on artificial intelligence

Abstract

The application provides an engineering safety evaluation method and system based on artificial intelligence, and belongs to the technical field of engineering structure health monitoring. The method collects time series data of physical parameters of an engineering structure through a distributed sensor network, synchronously acquires surface visual data, and processes the data; the processed data is input into a multi-modal artificial intelligence analysis model; an improved Mogrifier LSTM network processes the time series data through four rounds of iterative optimization and a sensor health perception mechanism, and an improved YOLOv8 network realizes damage positioning, classification and crack width quantification through a coordinate attention module and a decoupling detection head; the time series and visual features are fused through a cross-modal attention mechanism to generate a risk part feature vector; a safety risk index is calculated based on the risk part feature vector, and a three-level risk grade and a structured report are output. The application fuses multi-source data, and significantly improves the accuracy and real-time performance of engineering safety evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engineering structure health monitoring, in particular to an engineering safety evaluation method and system based on artificial intelligence. BACKGROUND

[0002] The field of engineering structure safety evaluation has long been faced with the core challenges of insufficient multi-source information fusion and difficult dynamic interference suppression. The current methods have the following problems:

[0003] Data coordination limitation: manual inspection and automatic monitoring systems usually operate independently, and the physical parameters (such as vibration, strain) collected by the sensor network and the visual appearance damage information lack effective correlation, making it difficult to build a overall understanding of the structure state;

[0004] Insufficient environmental robustness: changes in temperature and humidity on the engineering site are easy to cause sensor data drift, and existing filtering algorithms mostly rely on static thresholds, and the adaptability to dynamic environmental interference needs to be improved;

[0005] Limited depth of quantitative evaluation: although artificial intelligence-based damage identification models have been applied, the influence of the state of the sensing device itself on the reliability of the data is rarely considered in time series analysis, and visual detection mostly focuses on damage positioning while ignoring the automatic extraction of key quantitative parameters (such as crack width);

[0006] Decision support granularity is coarse: the existing risk output often stays at the overall safety level, lacking fine guidance of component-level positioning identification and response instructions.

[0007] The above bottlenecks restrict the real-time and engineering applicability of the evaluation results, and it is urgent to realize multi-modal data deep fusion, dynamic interference suppression and quantitative damage parameter generation through technological innovation. SUMMARY

[0008] The purpose of the present application is to provide an engineering safety evaluation method and system based on artificial intelligence, which realizes real-time intelligent diagnosis and risk warning of engineering structure damage by fusing multi-modal AI (sensor time series analysis + visual damage identification), and solves the problem of evaluation lag caused by artificial dependence and multi-source data fragmentation in traditional methods.

[0009] To achieve the above purpose, the present application provides an engineering safety evaluation method based on artificial intelligence, comprising the following steps:

[0010] Step S1, collecting physical parameter time series data of the engineering structure through a distributed sensor network;

[0011] Step S2, synchronously acquiring surface visual data of the engineering structure;

[0012] Step S3, dynamic noise filtering processing is performed on the physical parameter time series data, and geometric distortion correction is performed on the surface visual data;

[0013] Step S4, inputting the data processed in step S3 into a multi-modal artificial intelligence analysis model, specifically including:

[0014] The time series analysis branch uses an improved Mogrifier LSTM network to process sensor data;

[0015] The visual analysis branch uses an improved YOLOv8 network for damage identification;

[0016] Through a cross-modal attention mechanism, time series and visual features are fused to output a risk part feature vector;

[0017] Step S5, calculating a safety risk index based on the risk part feature vector;

[0018] Step S6, outputting a three-level risk grade and a structured risk report according to the safety risk index, the report including a risk component positioning identifier, a damage type and a quantitative parameter, and a hierarchical response instruction.

[0019] Preferably, in step S1, the sensor types include strain sensors, displacement sensors, and vibration sensors.

