Engineering safety assessment method and system based on artificial intelligence

Abstract

The invention provides an artificial intelligence-based engineering safety assessment method and system, and belongs to the technical field of engineering structure health monitoring. The method comprises the following steps: acquiring physical parameter time sequence data of an engineering structure through a distributed sensor network, synchronously acquiring surface visual data, and processing the data; inputting the processed data into a multi-modal artificial intelligence analysis model; processing time sequence data by an improved Mogrimer LSTM network through four-round iterative optimization and a sensor health sensing mechanism, and realizing damage positioning, classification and crack width quantification by an improved YOLOv8 network through a coordinate attention module and a decoupling detection head; fusing the time sequence and the visual features through a cross-modal attention mechanism, and generating a risk part feature vector; and calculating a safety risk index based on the risk part feature vector, and outputting a three-level risk level and a structured report. According to the method, multi-source data are fused, and the precision and real-time performance of engineering safety assessment are remarkably improved.

Description

Technical Field

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

[0002] The field of engineering structure safety assessment has long faced the core challenges of insufficient multi-source information fusion and difficulty in suppressing dynamic interference. Current methods have the following problems: Data collaboration limitations: Manual inspections and automated monitoring systems typically operate independently. Physical parameters (such as vibration and strain) collected by sensor networks lack effective correlation with visual damage information, making it difficult to establish a holistic understanding of the structural status. Insufficient environmental robustness: Temperature and humidity changes at the construction site can easily cause sensor data drift. Existing filtering algorithms mostly rely on static thresholds, and their adaptability to dynamic environmental interference needs to be improved. Limited depth of quantitative assessment: Although AI-based damage identification models have been applied, they rarely consider the impact of the sensor device's own status on data reliability in time series analysis. Visual inspection focuses more on damage location and ignores the automatic extraction of key quantitative parameters (such as crack width). Decision support is coarse-grained: Existing risk outputs often remain at the overall security level, lacking detailed guidance on component-level positioning identification and response instructions.

[0003] The above bottlenecks restrict the real-time performance and engineering applicability of the assessment results. There is an urgent need to achieve deep fusion of multimodal data, dynamic suppression of environmental interference, and generation of quantitative damage parameters through technological innovation. Summary of the Invention

[0004] The purpose of this invention is to provide an engineering safety assessment method and system based on artificial intelligence, which realizes real-time intelligent diagnosis and risk warning of engineering structure damage by integrating multimodal AI (sensor timing analysis + visual damage recognition), and solves the assessment lag problem caused by the traditional method's large reliance on manual labor and the fragmentation of multi-source data.

[0005] To achieve the above objectives, the present invention provides an engineering safety assessment method based on artificial intelligence, comprising the following steps: Step S1: collecting time series data of physical parameters of engineering structures through a distributed sensor network; Step S2, synchronously acquiring surface visual data of the engineering structure; Step S3: performing dynamic noise filtering on the physical parameter time series data and correcting the geometric distortion of the surface visual data; Step S4: Inputting the data processed in step S3 into a 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; The cross-modal attention mechanism is used to fuse temporal and visual features and output the risk part feature vector. Step S5: Calculate the safety risk index based on the risk location feature vector; Step S6: Output a three-level risk grade and a structured risk report based on the safety risk index. The report includes the location identification of the risk component, the damage type and quantitative parameters, and the graded response instructions.

[0006] Preferably, in step S1, the sensor types include strain sensors, displacement sensors and vibration sensors.

[0007] Preferably, in step S3, the dynamic noise filtering step is as follows: The collected engineering structure physical parameter time series signal is Layer decomposition, adaptively determine the wavelet denoising threshold according to signal characteristics to remove noise interference in 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, Represents the original physical parameter time series data, Indicates the environmental humidity of the project site. Indicates the ambient temperature of the project site. Represents the temperature coefficient of the sensor in microstrain per degree Celsius. Indicates the humidity coefficient of the sensor in microstrain per percent relative humidity.

