A building structure anomaly detection method based on internal and external defect cooperation

Abstract

The application provides a kind of building structure anomaly detection method based on internal and external defect cooperation, belongs to the abnormal detection technical field based on deep learning;First, construct digital twin, collect external defect image data and internal key node sensor data, and construct building anomaly structure representation dataset;Second, design an identification module containing an external identification channel and an internal identification channel, and output internal and external defect identification results;Third, build type-structure and position-structure datasets through digital twin simulation, train the mapping model, and obtain the sensitivity table of the quantitative defect type / position and the structure anomaly correlation strength combined with SHAP values;Finally, a joint sensitivity table is constructed, the structure anomaly probability is calculated based on the identification results, the overall early warning factor is obtained through adaptive weight and cooperative amplification factor, and three-level real-time early warning is realized based on the preset threshold;The application realizes accurate evaluation and adaptive response of building structure anomaly.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of deep learning-based anomaly detection, and particularly relates to a building structure anomaly detection method based on internal and external defect cooperation. BACKGROUND

[0002] With the continuous expansion of urban construction scale and the continuous emergence of high-rise and complex structure buildings, the traditional monitoring and maintenance means have been difficult to meet the demand for high-frequency and high-accuracy state evaluation of buildings in the whole life cycle. Building structures often face the joint action of various environmental and load factors in the service process, such as earthquake disturbance, wind load impact, foundation settlement and material aging, etc. These factors are prone to induce structure anomalies that are difficult to be directly observed, including micro-crack expansion, stress concentration, component connection loosening and other hidden diseases. Due to the limitations of artificial inspection cycle, instrument deployment range and subjective judgment, the traditional means are often difficult to identify potential risks in time, causing safety hazards to accumulate. The development of artificial intelligence technology provides a new solution path for building health monitoring. Existing building structure anomaly detection methods include the following methods:

[0003] (1) Artificial inspection-based method: artificial inspection is one of the most traditional and widely used building structure anomaly detection methods. Usually, professional inspectors regularly inspect the key parts of the building (such as columns, beams, connection nodes, etc.) on site or with the help of auxiliary tools for detection, such as crack gauges, hammering echoes, etc. This method is simple to operate, low in cost, suitable for early rough screening, but has the problems of long detection cycle, low efficiency and strong subjectivity, which is difficult to meet the demand for high-frequency and high-precision monitoring of modern complex buildings.

[0004] (2) Sensor monitoring-based method: this method collects real-time structural response data by deploying various sensors (such as strain gauges, accelerometers, optical fiber sensors, etc.) in the building structure, to monitor the changes of strain, displacement, vibration, etc. of the structure under load, and then judge whether there is an anomaly. This method has the advantages of strong real-time performance and high accuracy, and is widely used in bridges, high-rise buildings and other key infrastructure. However, its initial deployment cost is relatively high, and the maintenance is complex, and the monitoring range is limited by the deployment density of sensors.

[0005] (3) Image analysis and machine learning-based building structure anomaly detection method: With the development of intelligent technology, image analysis and machine learning-based building anomaly detection methods have attracted increasing attention. This method collects building images through unmanned aerial vehicles, cameras and other devices, combines edge detection, texture analysis and other image processing methods with traditional machine learning algorithms (such as SVM, K-nearest neighbor, random forest, etc.), and realizes automatic identification of defects such as cracks and corrosion. Compared with deep learning, this method has less dependence on data, flexible deployment, and is suitable for building primary intelligent inspection in resource-constrained scenarios, with good practicality and expandability. Although the method based on image analysis and traditional machine learning has realized preliminary automation, its feature extraction capability is limited, it is sensitive to environmental conditions, and it is difficult to handle the complexity and diversity of structural defects.

[0006] In summary, the existing methods generally lack the ability to fuse and analyze multi-source heterogeneous data and the ability to intelligently perceive complex structural risks, making it difficult to meet the high-precision, real-time and robustness requirements for building structure safety in complex environments. SUMMARY

[0007] To solve the above problems, the present application provides a building structure anomaly detection method based on internal and external defect collaboration, comprising the following steps:

[0008] S1, real-time acquisition of building external surface images; and based on the constructed building digital twin, real-time acquisition of multiple internal perception data of M key building nodes;

[0009] S2, inputting the data obtained in S1 into the trained building internal and external defect identification model, the model including a double-channel building defect identification module, an external structure defect identification channel inputting the external surface images, adopting a multi-scale external defect perception channel, outputting the detection results of the building external defects through multi-resolution feature fusion; an internal structure defect identification channel inputting the internal perception data, adopting a graph data calculation and fusion unit for component topological correlation characteristics, mining internal hidden defects based on component connection relationship, and outputting the identification results of the building internal defects;

[0010] S3, based on the constructed joint sensitivity table, combining the two identification results of S2 to calculate the multi-class structure anomaly probability, obtaining the overall early warning factor through adaptive weight and collaborative amplification factor, and obtaining multi-level early warning based on a preset threshold;

[0011] The joint sensitivity table is obtained by constructing a building defect risk degree prediction dataset through digital twinning, obtaining type-structure mapping distribution and location-structure mapping distribution, combining SHAP value to obtain type-structure anomaly sensitivity table and location-structure anomaly sensitivity table, and jointly constructing a joint sensitivity table.