[0020] Preferably, in step S3, the dynamic noise filtering step is as follows:

[0021] The collected engineering structure physical parameter time series signal is layered and decomposed, and a wavelet denoising threshold is adaptively determined according to the signal characteristics to remove noise interference in the signal;

[0022] Considering the influence of temperature and humidity on the engineering site, the denoised signal is compensated and corrected to obtain compensated physical parameter time series data ;

[0023] ;

[0024] Wherein, represents the original physical parameter time series data, represents the environmental humidity of the engineering site, represents the environmental temperature of the engineering site, represents the temperature coefficient of the sensor, with a unit of microstrain per degree Celsius, represents the humidity coefficient of the sensor, with a unit of microstrain per percentage relative humidity.

[0025] Preferably, in step S4, the improved Mogrifier LSTM network performs four rounds of iteration optimization on the input vector and hidden state to act on the LSTM core calculation, specifically as follows:

[0026] input vector corresponding to the time series data of physical parameters of the engineering structure collected by the sensor , after linear transformation by weight matrix layer normalization hidden state perform Hadamard product to obtain interaction transformation result ;

[0027] hidden state after transformation by weight matrix , and perform Hadamard product to obtain interaction transformation result , realizing bidirectional interaction transformation of the input vector and the hidden state; wherein, represents the hidden layer dimension set according to the engineering data processing requirement, represents the input vector dimension, represents element-wise multiplication, represents layer normalization operation;

[0028] An adjustment coefficient is introduced, and the decay factor is calculated by combining the engineering site environment temperature and the service time of the sensor, and the decay factor is applied to the forgetting gate calculation of the LSTM:

[0029] ;

[0030] ;

[0031] ;

[0032] wherein, represents the forgetting gate output, represents the weight matrix of the forgetting gate, represents the bias term of the forgetting gate, represents the Sigmoid activation function, represents the spliced vector, represents the decay factor, represents the engineering site environment temperature, represents the service time of the sensor, represents the adjustment coefficient, represents a very small constant.

[0033] Preferably, the improved Mogrifier LSTM network outputs sensor health perception related parameters , and predicts the health index of the sensor:

[0034] ;

[0035] wherein, represents the health index, a coefficient representing the comprehensive influence of the environment and the service time on the sensor, , are respectively the results of hidden state updates after the first to fourth iterations of the loop;

[0036] The decay factor is modified to obtain a modified decay factor :

[0037] ;

[0038] When the health index is less than 0.3, mark the sensor as faulty, and apply the modified decay factor to the LSTM forget gate calculation again.

[0039] Preferably, in step S4, the improved YOLOv8 network comprises:

[0040] When processing the visual data feature map of the engineering structure surface, a coordinate attention module operation is performed at the end of the Backbone;

[0041] The detection head is decoupled into three branches, including:

[0042] The damage location regression branch uses the EIoU loss to output the boundary box coordinates to locate the damage area for the engineering structure damage;

[0043] The damage type classification branch outputs a nine-class damage probability distribution;

[0044] The crack width regression branch outputs a pixel-level width map, which identifies and quantifies the damage on the surface of the engineering structure from multiple dimensions such as location, type, and extent, and outputs the processed visual data features.

[0045] Preferably, the crack width regression branch is implemented as follows:

[0046] For the input visual data feature map of the engineering structure surface, first, expand the channels through convolution to increase the number of feature map channels, and then obtain the feature map after pixel reorganization and upsampling through pixel shuffling operation;

[0047] The feature map after pixel reorganization and upsampling is optimized using the smooth L1 loss function.

[0048] Preferably, in step S4, the cross-modal attention mechanism is used to fuse the time series and visual features to output a risk part feature vector, as follows:

[0049] The time series feature obtained by the time series analysis branch and the visual feature Respective projection transformation is carried out, a space constraint matrix is constructed based on the space characteristics of the engineering structure, and a normalized correlation matrix is generated ;

[0050] Based on the normalized correlation matrix , execution is performed through a gated recurrent unit , and a fused feature vector is output .

[0051] Preferably, in step S5, the risk index calculation step is as follows:

[0052] The risk part feature vector is processed by a three-layer fully connected network to obtain the damage score of each component of the engineering structure.