[0008] Preferably, in step S4, the improved Mogrifier LSTM network performs four rounds of iterative optimization on the input vector and hidden state to act on the LSTM core calculation, as follows: The input vector corresponding to the time series data of the physical parameters of the engineering structure collected by the sensor , through the weight matrix After linear transformation, the hidden state is normalized with the layer Perform Hadamard product to get interactive transformation results ; Hidden State Weight matrix After transformation, Perform Hadamard product to get interactive transformation results , realizes the bidirectional interactive transformation of input vector and hidden state; among them, Indicates the hidden layer dimension set to meet the engineering data processing requirements, represents the input vector dimension, represents element-wise multiplication, Representation layer normalization operation; An adjustment coefficient is introduced, and the attenuation factor is calculated based on the ambient temperature of the project site and the service life of the sensor. The attenuation factor is then applied to the forget gate calculation of the LSTM: ; ; ; in, represents the output of the forget gate, represents the weight matrix of the forget gate, represents the bias term of the forget gate, represents the Sigmoid activation function, represents the concatenation vector, represents the attenuation factor, Indicates the ambient temperature of the project site. Indicates the service life of the sensor. represents the adjustment coefficient, Represents a very small constant.

[0009] Optimized, improved Mogrifier LSTM network outputs sensor health perception related parameters , predict the health index of the sensor: ; in, Represents the health index, Indicates the comprehensive impact coefficient of environment and service time on the sensor, , These are the results of hidden state updates after the first to fourth rounds of iterations; Correct the attenuation factor 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 again to the LSTM forget gate calculation.

[0010] Preferably, in step S4, the improved YOLOv8 network includes: When processing the feature map of the surface visual data of the engineering structure, the coordinate attention module operation is executed at the Backbone end; The detection head is decoupled into three branches, including: The damage location regression branch uses EIoU loss to output bounding box coordinates to locate the damage area for engineering structure damage; The damage type classification branch outputs nine types of damage probability distributions; The crack width regression branch outputs a pixel-level width map, identifies and quantifies surface damage of engineering structures from multiple dimensions such as location, type, and degree, and outputs processed visual data features.

[0011] Preferably, the crack width regression branch is implemented as follows: For the input engineering structure surface visual data feature map, it first performs channel expansion convolution to increase the number of feature map channels, and then performs pixel shuffling operation to obtain the feature map after pixel reorganization and upsampling; The smooth L1 loss function is used to optimize the model after pixel reorganization and upsampling of the feature map.

[0012] Preferably, in step S4, the temporal and visual features are fused through a cross-modal attention mechanism to output a risk part feature vector, specifically as follows: The timing characteristics obtained by the timing analysis branch , visual features obtained by the visual analysis branch Perform projection transformation respectively, build spatial constraint matrix based on the spatial characteristics of engineering structure, and then generate normalized correlation matrix ; Based on the normalized incidence matrix , performed by a gated recurrent unit , output the fused feature vector .

[0013] Preferably, in step S5, the risk index calculation steps are as follows: A three-layer fully connected network is used to process the eigenvectors of risk locations to obtain the damage scores of each component of the engineering structure. Based on the damage score, combined with the component weights and failure influencing factors determined in the engineering component design, the safety risk index of the engineering structure is calculated.

[0014] The present invention also provides an engineering safety assessment system based on artificial intelligence, comprising: include: Distributed sensor network module, used to collect time series data of physical parameters of engineering structures. Sensor types include strain sensors, displacement sensors, and vibration sensors. Visual acquisition module, used to synchronously acquire surface visual data of engineering structures; Data preprocessing module, including: Dynamic noise filtering unit, used for dynamic noise filtering processing of physical parameter time series data; A geometric distortion correction unit, used for correcting surface visual data; Multimodal artificial intelligence analysis module, including: The time series analysis unit uses an improved Mogrifier LSTM network to process sensor data. The network performs four rounds of iterative optimization of the input vector and hidden state, and calculates the attenuation factor based on the ambient temperature and the service life of the sensor, which is applied to the LSTM forget gate. The visual analysis unit uses an improved YOLOv8 network for damage identification. Its backbone terminal 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. Cross-modal fusion unit, used to fuse temporal features and visual features through a cross-modal attention mechanism to generate a risk part feature vector; The safety risk calculation module uses a three-layer fully connected network to process the risk location feature vectors and combines component weights with failure impact factors to calculate the safety risk index. The risk report generation module outputs three-level risk levels and a structured risk report based on the safety risk index. The report includes the location identification of risk components, damage types and quantitative parameters, and graded response instructions.