[0012] Preferably, the collection method of external structure data for training the building internal and external defect identification model is:

[0013] For the selected target building, the parameter information related to the building external contour and facade structure in the BIM model is retrieved, including building height, external facade material, component distribution data information; then the digital twin is constructed and the unmanned aerial vehicle is preset to collect the image surface picture;

[0014] During data labeling, the defect area of the building external surface is labeled by using a rectangular box, and on this basis, a structured label information is added for each labeled defect, including defect type Eer, specifically wall surface crack Eer 1 , external wall leakage Eer 2 , external component peeling Eer 3 , parapet wall cracking Eer 4 , door and window frame deformation Eer 5 , and defect occurrence position Eew, the position coordinates are (x, y, z).

[0015] Preferably, the collection method of internal structure data for training the building internal and external defect identification model is:

[0016] At the two ends of the concrete column, the two ends of the concrete beam, the connection joint of the steel structure, the beam-column joint, the main-secondary beam joint, the floor base, the floor span, and the root of the building cantilever component of the selected target building, M key building nodes are deployed to collect stress data Str1, vibration amplitude data Str2 and vibration frequency data Str3 are collected by deploying a vibration pickup, corrosion current density data Str4 of steel bars are collected by deploying an electrochemical sensor, building material reflection wave signal characteristics Str5 are collected by deploying an ultrasonic sensor, and static displacement Str6 is collected by deploying a laser displacement meter; and Str1, Str2, Str3, Str4, Str5, Str6 are combined to form building comprehensive perception data Stre j ;

[0017] During data labeling, the defect type Ier of the building internal structure defect includes internal honeycomb Ier 1 , internal structural crack Ier 2 , internal stress concentration of steel structure Ier 3 , prestress relaxation Ier 4 , foundation pile displacement Ier 5 , steel bar corrosion Ier 6 , and grouting loosening Ier 7 , and the defect occurrence position Iew, the defect occurrence position coordinates are (x, y, z).

[0018] Preferably, the external structure defect identification channel is specifically:

[0019] The input image data Datawin is first subjected to preliminary feature extraction of the building external surface image by the external defect detail perception unit to obtain a feature Fea1, and then subjected to the Inception feature extraction module to capture features of different sizes in parallel through multi-scale convolution kernels to calculate an output feature map Fea2;

[0020] Subsequently, the feature Fea2 is sent to the image attention segmentation unit: first, the feature Fea2 is subjected to local gradient amplitude calculation by the Sobel operator to generate an initial spatial attention mask, and the higher the gradient of the region, the stronger the corresponding relationship with the defect edge; subsequently, the feature is tilted to the feature region corresponding to the high gradient region through the spatial attention mask guided by the bidirectional gate mechanism, and finally an enhanced defect feature map Fea3 is output;

[0021] Subsequently, Fea3 is sent to the external defect detail parallel perception unit for external defect identification processing to calculate Feab1, Feab2, Feab3, and Feab4, respectively. The external defect detail parallel unit includes four external defect detail perception units arranged in parallel; and the image channel attention unit is used to complete adaptive integration of the features Feab1, Feab2, Feab3, and Feab4 to obtain a fused feature Fear.

[0022] The fused feature Fear is subjected to 3x3 convolution layer calculation again in sequence to obtain a feature Feawr, and then subjected to Dropout layer and two cascaded KNN classification layers for final processing to finally output the detection result Decwout of the building external defect.

[0023] Preferably, the internal structure defect identification channel is specifically:

[0024] First, the building abnormal structure internal defect data Dataiin is input into the graph data construction unit. In the graph data construction unit, the two ends of each concrete column, the two ends of each concrete beam, the connection nodes of steel structures, the beam-column nodes, the main-secondary beam joints, the floor base, the floor midspan, and the root of the building cantilever member are regarded as the nodes of the graph data, and the node features are the building comprehensive perception data Stre j collected by each group of sensors, j∈[1, M]; the building physical connection relationship between the sensor deployment positions is constructed as an edge, thereby forming a building representation graph data Gradata with physical topological prior;

[0025] Subsequently, the building representation graph data Gradata is input into three groups of structurally differentiated graph convolution channels at the same time, each group of channels including three layers of graph convolution operations, respectively, a graph convolution path based on Chebyshev polynomial is used to extract high-order features Feag1, a graph convolution path based on a neighbor aggregation mechanism is used to extract heterogeneous neighborhood features Feag2, and a standard graph convolution path is used to extract basic structural features Feag3;

[0026] The features Feag1, Feag2 and Feag3 are subjected to multi-channel fusion in a feature splicing layer and are subjected to nonlinear mapping through an Elu activation layer to obtain intermediate features Feagz; subsequently, the features Feagz are sent to a deep graph convolution fusion module, two layers of Chebyshev graph convolution calculation are performed to obtain features Feaq, two layers of aggregation-type graph convolution calculation are performed to obtain features Feaj, and two layers of standard graph convolution calculation are performed to obtain features Feab;

[0027] On this basis, an internal-external defect interaction unit is introduced to fuse the features Feawr calculated in the building external structural defect recognition path and the features Feaq, Feaj and Feab to realize the correlation modeling between the internal state and the external performance of the structure; in the internal-external defect interaction unit, two mutually connected channel attention layers are included, in the first three-channel attention layer, the features Feaq, Feaj and Feab are adaptively fused to obtain Fean1, and then a double-channel attention layer is used to internally fuse the features Fean1 and the features Feawr to output the features Feazsr.