[0053] Based on the damage score, the component weight and failure influence factor determined in the engineering component design are combined to calculate the safety risk index of the engineering structure.

[0054] The application also provides an engineering safety evaluation system based on artificial intelligence, comprising:

[0055] Comprising:

[0056] A distributed sensor network module is used to collect physical parameter time series data of the engineering structure, and the sensor types include strain sensors, displacement sensors and vibration sensors.

[0057] A visual acquisition module is used to synchronously acquire surface visual data of the engineering structure.

[0058] A data preprocessing module comprises:

[0059] A dynamic noise filtering unit is used for dynamic noise filtering processing of the physical parameter time series data.

[0060] A geometric distortion correction unit is used for correction of the surface visual data.

[0061] A multi-modal artificial intelligence analysis module comprises:

[0062] A time series analysis unit uses an improved Mogrifier LSTM network to process sensor data, which performs four rounds of iterative optimization of input vectors and hidden states, and calculates a decay factor based on environmental temperature and sensor service time, which is applied to the LSTM forget gate.

[0063] A visual analysis unit uses an improved YOLOv8 network for damage identification, which integrates a coordinate attention module at the end of the Backbone, and decouples the detection head into a damage position regression branch, a damage type classification branch and a crack width regression branch.

[0064] The cross-modal fusion unit is configured to fuse the time sequence features and the visual features through a cross-modal attention mechanism to generate a risk part feature vector.

[0065] The safety risk calculation module is configured to process the risk part feature vector by using a three-layer fully connected network, and calculate a safety risk index by combining a component weight and a failure influence factor.

[0066] The risk report generation module is configured to output a three-level risk grade and a structured risk report according to the safety risk index, wherein the report includes a risk component positioning identifier, a damage type and a quantitative parameter, and a hierarchical response instruction.

[0067] Therefore, the engineering safety evaluation method and system based on artificial intelligence have the following beneficial technical effects:

[0068] (1) Improved evaluation accuracy and real-time performance: The method fuses multi-source data, including physical parameter time sequence data collected by a distributed sensor network and surface visual data synchronously acquired, and inputs the processed data into a multi-modal artificial intelligence analysis model. The improved Mogrifier LSTM network and the improved YOLOv8 network are respectively used to process the time sequence data and the visual data, and the features are fused through a cross-modal attention mechanism, so that the safety risk of the engineering structure can be more accurately evaluated. Compared with the traditional method, the accuracy and real-time performance are significantly improved, and the evaluation lag problem caused by the fragmentation of multi-source data in the traditional method is solved.

[0069] (2) Enhanced environmental adaptability: In the dynamic noise filtering process, the influence of the temperature and humidity of the engineering site is considered, and the compensated and corrected signal after denoising is obtained, so that the sensor data can better adapt to the complex and variable engineering site environment, and the reliability of the data is improved.

[0070] (3) Deeply quantifying damage parameters: The improved YOLOv8 network can not only realize damage positioning, but also output nine types of damage probability distribution and pixel-level width map, so as to identify and quantify the surface damage of the engineering structure from multiple dimensions such as position, type and degree, and provide more rich quantitative parameters for engineering safety evaluation, which is conducive to more accurately judging the severity of damage and the influence on the structure safety.

[0071] (4) Ensuring the reliability of sensor data: The improved Mogrifier LSTM network can output sensor health perception correlation parameters, predict the health index of the sensor, and mark the fault when the health index of the sensor is lower than the threshold. At the same time, the decay factor is corrected and applied to the LSTM forgetting gate calculation again, so as to effectively consider the influence of the state of the sensor on the reliability of the data, ensure the quality of the sensor data input into the analysis model, and improve the stability and accuracy of the entire evaluation system. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 Flowchart of the engineering safety assessment method based on artificial intelligence of the present application;

[0073] Figure 2 Multimodal artificial intelligence analysis model architecture diagram;

[0074] Figure 3 Improved YOLOv8 network detail diagram. DETAILED DESCRIPTION

[0075] The technical solutions of the present application are further described below through the drawings and examples.