[0015] Therefore, the present invention adopts the above-mentioned engineering safety assessment method and system based on artificial intelligence, and the beneficial technical effects are as follows: (1) Improve assessment accuracy and real-time performance: Fusion of multi-source data, including physical parameter time series data collected by distributed sensor networks and surface visual data acquired synchronously, after processing, inputs into the multimodal artificial intelligence analysis model. The improved Mogrifier LSTM network and the improved YOLOv8 network process the time series data and visual data respectively, and fuse features through the cross-modal attention mechanism, so as to more accurately assess the safety risks of engineering structures. Compared with traditional methods, the accuracy and real-time performance are significantly improved, solving the assessment lag problem caused by the fragmentation of multi-source data in traditional methods. (2) Enhanced environmental adaptability: During the dynamic noise filtering process, the influence of temperature and humidity at the engineering site is taken into account, and compensation correction is performed on the denoised signal, so that the sensor data can better adapt to the complex and changeable engineering site environment, thereby improving the reliability of the data; (3) Deep quantification of damage parameters: The improved YOLOv8 network can not only locate damage, but also output nine types of damage probability distribution and pixel-level width maps. It can identify and quantify surface damage of engineering structures from multiple dimensions such as location, type, and degree, providing richer quantitative parameters for engineering safety assessment, which is conducive to more accurate judgment of the severity of damage and its impact on structural safety. (4) Ensuring the reliability of sensor data: The improved Mogrifier LSTM network can output sensor health perception related parameters, predict the sensor health index, and mark a fault when the sensor health index is lower than the threshold. At the same time, the attenuation factor is corrected and applied again to the LSTM forget gate calculation, thereby effectively considering the impact of the sensor's own state on data reliability, ensuring the quality of the sensor data input into the analysis model, and improving the stability and accuracy of the entire evaluation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of an engineering safety assessment method based on artificial intelligence of the present invention; Figure 2 This is a diagram of the architecture of the multimodal artificial intelligence analysis model; Figure 3 Detailed diagram of the improved YOLOv8 network. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0018] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0019] Example 1 like Figure 1 As shown in the figure, an engineering safety assessment method based on artificial intelligence is used, taking an actual bridge project in Beijing as an example. The specific implementation process is as follows: Step S1: collecting time series data of physical parameters of engineering structures through a distributed sensor network.

[0020] 150 strain sensors, 30 displacement sensors, and 50 vibration sensors were deployed in key locations on a large cable-stayed bridge, including its towers, main beams, and piers. These sensors were evenly distributed and focused on areas prone to problems, such as mid-span and near supports. The sensor sampling frequency was set to 100 Hz.

[0021] Step S2: synchronously acquiring surface visual data of the engineering structure.

[0022] Four high-definition cameras were installed on both sides of the bridge and in key locations: on the east, west, south, and north sides. With a resolution of 4K and a frame rate of 30fps, these cameras can clearly capture surface damage, such as cracks and spalling.

[0023] Step S3: performing dynamic noise filtering on the physical parameter time series data and correcting the geometric distortion of the surface visual data; (1) The steps of dynamic noise filtering are as follows: The collected time series signals of engineering structure physical parameters are decomposed into five layers, and the wavelet denoising threshold is adaptively determined based on the signal characteristics to remove noise interference in 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, Represents the original physical parameter time series data, Indicates the environmental humidity of the project site. Indicates the ambient temperature of the project site. Represents the temperature coefficient of the sensor in microstrain per degree Celsius. Indicates the humidity coefficient of the sensor in microstrain per percent relative humidity.

[0024] (2) The steps for geometric distortion correction are as follows: Correct the geometric distortion of the captured visual images. For example, use the Zhang Zhengyou calibration method to calibrate the cameras, obtaining each camera's intrinsic parameter matrix (focal length, principal point coordinates, etc.) and extrinsic parameter matrix (rotation matrix and translation vector). Then, use the undistort function in the OpenCV library to correct the image, eliminating image distortion caused by camera lens distortion.

[0025] Step S4: Figure 2 As shown, the data processed in step S3 is input into the multimodal artificial intelligence analysis model, specifically including: (1) The time series analysis branch uses an improved Mogrifier LSTM network to process sensor data; The improved Mogrifier LSTM network performs four rounds of iterative optimization on the input vector and hidden state to act on the LSTM core calculation, as follows: The input vector corresponding to the time series data of the physical parameters of the engineering structure collected by the sensor , through the weight matrix After linear transformation, the hidden state is normalized with the layer Perform Hadamard product to get interactive transformation results ; Hidden State Weight matrix After transformation, Perform Hadamard product to get interactive transformation results , realizes the bidirectional interactive transformation of input vector and hidden state; among them, Indicates the hidden layer dimension set to meet the engineering data processing requirements, represents the input vector dimension, represents element-wise multiplication, Representation layer normalization operation; An adjustment coefficient is introduced, and the attenuation factor is calculated based on the ambient temperature of the project site and the service life of the sensor. The attenuation factor is then applied to the forget gate calculation of the LSTM: ; ; ; in, represents the output of the forget gate, represents the weight matrix of the forget gate, represents the bias term of the forget gate, represents the Sigmoid activation function, represents the concatenation vector, represents the attenuation factor, Indicates the ambient temperature of the project site. Indicates the service life of the sensor. represents the adjustment coefficient, Represents a very small constant, with a value less than 10 -6 .