[0028] After the interaction output features Feazsr are subjected to ReLU activation function and Dropout layer processing, the building internal defect recognition result Deciout is finally given by the output layer.

[0029] Preferably, the type-structure mapping distribution and the location-structure mapping distribution are obtained by constructing a building defect risk degree speculation dataset through digital twinning, and the specific process is as follows:

[0030] Firstly, the following several structural anomaly types Su are defined: local buckling, member load capacity decline, material performance degradation, structural deviation, uneven interlayer stiffness and local high stress aggregation; then, based on the constructed building digital twin, the building operating conditions when the building appears the above several types of structural anomalies are simulated, and the following data are recorded respectively:

[0031] Whether each type of external defect occurs is recorded, and the external defect type data Esd is obtained; similarly, the internal defect type data Isd is obtained, and the Esd and Isd are spliced to obtain the complete defect type state data Csd;

[0032] Record the position of the external defect occurrence Eld and the position of the internal defect occurrence Ild, and splice Eld and Ild to obtain complete defect position data Cld;

[0033] Based on the above method, the data set is constructed, the building defect risk degree mapping model is trained, and the model parameters Param for type-structure after training are obtained cls and the model parameters Param for position-structure loc .

[0034] Preferably, the type-structure abnormal sensitivity table and the position-structure abnormal sensitivity table are obtained in combination with the SHAP value, and the specific process is as follows:

[0035] Based on the parameters Param cls , the SHAP algorithm is used to calculate the SHAP value of each defect type in the process of judging each structure abnormal type, and the original SHAP sensitivity value is calculated by the minimum-maximum normalization formula to obtain the sensitivity value, which is in the range of [-1, 1], and the greater the value, the greater the positive contribution, and vice versa, the greater the negative contribution; the sensitivity values of all types to each structure abnormality form a type-structure abnormality sensitivity table Sen cls ;

[0036] Similarly, based on the parameters Param loc , the SHAP algorithm is used to calculate the SHAP value of each defect position in the process of judging each structure abnormal type, and the original SHAP sensitivity value is calculated by the minimum-maximum normalization formula to obtain the sensitivity value; the sensitivity values of all positions to each structure abnormality form a position-structure abnormality sensitivity table Sen loc .

[0037] Preferably, the joint sensitivity table is constructed by an unsupervised optimization method:

[0038] The optimization objectives include:

[0039] The structure abnormality sensitivity minimum fluctuation objective: by weighting the type sensitivity and the position sensitivity, for each structure abnormality, the average fluctuation value of the joint sensitivity output of the abnormality under different data inputs is calculated, and the objective term is the average fluctuation degree of all structure abnormality sensitivity outputs;

[0040] The structure abnormality sensitivity maximum discrimination objective: in the constructed joint sensitivity table, for each defect type-position combination, the sensitivity value distribution of each structure abnormality is observed; the discrimination degree of the distribution is calculated, and the discrimination objective is the weighted average value of the difference, and the objective is to make it as large as possible in optimization to ensure that the model can distinguish different structure abnormality risk characteristics;

[0041] Based on the above two optimization objectives, the particle swarm optimization algorithm is selected for unsupervised solution, and finally the joint sensitivity table Sen is obtained, which can fully represent the two risk sensitivity distributions jit .

[0042] Preferably, the S3 specific process includes:

[0043] The identified external defect recognition result Decwout and internal defect recognition result Deciout are integrated to obtain the complete defect recognition result Det cmb ;

[0044] According to the actual defect type and location, the corresponding sensitivity value is extracted from the joint sensitivity table Sen jit , and a weighted combination is made to construct the risk response score Sc1, Sc2,..., Sc6 of each type of structural anomaly, respectively corresponding to local buckling, member load capacity reduction, material performance degradation, structural displacement, interlayer stiffness unevenness and local high stress concentration;

[0045] The above score is converted into the probability Ph1, Ph2, …, Ph6 of structural anomaly occurrence through a soft classification mechanism;

[0046] According to the structural anomaly probability Ph1, Ph2,..., Ph6, an adaptive probability weight W w is assigned to each structural anomaly probability, w∈[1, 6], and then a weighted sum is obtained to obtain the basic severity Sev base ;

[0047] The number Num fau of anomalies in the probability set probability Ph1, Ph2,..., Ph6 is counted, and the synergistic amplification factor Emk=1+0.2×(Num fau -1);

[0048] Then, the basic severity Sev base and the synergistic amplification factor are multiplied to obtain the final building structure overall early warning factor Sev fin ;

[0049] Based on the early warning factor Sev fin and the set safety threshold, a multi-level real-time warning is obtained.

[0050] Preferably, a three-level building real-time warning mechanism is constructed, and the safety threshold Thr safe , the alarm threshold Thr ris and the emergency alarm threshold Thr best are set; according to the building structure overall early warning factor Sev fin calculated in real time and the above threshold, when Sevfin Less than Thr safe At this time, the building is in a safe state, and no early warning signal is issued; when Sev fin Greater than or equal to Thr safe And less than Thr ris A slight early warning signal is issued, when Sev fin Greater than or equal to Thr ris And less than Thr best An intermediate early warning signal is issued, when Sev fin Greater than or equal to Thr best An emergency early warning signal is issued.