[0076] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those skilled in the art to which the present application belongs.

[0077] Example 1

[0078] As shown in Figure 1 , an engineering safety assessment method based on artificial intelligence is taken for a certain actual bridge engineering in Beijing, and the specific implementation process is as follows:

[0079] Step S1, collect the time series data of the physical parameters of the engineering structure through a distributed sensor network.

[0080] Deploy 150 strain sensors, 30 displacement sensors and 50 vibration sensors at key parts of the bridge tower, main beam and piers of a certain large cable-stayed bridge. These sensors are evenly distributed and are mainly arranged in areas prone to problems, such as the midspan, the vicinity of the support, etc. The sampling frequency of the sensors is set to 100 Hz.

[0081] Step S2, synchronously acquire the surface visual data of the engineering structure.

[0082] Install 4 high-definition cameras on both sides of the bridge and at key locations, respectively on the east, west, south and north sides of the bridge. The resolution of the camera is 4K, and the frame rate is 30 fps, which can clearly capture the damage conditions on the surface of the bridge, such as cracks, peeling, etc.

[0083] Step S3, dynamic noise filtering processing is performed on the time series data of the physical parameters, and geometric distortion correction is performed on the surface visual data;

[0084] (1) The dynamic noise filtering step is as follows:

[0085] The collected time series signals of the physical parameters of the engineering structure are decomposed into 5 layers, and the wavelet denoising threshold is determined adaptively according to the signal characteristics to remove noise interference in the signal;

[0086] The denoised signal is compensated and corrected considering the influence of the temperature and humidity of the engineering site, to obtain compensated physical parameter time series data ;

[0087] ;

[0088] wherein, represents the original physical parameter time series data, represents the environmental humidity of the engineering site, represents the environmental temperature of the engineering site, represents the temperature coefficient of the sensor, in units of micro-strain per degree Celsius, represents the humidity coefficient of the sensor, in units of micro-strain per percentage relative humidity.

[0089] (2) The geometric distortion correction step is as follows:

[0090] The collected visual image is geometrically corrected. For example, the camera is calibrated using Zhang Zhengyou calibration method to obtain the intrinsic matrix (focal length, principal point coordinates, etc.) and extrinsic matrix (rotation matrix and translation vector) of each camera. Then, the undistort function in the OpenCV library is used to correct the image and eliminate the image distortion caused by the camera lens distortion.

[0091] Step S4, as shown in Figure 2 , input the data processed in step S3 into a multi-modal artificial intelligence analysis model, specifically including:

[0092] (1) The time series analysis branch uses an improved Mogrifier LSTM network to process sensor data;

[0093] The improved Mogrifier LSTM network performs four rounds of iterative optimization on the input vector and the hidden state to act on the LSTM core calculation, specifically as follows:

[0094] For the input vector corresponding to the physical parameter time series data of the engineering structure collected by the sensor, after linear transformation by the weight matrix , the layer normalized hidden state performs Hadamard product to obtain the interaction transformation result ;

[0095] The hidden state is transformed by the weight matrix , and performs Hadamard product to obtain the interaction transformation result , realizing the bidirectional interaction transformation of the input vector and the hidden state; wherein, represents the hidden layer dimension set to adapt to the engineering data processing requirements, represents input vector dimension, represents element-wise multiplication, represents layer normalization operation;

[0096] An adjustment coefficient is introduced, combined with the engineering site environment temperature and the service time of the sensor, to calculate the attenuation factor, and the attenuation factor is applied to the forgetting gate calculation of LSTM:

[0097] ;

[0098] ;

[0099] ;

[0100] wherein, represents the output of the forgetting gate, represents the weight matrix of the forgetting gate, represents the bias term of the forgetting gate, represents the Sigmoid activation function, represents the spliced vector, represents the attenuation factor, represents the engineering site environment temperature, represents the service time of the sensor, represents the adjustment coefficient, represents a very small constant, with a value less than 10 -6 .