[0026] Improved Mogrifier LSTM network outputs sensor health perception related parameters , predict the health index of the sensor: ; in, Represents the health index, Indicates the comprehensive impact coefficient of environment and service time on the sensor, , These are the results of hidden state updates after the first to fourth rounds of iterations; Correct the attenuation factor 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 The LSTM forget gate calculation is applied again. In this embodiment, a strain sensor 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 normal working condition and the data is reliable.

[0027] (2) The visual analysis branch uses an improved YOLOv8 network for damage identification; like Figure 3 As shown, the improved YOLOv8 network includes: When processing the feature map of visual data on the surface of engineering structures, the coordinate attention module operation is executed at the Backbone end to enhance the model's ability to locate damaged areas by redistributing the channel and spatial information of the feature map.

[0028] The detection head is decoupled into three branches, including: The damage location regression branch uses EIoU loss to output bounding box coordinates to locate the damaged area for engineering structure damage. For example, the bounding box coordinates of a crack detected are (100, 150, 300, 400), indicating the location of the crack in the image. The damage type classification branch outputs nine types of damage probability distributions; These include cracks, spalling, exposed reinforcement, holes, honeycomb surface, steel bar corrosion, network cracks, concrete aging, and support damage. Taking a certain damaged area as an example, the classification results are as follows: Crack: The probability is 0.92, indicating that the damaged area is likely to be a crack, which may be caused by factors such as structural stress, temperature change, or material shrinkage.

[0029] Spalling: The probability is 0.05, indicating that there is a certain possibility of spalling. It may be that the protective layer on the concrete surface has fallen off due to aging or external impact.

[0030] Exposed rebar: The probability is 0.01, which is low, but attention should be paid to whether there is exposed rebar, which may be caused by insufficient or damaged concrete cover.

[0031] Holes: The probability is 0.01. The possibility of holes is small, but it is necessary to check whether there are voids inside the concrete, which may affect the integrity of the structure.

[0032] Honeycomb and pitting: The probability is 0.005, which is very low, but attention should be paid to whether there are honeycomb or pitting defects on the concrete surface, which may affect the durability of the structure.

[0033] Rebar corrosion: The probability is 0.003, which is very low, but you still need to be alert to signs of rebar corrosion, which may be caused by damage to the concrete cover or environmental corrosion factors.

[0034] Network cracks: The probability is 0.001, which is extremely low, indicating that the possibility of network cracks is very small and may be caused by concrete shrinkage or temperature changes.

[0035] Concrete aging: The probability is 0.001, which is extremely low, indicating that the concrete aging phenomenon is not obvious, but attention should still be paid to the impact of long-term environmental factors on concrete performance.

[0036] Support damage: The probability is 0.001, which is extremely low, but attention should be paid to whether the support has damage such as aging, deformation or displacement, which may affect the normal force transmission of the structure.

[0037] The crack width regression branch outputs a pixel-level width map, identifies and quantifies surface damage of engineering structures from multiple dimensions such as location, type, and degree, and outputs processed visual data features.

[0038] The steps to implement the crack width regression branch are as follows: The input feature map first undergoes channel dilation convolution, increasing the number of channels from 256 to 512. A pixel shuffling operation is then performed to obtain a pixel-wise upsampled feature map. The model is optimized using a smoothed L1 loss function, outputting a pixel-level width map. For example, in a crack region, the pixel values ​​in the width map range from 3 to 7, corresponding to an actual crack width of 3 to 7 mm. Analysis of the width map reveals the crack width distribution, providing a quantitative basis for assessing crack severity.

[0039] (3) By fusing temporal and visual features through the cross-modal attention mechanism, the risk part feature vector is output as follows: The timing characteristics obtained by the timing analysis branch , visual features obtained by the visual analysis branch Perform projection transformation respectively, build a spatial constraint matrix based on the spatial structural characteristics of the bridge and the connection relationship between the components, and then generate a normalized correlation matrix ; Based on the normalized incidence matrix , performed by a gated recurrent unit , output the fused feature vector .