[0051] Compared with the prior art, the present application has the following beneficial effects:

[0052] (1) Internal and external defect joint representation and double-path recognition mechanism: This mechanism breaks through the limitation of traditional early warning of internal and external defect separation recognition, can accurately identify visible defects of external facade, and can also excavate hidden defects in component topological correlation, so that the intelligent early warning system can more comprehensively perceive the abnormality of building structure, reduce the early warning omission caused by information fragmentation, and improve the accuracy of early warning.

[0053] (2) Defect recognition mechanism of double-path cooperation: An external multi-scale feature fusion path and an internal graph network topology perception path are designed, and the dynamic correlation of internal and external defect features is realized through an interaction unit, so as to improve the recognition ability of hidden and associated defects. Solve the problem that internal structural defects are difficult to accurately capture in traditional early warning, so that the early warning system can discover potential risk points earlier, especially for external representation abnormalities caused by internal defects, and realize early warning.

[0054] (3) Joint sensitivity driven risk early warning mechanism: a joint sensitivity table integrating defect type and position correlation strength is constructed, the overall early warning factor is calculated by combining adaptive weight and synergistic amplification factor, and dynamic grading response is realized through three-level threshold. Solve the problem that multiple defect concurrent risks are difficult to quantify in traditional evaluation, so that the early warning system can accurately capture superimposed risks, especially for coupled structural abnormalities, and realize targeted early warning. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The overall implementation process flowchart of the present application.

[0056] Figure 2 The overall framework diagram of the building defect recognition module.

[0057] Figure 3 The accuracy model performance comparison chart in the embodiment.

[0058] Figure 4 The PR-AUC model performance comparison chart in the embodiment.

[0059] Figure 5 Defect joint risk degree heat map for the embodiment. DETAILED DESCRIPTION

[0060] The present application proposes a building structure anomaly detection method based on internal and external defect cooperation, and the overall technical route flow chart is as shown in Figure 1 The method comprises:

[0061] Building anomaly structure representation dataset construction: in order to realize the accurate representation of the defects of the internal and external structures of the building, according to the difference between the identification methods of the two, the building external structure defect identification dataset and the digital twin driven internal structure defect identification dataset are constructed respectively, wherein the digital twin dataset maps the internal component defect features through building three-dimensional modeling, forming an anomaly representation system covering the building structure;

[0062] Building internal and external defect identification model construction: a building defect identification module containing double channels is designed to adapt to the sensing needs of internal and external defects. In the external structure defect identification channel, a multi-scale external defect sensing channel is innovatively designed to improve the recognition accuracy of complex external facade defects through multi-resolution feature fusion; in the internal structure defect identification channel, a graph data calculation and fusion unit is innovatively designed according to the component topology correlation characteristics, and internal hidden defects are mined based on the component connection relationship;

[0063] Risk sensitivity table based on building defect risk degree mapping model: this module first constructs a dataset of defect types, locations and structure anomalies through digital twinning; then uses LightGBM to train the mapping model, and combines SHAP algorithm to obtain a sensitivity table of quantitative risk association; finally, the quantitative evaluation of defect and structure anomaly mapping is realized, and the risk degree reference basis for early warning is provided;

[0064] Building structure anomaly early warning unit design: this early warning unit first constructs a joint sensitivity table of defect types and locations through unsupervised optimization, and fuses the influence of the two on structure anomalies; then combines the actual defect identification results, the overall early warning factor, and finally realizes three-level real-time early warning according to the preset threshold, completing the closed loop from risk perception to graded response.

[0065] The specific implementation process of the present application will be further described in combination with specific embodiments.

[0066] I. Building anomaly structure representation dataset construction

[0067] The present application aims to accurately perceive various defects existing in the use process of buildings. Since the external defects of buildings are the surface representation of internal defects, simultaneously accurately perceiving internal and external structural defects helps to make accurate assessment of the overall defects of buildings. Based on this, in the construction phase of the building abnormal structure representation data set, the data collected by the present application includes building internal structure defect identification data and building external structure defect identification data;

[0068] 1. Building digital twin construction: The present application aims to achieve accurate identification of internal defects by analyzing sensor data of key nodes of buildings. However, the existing technology has obvious limitations: on the one hand, the number and dimension of sensor records of existing buildings are insufficient, making it difficult to directly establish the mapping relationship between sensor data and internal structural defects; on the other hand, for the built entity buildings, even if a sensor network is deployed to collect various state information, it is difficult to safely and controllably simulate internal structural defect scenarios, and it is difficult to ensure the consistency of building internal and external defect identification data collection.

[0069] To solve the above problems and ensure the convenience and safety of data collection scenarios, the present application is based on the BIM model of the building, constructs a high-precision building digital twin on the Autodesk Revit platform, and relies on the digital twin to carry out subsequent data collection work.