[0101] The improved Mogrifier LSTM network outputs the sensor health perception correlation parameters , and predicts the health index of the sensor:

[0102] ;

[0103] wherein, represents the health index, represents the comprehensive influence coefficient of the environment and service time on the sensor, , are the results of hidden state update after the first to fourth iterations, respectively;

[0104] The attenuation factor is corrected to obtain the corrected attenuation factor :

[0105] ;

[0106] When the health index is less than 0.3, mark the sensor failure, and at the same time, the corrected attenuation factor Again applied to the LSTM forget gate calculation. The strain sensor in this embodiment has been in service for 3 years, and its calculated health index is 0.51, which is greater than 0.3, indicating that the sensor is currently in a normal working state, and the data is reliable.

[0107] (2) The visual analysis branch uses an improved YOLOv8 network for damage identification;

[0108] As shown in Figure 3 , the improved YOLOv8 network includes:

[0109] When processing the visual data feature map of the engineering structure surface, a coordinate attention module operation is performed at the end of the Backbone, which enhances the model's positioning ability for damage areas by redistributing the channel and spatial information of the feature map.

[0110] The detection head is decoupled into three branches, including:

[0111] The damage location regression branch uses the EIoU loss to output the bounding box coordinates to locate the damage area; for example, the bounding box coordinates of a crack are (100, 150, 300, 400), indicating the location of the crack in the image;

[0112] The damage type classification branch outputs a nine-class damage probability distribution;

[0113] including cracks, spalling, exposed reinforcement, holes, honeycomb pitting, steel corrosion, net-like cracks, concrete aging, and support damage. Taking a damage area as an example, the classification results are as follows:

[0114] Crack: probability of 0.92, indicating that the damage area is likely to be a crack, possibly caused by factors such as structural stress, temperature changes, or material shrinkage.

[0115] Spalling: probability of 0.05, indicating a certain possibility of spalling, possibly due to the aging or impact of the protective layer on the concrete surface.

[0116] Exposed reinforcement: probability of 0.01, low possibility, but attention should be paid to whether there is exposed reinforcement, possibly due to insufficient concrete protection or damage.

[0117] Hole: probability of 0.01, low possibility of hole occurrence, but check whether there are internal cavities in the concrete, which may affect the integrity of the structure.

[0118] Honeycomb pitting: probability of 0.005, very low possibility, but attention should be paid to whether there are honeycomb or pitted defects on the concrete surface, which may affect the durability of the structure.

[0119] Steel bar corrosion: probability 0.003, low possibility, but still need to be vigilant about the signs of steel bar corrosion, which may be caused by concrete cover damage or environmental corrosion factors.

[0120] Net crack: probability 0.001, extremely low possibility, indicating a very small chance of net crack, which may be caused by concrete shrinkage or temperature change.

[0121] Concrete aging: probability 0.001, extremely low possibility, indicating that concrete aging is not obvious, but long-term environmental factors still need to be considered for their impact on concrete performance.

[0122] Support damage: probability 0.001, extremely low possibility, but attention should be paid to whether the support has aging, deformation or displacement damage, which may affect the normal force transmission of the structure.

[0123] The crack width regression branch outputs a pixel-level width map, which identifies and quantifies the surface damage of the engineering structure from multiple dimensions such as location, type and degree, and outputs the processed visual data features.

[0124] The crack width regression branch implementation steps are as follows:

[0125] The input feature map is first expanded by channel convolution, increasing the number of channels from 256 to 512, and then the pixel shuffle operation is performed to obtain the feature map after pixel reorganization and upsampling. The model is optimized using the smooth L1 loss function, and the pixel-level width map is output. For example, in a certain crack area, the pixel value range in the width map is 3-7, and the actual crack width is 3-7mm. By analyzing the width map, the width distribution of the crack can be obtained, providing a quantitative basis for evaluating the severity of the crack.