[0040] Step S5: Calculate the safety risk index based on the risk location feature vector: A three-layer fully connected network is used to process the feature vectors of risky locations. The input feature vector has a dimension of 512. The first layer of the fully connected network maps it to 256 dimensions, the second layer to 128 dimensions, and the third layer to 64 dimensions. This yields damage scores for each component of the engineering structure. For example, a bridge pier has a damage score of 0.72, indicating relatively severe damage that could significantly impact the overall safety of the bridge. A main beam has a damage score of 0.45, indicating relatively minor damage but still requiring attention.

[0041] The safety risk index of the engineering structure is calculated based on the damage score, combined with the component weights and failure impact factors determined during the engineering component design. The weights and failure impact factors of each component are determined based on the bridge's design documents and structural analysis results. For example, the weight of a pier is 0.3 and its failure impact factor is 0.8; the weight of a main beam is 0.4 and its failure impact factor is 0.7. Substituting these into the formula to calculate the contribution of each component to the overall safety risk index, the pier's contribution is 0.72 × 0.3 × 0.8 = 0.1728, and the main beam's contribution is 0.45 × 0.4 × 0.7 = 0.126. Taking into account the contributions of all components, the overall safety risk index of the bridge is 0.65.

[0042] Step S6: Based on the safety risk index of 0.65 and the risk classification criteria (low risk 0-0.3, medium risk 0.3-0.7, high risk 0.7-1.0), the bridge is determined to be at a medium risk level. This indicates that the bridge currently has a certain safety risk and requires appropriate management and maintenance measures.

[0043] Data from three different time periods on the same bridge were compared using the traditional method (using only manual inspections and single sensor monitoring), the non-fusion method (using only sensor data without multimodal fusion), and the proposed method (an artificial intelligence analysis method that integrates multimodal data). The results are shown in Table 1: Table 1 Comparison results ;

[0044] Comparative results show that the proposed method significantly outperforms traditional and non-fusion methods in terms of data acquisition and processing efficiency, damage identification accuracy, and risk assessment accuracy. The proposed method requires only 8 hours of data acquisition and 15 minutes of processing, achieves 95% damage identification accuracy, and 90% risk assessment accuracy. It also comprehensively monitors sensor health and automatically measures crack width with an error of only ±0.5 mm.

[0045] Example 2 An engineering safety assessment system based on artificial intelligence, comprising: Distributed sensor network module, used to collect time series data of physical parameters of engineering structures. Sensor types include strain sensors, displacement sensors, and vibration sensors. Visual acquisition module, used to synchronously acquire surface visual data of engineering structures; Data preprocessing module, including: Dynamic noise filtering unit, used for dynamic noise filtering processing of physical parameter time series data; A geometric distortion correction unit, used for correcting surface visual data; Multimodal artificial intelligence analysis module, including: The time series analysis unit uses an improved Mogrifier LSTM network to process sensor data. The network performs four rounds of iterative optimization of the input vector and hidden state, and calculates the attenuation factor based on the ambient temperature and the service life of the sensor, which is applied to the LSTM forget gate. The visual analysis unit uses an improved YOLOv8 network for damage identification. Its backbone terminal 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. Cross-modal fusion unit, used to fuse temporal features and visual features through a cross-modal attention mechanism to generate a risk part feature vector; The safety risk calculation module uses a three-layer fully connected network to process the risk location feature vectors and combines component weights with failure impact factors to calculate the safety risk index. The risk report generation module outputs three-level risk levels and a structured risk report based on the safety risk index. The report includes the location identification of risk components, damage types and quantitative parameters, and graded response instructions.

[0046] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0047] Therefore, the present invention adopts the above-mentioned artificial intelligence-based engineering safety assessment method and system, and realizes real-time intelligent diagnosis and risk warning of engineering structure damage by integrating multimodal AI (sensor timing analysis + visual damage recognition), thereby solving the assessment lag problem caused by the traditional method's large manual reliance and multi-source data fragmentation.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An engineering safety assessment method based on artificial intelligence, characterized in that: The following steps are involved: Step S1: collecting time series data of physical parameters of engineering structures through a distributed sensor network; Step S2, synchronously acquiring surface visual data of the engineering structure; Step S3: performing dynamic noise filtering on the physical parameter time series data and correcting the geometric distortion of the surface visual data; Step S4: Inputting the data processed in step S3 into a 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; The cross-modal attention mechanism is used to fuse temporal and visual features and output the risk part feature vector. Step S5: Calculate the safety risk index based on the risk location feature vector; Step S6: Output three-level risk levels and a structured risk report based on the security risk index.