[0070] 2. Building external structural defect identification data collection:

[0071] (1) Building external image collection: For the selected target building, the parameter information related to the building contour and facade structure in the BIM model of the building is retrieved, including building height, external facade material, component distribution data information. Then, a drone is preset in the digital twin constructed by Autodesk Revit and the collection of image surface pictures is carried out. After the collection is completed, the pictures are filtered by the external facade model of the digital twin, and only the building surface pictures with external defects are retained. Finally, N building surface pictures Fig i ,i∈[1,N] are collected as building abnormal structure external defect input data Datawin;

[0072] (2) Building external structural defect labeling: Based on the building surface pictures obtained through the above image collection process, the present application organizes professionals in the field of building to carry out professional image labeling work. The labeling process is as follows: first, a rectangular frame is used to accurately label the defect area of the building surface, and on this basis, structured label information is added to each labeled defect, including defect type Eer 1 , external wall leakage Eer 2 , external component peeling Eer 3 , parapet wall cracking Eer4 and the door and window frame deformation Eer 5 , defect occurrence position Eew, specifically defect occurrence position coordinates (x, y, z). The coordinate system takes the intersection point of the outer edge line of the outer wall at the lower left corner of the first floor plan of the building as the coordinate origin (O); the X-axis is parallel to the longitudinal axis of the building (i.e. the length direction of the building), and the positive direction extends from the origin to the other end of the building; the Y-axis is parallel to the transverse axis of the building (i.e. the width direction of the building), and the positive direction extends from the origin to the other side of the building; the Z-axis is the vertical height, with the ground elevation of the first floor as Z=0, and the upward direction as the positive direction; the coordinate units are all meters (m).

[0073] (3) Based on the building exterior structure defect labeling method described in the above (2), the N building exterior surface images Fig i are sequentially labeled and integrated, and the labeled images are integrated to obtain building abnormal structure exterior defect output data Datawout;

[0074] 3. Building interior structure defect identification data collection:

[0075] (1) Sensor perception network arrangement and data collection: based on the digital twin constructed in the above process, piezoelectric sensors are deployed at the two ends of the concrete column, the two ends of the concrete beam, the connection nodes of the steel structure, the beam-column joints, the main and secondary beam intersections, the floor base and the floor span, and the roots of the building cantilever members to collect stress data Str1, vibration amplitude data Str2 and vibration frequency data Str3 are collected by deploying vibration pickups, corrosion current density data Str4 of steel bars are collected by deploying electrochemical sensors, ultrasonic wave sensor is deployed to collect building material reflection wave signal characteristics Str5, static displacement Str6 is collected by deploying laser displacement meter. And Str1, Str2, Str3, Str4, Str5, Str6 are combined to form building comprehensive perception data Stre j ,

[0076] (2) Building interior structure defect scene simulation: after completing the sensor perception network arrangement described in the above process, the building digital twin constructed in the above process is used to simulate the interior structure defect scene for data collection. And record the defect type Ier of the building interior structure defect at this time, which includes the interior honeycomb Ier 1 of the concrete structure, the interior structural crack Ier 2 of the component, the interior stress concentration Ier 3 of the steel structure, the prestress relaxation Ier 4 , foundation pile body displacement Ier 5 , steel bar corrosion Ier6 and loose grouting 7 The defect location is specified as Iew, specifically the coordinates (x, y, z). This coordinate system uses the intersection of the outer edges of the outer walls at the lower left corner of the building's first-floor plan as its origin (O). The X-axis is parallel to the building's longitudinal axis (length direction), extending positively from the origin to the other end of the building. The Y-axis is parallel to the building's transverse axis (width direction), extending positively from the origin to the other side of the building. The Z-axis represents vertical height, with the first-floor ground level at Z=0, and upwards as the positive direction. All coordinate units are meters (m).

[0077] Both Ier and Iew are arrays because the number of internal structural defects is v when simulating internal structural defects. Ier = [Ier1, Ier2, ..., Ier...] v ], Iew=[Iew1, Iew2,..., Iew v The data Ier and Iew are then combined to form the output data Dataiout,Dataiout=[Ier,Iew] for internal defects in the building's abnormal structure.

[0078] Simultaneously, record the comprehensive building perception data St collected from M building nodes when internal structural defects occur. re j, and combine the M sets of building comprehensive perception data to form the input number of internal defects in the abnormal building structure, Dataiin, where Dataiin = [Stre 1 Stre 2 ,...,Stre M ].

[0079] 4. Acquisition of abnormal building structure characterization data: Combine the above-mentioned external defect input data Datawin and internal defect input data Dataiin of abnormal building structure to obtain a complete input data Datain for a building abnormal structure characterization data, Datain = [Dataiin, Datawin];

[0080] The external defect output data (Datawout) and internal defect output data (Dataiout) of the above-mentioned abnormal building structure are combined to obtain a complete output data (Dataout) representing an abnormal building structure, where Dataout = [Dataiout, Datawout]. A complete output data (Databz) representing an abnormal building structure is formed by concatenating the data (Datain) and data (Dataout), where Databz = [Datain, Dataout].

[0081] 5、Building abnormal structure characterization dataset construction: based on the above specific description of the data acquisition method, the present application has carried out abnormal structure characterization data acquisition work on T building in total, and the building abnormal structure characterization data Databz t , t e [1, T] are integrated, so as to complete the construction process of building abnormal structure characterization dataset Data, Data = [Databz1, Databz2,..., Databz T ].

[0082] II. Building internal and external defect identification model construction

[0083] The building internal and external defect identification model designed by the present application comprises building external abnormal structure identification path and building internal abnormal structure identification path respectively, and since the external structure abnormality of the building and the internal structure abnormality of the building have a whole structure relationship, the internal and external structure abnormality interaction path is also innovatively designed in the building abnormal structure identification module to further improve the identification accuracy of the building structure abnormality. The overall framework diagram of the model is shown in Figure 2 .