[0126] (3) Fuse the timing and visual features through the cross-modal attention mechanism to output the risk part feature vector, as follows:

[0127] Project the timing features obtained by the timing analysis branch and the visual features obtained by the visual analysis branch respectively, construct a spatial constraint matrix according to the spatial structure characteristics of the bridge and the connection relationship between the components, and then generate a normalized correlation matrix ;

[0128] Based on the normalized correlation matrix , execute through the gated recurrent unit to output the fused feature vector .

[0129] Step S5, calculate the safety risk index based on the risk part feature vector:

[0130] The three-layer fully connected network is used to process the risk part feature vector. The dimension of the input feature vector is 512, the first layer of fully connected network maps it to 256 dimensions, the second layer maps it to 128 dimensions, and the third layer maps it to 64 dimensions, to obtain the damage score of each component of the engineering structure. For example, the damage score of a pier is 0.72, indicating that the damage degree of the pier is relatively heavy, which may have a greater impact on the overall safety of the bridge; the damage score of a main beam is 0.45, indicating that the damage degree is relatively light, but still needs attention.

[0131] Based on the damage score, the component weight and failure influence factor determined in the engineering component design are combined to calculate the safety risk index of the engineering structure. According to the design documents and structural analysis results of the bridge, the weight and failure influence factor of each component are determined. For example, the weight of the pier is 0.3, and the failure influence factor is 0.8; the weight of the main beam is 0.4, and the failure influence factor is 0.7. Substitute the formula to calculate the contribution of each component to the overall safety risk index, the contribution of the pier is 0.72x0.3x0.8=0.1728, and the contribution of the main beam is 0.45x0.4x0.7=0.126. The contributions of all components are integrated to obtain the overall safety risk index of the bridge, which is 0.65.

[0132] Step S6, according to the safety risk index 0.65, combined with the risk level division standard (low risk 0-0.3, medium risk 0.3-0.7, high risk 0.7-1.0), it is determined that the bridge is in the medium risk level. This shows that the bridge currently has certain safety risk, and corresponding measures need to be taken for management and maintenance.

[0133] The data of the same bridge at three different time periods are selected for comparison, which are the traditional method (only using manual inspection and single sensor monitoring), the non-fusion method (only using sensor data without multi-modal fusion) and the method of the present application (multi-modal data fusion artificial intelligence analysis method). The results are shown in Table 1:

[0134] Table 1 Comparison results

[0135] ;

[0136] The comparison results show that the method of the present application is significantly better than the traditional method and the non-fusion method in terms of data acquisition and processing efficiency, damage identification accuracy and risk assessment accuracy. The acquisition time of the present application is only 8 hours, the processing time is 15 minutes, the damage identification accuracy is 95%, the risk assessment accuracy is 90%, and the sensor health status can be monitored comprehensively, and the error of automatic measurement of crack width is only ±0.5mm.

[0137] Example 2

[0138] An engineering safety evaluation system based on artificial intelligence, comprising:

[0139] a distributed sensor network module for collecting time series data of physical parameters of the engineering structure, the sensor types including strain sensors, displacement sensors and vibration sensors;

[0140] a visual acquisition module for synchronously acquiring surface visual data of the engineering structure;

[0141] a data preprocessing module, comprising:

[0142] a dynamic noise filtering unit for performing dynamic noise filtering processing on the time series data of physical parameters;

[0143] a geometric distortion correction unit for performing correction on the surface visual data;

[0144] a multi-modal artificial intelligence analysis module, comprising:

[0145] a time series analysis unit for processing sensor data by using an improved Mogrifier LSTM network, which performs four rounds of iterative optimization of input vectors and hidden states, and calculates a decay factor based on ambient temperature and service time of the sensor and applies it to the forgetting gate of the LSTM;

[0146] a visual analysis unit for damage identification by using an improved YOLOv8 network, which integrates a coordinate attention module at the end of the Backbone and decouples the detection head into a damage location regression branch, a damage type classification branch and a crack width regression branch;

[0147] a cross-modal fusion unit for fusing time series features and visual features by using a cross-modal attention mechanism to generate a risk part feature vector;

[0148] a safety risk calculation module for processing the risk part feature vector by using a three-layer fully connected network, combining a component weight and a failure influence factor to calculate a safety risk index;

[0149] a risk report generation module for outputting a three-level risk grade and a structured risk report according to the safety risk index, the report including a risk component positioning identifier, a damage type and quantitative parameters and a graded response instruction.