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 collected engineering structure physical parameter time series signal is Layer decomposition, adaptively determine the wavelet denoising threshold according to signal characteristics to remove noise interference in 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, Represents the original physical parameter time series data, Indicates the environmental humidity of the project site. Indicates the ambient temperature of the project site. represents the temperature coefficient of the sensor, 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 Mogrifier LSTM network performs four rounds of iterative optimization on the input vector and hidden state to act on the LSTM core calculation, as follows: The input vector corresponding to the time series data of the physical parameters of the engineering structure collected by the sensor , through the weight matrix After linear transformation, the hidden state is normalized with the layer Perform Hadamard product to get interactive transformation results ; Hidden State Weight matrix After transformation, Perform Hadamard product to get interactive transformation results , realizes the bidirectional interactive transformation of input vector and hidden state; among them, Indicates the hidden layer dimension set to meet the engineering data processing requirements, represents the input vector dimension, represents element-wise multiplication, Representation layer normalization operation; An adjustment coefficient is introduced, and the attenuation factor is calculated based on the ambient temperature at the project site and the service life of the sensor. The attenuation factor is then applied to the forget gate calculation of the LSTM: ； ； ； in, represents the output of the forget gate, represents the weight matrix of the forget gate, represents the bias term of the forget gate, represents the Sigmoid activation function, represents the concatenation vector, represents the attenuation factor, Indicates the ambient temperature of the project site. Indicates the service life of the sensor. represents the adjustment coefficient, Represents a very small constant.

5. The engineering safety assessment method based on artificial intelligence according to claim 4 is characterized in that: Improved Mogrifier LSTM network outputs sensor health perception related parameters , predict the health index of the sensor: ； in, Represents the health index, Indicates the comprehensive impact coefficient of environment and service time on the sensor, , These are the results of hidden state updates after the first to fourth rounds of iterations; Correct the attenuation factor 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 again to the LSTM forget gate calculation.

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

7. The engineering safety assessment method based on artificial intelligence according to claim 6 is characterized in that: The steps to implement the crack width regression branch are as follows: For the input engineering structure surface visual data feature map, it first performs channel expansion convolution to increase the number of feature map channels, and then performs pixel shuffling operation to obtain the feature map after pixel reorganization and upsampling; The smooth L1 loss function is used to optimize the model after pixel reorganization and upsampling of the feature map.

8. The engineering safety assessment method based on artificial intelligence according to claim 1 is characterized in that: In step S4, the temporal and visual features are fused through the cross-modal attention mechanism to output the risk part feature vector, as follows: The timing characteristics obtained by the timing analysis branch , visual features obtained by the visual analysis branch Perform projection transformation respectively, build spatial constraint matrix based on the spatial characteristics of engineering structure, and then generate normalized correlation matrix ; Based on the normalized incidence matrix , performed by a gated recurrent unit , output the fused feature vector .

9. The engineering safety assessment method based on artificial intelligence according to claim 1 is 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 eigenvectors of risk locations to obtain the damage scores of each component of the engineering structure. Based on the damage score, combined with the component weights and failure influencing factors determined in the engineering component design, the safety risk index of the engineering structure is calculated.

10. An engineering safety assessment system based on artificial intelligence, characterized in that: include: Distributed sensor network module, used to collect time series data of physical parameters of engineering structures. Sensor types include strain sensors, displacement sensors, and vibration sensors. Visual acquisition module, used to synchronously acquire surface visual data of engineering structures; Data preprocessing module, including: Dynamic noise filtering unit, used for dynamic noise filtering processing of physical parameter time series data; A geometric distortion correction unit, used for correcting surface visual data; Multimodal artificial intelligence analysis module, including: The time series analysis unit uses an improved Mogrifier LSTM network to process sensor data. The network performs four rounds of iterative optimization of the input vector and hidden state, and calculates the attenuation factor based on the ambient temperature and the service life of the sensor, which is applied to the LSTM forget gate. The visual analysis unit uses an improved YOLOv8 network for damage identification. Its backbone terminal 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. Cross-modal fusion unit, used to fuse temporal features and visual features through a cross-modal attention mechanism to generate a risk part feature vector; The safety risk calculation module uses a three-layer fully connected network to process the risk location feature vectors and combines component weights with failure impact factors to calculate the safety risk index. The risk report generation module outputs three-level risk levels and a structured risk report based on the safety risk index. The report includes the location identification of risk components, damage types and quantitative parameters, and graded response instructions.

Images