[0084] 1、Building external structure defect identification path design: in the building external structure abnormality identification path, the input is the collected building abnormal structure external defect input image data Datawi n , then the external defect detail perception unit designed by the present application is used to preliminarily extract the features of the building external surface image, and the feature Fea1 is obtained. On the basis of this feature extraction, in order to further enhance the efficiency of building external surface feature extraction, the Inception feature extraction module is introduced into the building external structure abnormality identification path designed by the present application. The module captures features of different sizes in parallel through multi-scale convolution kernels, and calculates to obtain the output feature map Fea2.

[0085] Then the feature Fea2 is sent to the image attention segmentation unit for further processing to strengthen the specificity of the defect feature: first, the local gradient amplitude is calculated by Sobel operator on the Inception output feature Fea2 to generate an initial spatial attention mask, and the higher the gradient of the region, the stronger the corresponding relationship with the defect edge; then the spatial attention mask guides the features to tilt to the feature region corresponding to the high gradient region through the bidirectional gate mechanism, and finally the enhanced defect feature map Fea3 is output.

[0086] The output enhanced defect feature map Fea3 is then sent to an external defect detail parallel perception unit for further external defect recognition processing to obtain Feab1, Feab2, Feab3, and Feab4, respectively. The external defect detail parallel unit includes four external defect detail perception units arranged in parallel. The image channel attention unit designed in the application is used to complete the adaptive integration of the features Feab1, Feab2, Feab3, and Feab4, and obtain the fused feature Fear.

[0087] The fused feature Fear is sequentially subjected to a 3x3 convolution layer (the channel number is adjusted to 128 to retain key features) to obtain the feature Feawr, and then subjected to a Dropout layer (dropout rate 0.3 to suppress overfitting) and two cascaded KNN classification layers for final processing (the first layer K=5 to screen candidate defect regions, and the second layer K=3 to accurately determine the defect category), and finally outputs the detection result Decwout of the building external defect.

[0088] In the external defect detail perception unit, two parallel feature processing sub-paths are included: feature processing sub-path one first performs preliminary feature extraction through a 5x5 hollow convolution layer, and then completes feature standardization through a batch normalization (BN) layer and introduces a nonlinear transformation through a ReLU activation function; feature processing sub-path two performs preliminary feature extraction through a 3x3 hollow convolution layer, and then completes standardization and nonlinear activation through a batch normalization (BN) layer and a Silu activation function in sequence. The two features are fused through a channel splicing operation, and then subjected to feature dimension adjustment and depth calculation through a 1x1 convolution layer to finally form a complete external defect detail perception unit.

[0089] The construction process of the image channel attention unit is as follows: the input multi-channel features are first sent to a fully connected layer to complete feature fusion, and the fused features are processed in two parallel paths: path one sequentially compresses the spatial dimension through a max-pooling layer and extracts local correlation features through a 3x3 convolution layer, and path two aggregates global information through an average-pooling layer and is also processed through a 3x3 convolution layer; the output features of the two paths are sent to a channel attention layer for weighted fusion, and finally the feature depth is integrated through a 3x3 convolution layer to form a complete image channel attention unit.

[0090] 2. Building internal structure defect recognition path design: in the application, for the recognition task of building internal structure defects, a multi-path feature fusion recognition path based on a graph neural network is proposed.

[0091] The path first inputs the building internal defect structure data Dataiin into the graph data construction unit, in which each end of the concrete column, each end of the concrete beam, the connection node of the steel structure, the beam-column joint, the main and secondary beam intersection, the floor base and the floor span, and the root of the building cantilever component are regarded as the nodes of the graph data, and the node features are the building comprehensive perception data Stree collected by each group of sensors. j , j ∈ [1, M]; and the edges are constructed according to the building physical connection relationship between the sensor deployment positions, thereby forming the building representation graph data Gradata with physical topology priori;

[0092] Subsequently, the building representation graph data Gradata is simultaneously input into three groups of structure-differentiated graph convolution paths, each of which includes three layers of graph convolution operations, i.e., a graph convolution path based on Chebyshev polynomial to extract high-order features Feag1, a graph convolution path based on neighbor aggregation mechanism to extract heterogeneous neighborhood features Feag2, and a standard graph convolution path to extract basic structure features Feag3.

[0093] The features Feag1, Feag2 and Feag3 are subjected to multi-channel fusion in the feature splicing layer and nonlinear mapping through the Elu activation layer to obtain intermediate features Feagz. Subsequently, the features Feagz are sent to a deep graph convolution fusion module, which specifically includes two layers of Chebyshev graph convolution calculation processing to obtain features Feaq, two layers of aggregation-type graph convolution calculation processing to obtain features Feaj, and two layers of standard graph convolution calculation processing to obtain features Feab, for improving the expression depth and recognition accuracy of the structure defect recognition.

[0094] On this basis, by introducing an internal-external defect interaction unit, the features Feawr, Feaq, Feaj and Feab calculated in the building external structure defect recognition path are fused to realize the correlation modeling between the internal state and the external performance of the structure and enhance the perception ability of the model to the internal structure defects of the building. In the internal-external defect interaction unit: two interconnected channel attention layers are included, in the first three-channel attention layer, the features Feaq, Feaj and Feab are adaptively fused to obtain Fean1, and then the double-channel attention layer is used to internally fuse the features Fean1 and Feawr, and the interactive output feature Feazsr is obtained.