[0150] It should be noted that the contents not elaborated in the present application are all prior art and are well known to those skilled in the art.

[0151] Therefore, the present application adopts the above-mentioned engineering safety evaluation method and system based on artificial intelligence, realizes real-time intelligent diagnosis and risk warning of engineering structure damage by fusing multi-modal AI (sensor time series analysis + visual damage identification), and solves the problems of manual dependence and evaluation lag caused by fragmented multi-source data in traditional methods.

[0152] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. An artificial intelligence-based engineering safety assessment method, characterized in that, Includes the following steps: Step S1: Collect time-series data of physical parameters of the engineering structure through a distributed sensor network; Step S2: Synchronously acquire surface visual data of the engineering structure; Step S3: Perform dynamic noise filtering on the time series data of physical parameters and correct geometric distortion of surface visual data; Step S4: Input the data processed in step S3 into the multimodal artificial intelligence analysis model, specifically including: The time-series analysis branch uses an improved Mogrifier LSTM network to process sensor data; The visual analysis branch uses an improved YOLOv8 network for damage identification; By fusing the temporal features obtained from the temporal analysis branch and the visual features obtained from the visual analysis branch through a cross-modal attention mechanism, a feature vector of the risk location is output. Step S5: Calculate the safety risk index based on the feature vector of the risky location; Step S6: Output a three-level risk rating and a structured risk report based on the security risk index; In step S4, the improved Mogrifier LSTM network performs four rounds of iterative optimization on the input vector and hidden state to affect the core computation of the LSTM, as follows: The input vector corresponding to the time-series data of the physical parameters of the engineering structure acquired by the sensor. Through the weight matrix After linear transformation, the hidden state is normalized to the layer. Performing the Hadamard product yields the result of the interactive transformation. ; Hidden state Through weight matrix After transformation, with Performing the Hadamard product yields the result of the interactive transformation. This enables bidirectional interactive transformation between the input vector and the hidden state; where, This indicates the hidden layer dimension set to meet the needs of engineering data processing. Indicates the dimension of the input vector. This indicates element-wise multiplication. Presentation layer normalization operation; An adjustment coefficient is introduced, and an attenuation factor is calculated by combining the ambient temperature at the engineering site and the service life of the sensor. This attenuation factor is then applied to the forget gate calculation of the LSTM. ； ； ； in, Indicates the output of the forget gate. The weight matrix represents the forget gate. The bias term representing the forget gate. This represents the Sigmoid activation function. This indicates the concatenation of vectors. Indicates the attenuation factor. Indicates the ambient temperature at the construction site. Indicates the service life of the sensor. Indicates the adjustment factor. Represents a minimal constant; Improved Mogrifier LSTM network outputs sensor health sensing correlation parameters Predicting the health index of sensors: ； in, Indicates health index, This represents the combined impact coefficient of environment and service life on the sensor. , These are the results of the hidden state updates after the first to fourth iterations, respectively. The attenuation factor is corrected to obtain the corrected attenuation factor. : ； When the health index is less than 0.3, the sensor is marked as faulty, and the corrected attenuation factor is applied. This is then applied again to LSTM forget gate calculation.

2. The engineering safety assessment method based on artificial intelligence according to claim 1, characterized in that, In step S1, the sensor types include strain sensors, displacement sensors, and vibration sensors.

3. The engineering safety assessment method based on artificial intelligence according to claim 1, characterized in that, In step S3, the dynamic noise filtering steps are as follows: The time-series signals of the physical parameters of the engineering structure were collected and processed. Layer decomposition, adaptively determining the wavelet denoising threshold based on signal characteristics, to remove noise interference from the signal; Considering the influence of temperature and humidity at the engineering site, the denoised signal is compensated and corrected to obtain the compensated physical parameter time series data. ; ； in, This represents the time series data of the original physical parameters. Indicates the humidity of the construction site environment. Indicates the ambient temperature at the construction site. Indicates the temperature coefficient of the sensor. This indicates the humidity coefficient of the sensor.

4. The engineering safety assessment method based on artificial intelligence according to claim 1, characterized in that, In step S4, the improved YOLOv8 network includes: When processing visual data feature maps of engineering structure surfaces, coordinate attention module operations are performed at the end of the backbone. The detection head is decoupled into three branches, including: The damage location regression branch targets damage to engineering structures, using EIoU loss to output bounding box coordinates to locate the damage area. The damage type classification branch outputs the probability distribution of nine types of damage. The crack width regression branch outputs a pixel-level width map, which identifies and quantifies surface damage to engineering structures from multiple dimensions such as location, type, and degree, and outputs the processed visual data features.

5. The engineering safety assessment method based on artificial intelligence according to claim 4, characterized in that, The steps for implementing the crack width regression branch are as follows: The input visual data feature map of the engineering structure surface is first subjected to channel expansion convolution to increase the number of feature map channels, and then the pixel shuffling operation is used to obtain the feature map after pixel recombination and upsampling. The feature map after pixel recombination and upsampling is then used to optimize the model using a smooth L1 loss function.

6. The engineering safety assessment method based on artificial intelligence according to claim 1, characterized in that, In step S4, temporal and visual features are fused through a cross-modal attention mechanism to output a feature vector of the risk area, as follows: Temporal features obtained from the temporal analysis branch Visual features obtained from the visual analysis branch Projection transformations are performed separately, and a spatial constraint matrix is ​​constructed based on the spatial characteristics of the engineering structure, thereby generating a normalized correlation matrix. ; Based on normalized correlation matrix Executed through the gated loop unit Output the fused feature vector .

7. The engineering safety assessment method based on artificial intelligence according to claim 1, characterized in that, In step S5, the risk index calculation steps are as follows: A three-layer fully connected network is used to process the feature vectors of risky parts to obtain the damage scores of each component of the engineering structure. Based on the damage score, and combined with the component weights and failure impact factors determined in the engineering component design, the safety risk index of the engineering structure is calculated.

8. An engineering safety assessment system based on artificial intelligence, characterized in that, The method for performing the AI-based engineering safety assessment method as described in any one of claims 1-7 includes: Distributed sensor network modules are used to collect time-series data of physical parameters of engineering structures. Sensor types include strain sensors, displacement sensors, and vibration sensors. The vision acquisition module is used to synchronously acquire surface visual data of the engineering structure; The data preprocessing module includes: The dynamic noise filtering unit is used to perform dynamic noise filtering on physical parameter time-series data. Geometric distortion correction unit, used to correct surface visual data; The multimodal artificial intelligence analysis module includes: The timing analysis unit uses an improved Mogrifier LSTM network to process sensor data. This network performs four rounds of iterative optimization of the input vector and hidden state, and calculates the decay factor based on the ambient temperature and sensor service life, which is then applied to the LSTM forget gate. The visual analysis unit uses an improved YOLOv8 network for damage recognition. Its backbone end integrates a coordinate attention module, and the detection head is decoupled into a damage location regression branch, a damage type classification branch, and a crack width regression branch. The cross-modal fusion unit is used to fuse temporal and visual features through a cross-modal attention mechanism to generate a feature vector of risk sites; The safety risk calculation module uses a three-layer fully connected network to process the feature vectors of risky parts, and combines the component weights and failure impact factors to calculate the safety risk index. The risk report generation module outputs a three-level risk level and a structured risk report based on the safety risk index. The report includes the location identifier of the risky component, the damage type and quantitative parameters, and the graded response instructions.

Images