[0095] After the interactive output feature Feazsr is processed by the ReLU activation function and the Dropout layer, the final building internal defect recognition result Deciout is given by the output layer.

[0096] III. Construction of position / type-structure abnormal sensitivity table based on building defect risk degree mapping model

[0097] To fit the direct mapping between building defect types and defect locations and structural anomalies through modeling, this invention constructs a building defect risk prediction dataset using digital twins, ultimately obtaining type-structure mapping distributions and location-structure mapping distributions for final structural anomaly assessment and early warning. The specific steps include:

[0098] 1. Construction of a Dataset for Estimating Building Defect Risk: First, the following structural anomaly types, Su, are defined: local buckling, decreased component bearing capacity, material performance degradation, structural displacement, uneven inter-story stiffness, and localized high stress accumulation. Then, based on the constructed digital twin of the building, the building's operational conditions under these structural anomalies are simulated, and the following data are recorded respectively:

[0099] (1) Record whether various external defects occur and obtain external defect type data Esd; similarly, obtain internal defect type data Isd, and concatenate Esd and Isd to obtain complete defect type status data Csd.

[0100] (2) Record the location of the external defect Eld and the location of the internal defect Ild, and concatenate Eld and Ild to obtain the complete defect location data Cld;

[0101] By taking defect type status data Csd as input and structural anomaly type Su as output, a set of type-structure datasets is obtained. Based on this principle, a large amount of data is collected, and finally a complete type-structure dataset is obtained.

[0102] Using defect location data Cld as input and structural anomaly type Su as output, a set of location-structure datasets is obtained. Based on this principle, a large amount of data is collected, and finally a complete location-structure dataset is obtained.

[0103] Type-structure datasets and location-structure datasets are used to obtain the direct mapping relationship between type / location and structural anomalies, thereby assessing the overall structural anomalies of the building based on the mapping relationship.

[0104] 2. Design of Building Defect Risk Mapping Model: This model uses the LightGBM classifier to map type / location to structural anomalies. The model can output the classification results of each structural anomaly in parallel.

[0105] The LightGBM classifier was trained using the type-structure dataset and the location-structure dataset respectively, and the trained model parameters Param were obtained for the type-structure dataset. cls And the model parameters Param for location-structure loc ;

[0106] 3. Type / position sensitivity acquisition based on SHAP value: for realizing quantitative analysis of the degree of influence of defect type and position on structural anomaly identification, SHAP (SHapley Additive exPlanations) algorithm is introduced to explain the type-structure model parameter Param cls and the position-structure model parameter Param loc , so as to measure the marginal contribution of each defect type and position to the structural anomaly; the specific method includes:

[0107] (1) based on the parameter Param cls , the SHAP value of each defect type in the structural anomaly type discrimination process is calculated by using the SHAP algorithm, and the sensitivity value is calculated by the minimum-maximum normalization formula, the value range is [-1, 1], and the greater the value, the greater the positive contribution, on the contrary, the greater the negative contribution; the sensitivity values of all types to each structural anomaly constitute a type-structural anomaly sensitivity table Sen cls ;

[0108] (2) similarly, based on the parameter Param loc , the SHAP value of each defect position in the structural anomaly type discrimination process is calculated by using the SHAP algorithm, and the sensitivity value is calculated by the minimum-maximum normalization formula; the sensitivity values of all positions to each structural anomaly constitute a position-structural anomaly sensitivity table Sen loc ;

[0109] Therefore, Sen cls and Sen loc quantify the risk association strength between each defect type and defect position and the structural anomaly.

[0110] Four, building structure anomaly early warning

[0111] In order to realize the adaptive response and risk warning of the overall structure state of the building, the building defect risk degree estimation module is further designed. The module combines the internal and external defect results Decwout and Deciout obtained by identification, and the risk sensitivity tables Sen cls and Sen loc constructed, and the sensitivity table is constructed to realize the probability estimation and risk response of the building structure anomaly. Specifically, the following steps are included:

[0112] 1. Joint sensitivity table construction: Sen cls and Sen locRespectively used to represent the influence degree of different defect types / positions on structure abnormality, for further unified expression of the two types of sensitivity, the application constructs a joint sensitivity table to realize the joint contribution evaluation of defect types and positions on structure abnormality.

[0113] Specifically, the application proposes a method of unsupervised optimization to construct the joint sensitivity table, first, the objective function includes:

[0114] (1) Structure abnormality sensitivity minimum fluctuation target: the target aims to ensure that the output fluctuation of the generated joint sensitivity value in different samples or data batches is as small as possible, and the calculation method is: by weighting, the type sensitivity and the position sensitivity are fused, for each type of structure abnormality, the average fluctuation value of the joint sensitivity output of the abnormality under different data inputs is calculated, and the objective function term is the average fluctuation degree of all structure abnormality sensitivity outputs;

[0115] (2) Structure abnormality sensitivity maximum discrimination target: aims to improve the distinguishability of the risk contribution degree of each type of structure abnormality, the calculation method is: in the constructed joint sensitivity table, for each defect type-position combination, observe its sensitivity value distribution to each type of structure abnormality; calculate the discrimination degree of the distribution, and the discrimination target is the weighted average value of the difference, the target is to make it as large as possible in optimization to ensure that the model can distinguish different structure abnormality risk characteristics.

[0116] Based on the above two optimization targets, the particle swarm optimization algorithm is selected for unsupervised solution, and finally the joint sensitivity table Sen jit solving result.

[0117] 2, structure abnormality probability estimation:

[0118] After obtaining the joint sensitivity table Sen jit , combined with the actual defect recognition result Decwout and Deciout, the structure abnormality probability is calculated, specifically including:

[0119] (1) The internal and external defect recognition results Decwout and Deciout obtained by recognition are integrated to obtain the complete defect recognition result Det cmb ;

[0120] (2) According to the actual occurrence of the defect type and position, the corresponding sensitivity value is extracted from the joint sensitivity table Sen jit , and a weighted combination is constructed to build the risk response score Sc1, Sc2,..., Sc6 (respectively corresponding to local buckling, component load capacity decline, material performance degradation, structure offset, interlayer stiffness unevenness and local high stress aggregation) of each type of structure abnormality;

[0121] (3) The above scores are converted into structural anomaly occurrence probabilities Ph1, Ph2,..., Ph6 (corresponding to the occurrence probabilities of six structural anomalies) through a soft classification mechanism, which are used for subsequent early warning triggering.

[0122] 3. Building structure overall early warning factor calculation: According to the calculated six types of structural anomaly probabilities Ph1, Ph2,..., Ph6, an adaptive probability weight W is assigned to each structural anomaly probability w , w ∈ [1, 6], and then a weighted sum is obtained to obtain the basic severity Sev base . When multiple high-risk anomalies exist simultaneously, their combined effect may be much greater than that of individual action. Therefore, to further enhance the robustness of subsequent early warning and to enhance the accuracy of overall early warning factor calculation, a synergistic amplification factor Emk is further designed. First, count the number of anomalies Num fau in the probability set Ph1, Ph2,..., Ph6 whose values are ≥ 0.5 fau , and the synergistic amplification factor Emk = 1 + 0.2 × (Num base - 1). Then multiply the basic severity Sev fin and the synergistic amplification factor to obtain the final building structure overall early warning factor Sev safe .

[0123] 4. Building classification real-time early warning implementation:

[0124] A three-level building real-time early warning mechanism is constructed, and safety threshold Thr safe , alarm threshold Thr ris and emergency alarm threshold Thr best are set. According to the building structure overall early warning factor Sev fin calculated in real time and the above thresholds, the following judgments are made. When Sev fin is less than Thr safe , the building is in a safe state and no early warning signal is issued; when Sev fin is greater than or equal to Thr safe and less than Thr ris , a slight early warning signal is issued; when Sev fin is greater than or equal to Thr ris and less than Thr best , a moderate early warning signal is issued; and when Sev fin is greater than or equal to Thr best , an emergency early warning signal is issued.

[0125] Five, experimental result analysis

[0126] To verify the structural defect detection accuracy and early warning accuracy of the building structure anomaly detection and early warning method proposed in the application, comparative experiments are carried out to evaluate the performance, and the performance is compared and evaluated for two main technical links: (1) internal defect identification accuracy of the building anomaly structure identification module (2) joint structure anomaly sensitivity table modeling test.

[0127] 1. Defect identification accuracy analysis of the building anomaly structure identification module

[0128] For the building anomaly structure representation data set constructed, the prediction accuracy of the building defect identification model proposed in the application based on graph data analysis and fusion of internal and external unit interaction and two types of typical comparative models on five different internal defects (internal honeycomb of concrete head, internal structural crack of component, internal stress concentration of steel structure, prestress relaxation, and foundation pile body displacement) is evaluated. The comparative models include:

[0129] (1) VGG16 model: taking multi-modal sensing data (such as stress wave and electromagnetic signal) as input, outputting defect category probability through full connection layer, without introducing component topological correlation feature;

[0130] (2) MobileNetV2 model: adopting deep separable convolution to realize lightweight design, which can classify internal defects, but has low degree of classification of fine defects such as steel bar corrosion and prestress relaxation;

[0131] The evaluation indexes are classification accuracy (Accuracy) and precision-recall area under the curve (PR-AUC), wherein the accuracy reflects the overall ability of the model to correctly classify the defect category, and the higher the accuracy, the better the classification effect; the PR-AUC score comprehensively measures the identification accuracy and comprehensiveness of the model for each defect, and the higher the PR-AUC score, the stronger the balance of the category determination.

[0132] From the above result comparison image, Figure 3 , Figure 4 It can be seen that the building defect identification model proposed in the application is better than the comparative model in terms of Accuracy and PR-AUC for the identification of five different internal defects. The error in internal defect identification is obviously reduced, and the stability and robustness of the multi-channel graph convolution operation and the internal and external defect combination unit in building internal defect identification are embodied.

[0133] 2. Heat map display of joint structure anomaly sensitivity table modeling

[0134] To further illustrate the quantitative ability of the proposed optimization algorithm in the aspect of structural joint risk degree, the application selects the normalized contribution degrees of five different defect types and five different building positions from the final modeling results, and drawsFigure 5 The heat map is shown.

[0135] The normalized joint risk degree heat map is a core visualization tool of the building intelligent early warning system, deeply integrates digital twin simulation data and model output type / location sensitivity features, and provides precise targeting for early warning by quantifying the coupling risk of "defect type / location". The intelligent and precision of building structure anomaly early warning are significantly improved, and intuitive and quantitative decision basis is provided for active prevention of structure risk.

[0136] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0137] Although the specific embodiments of the present application are described above, they are not intended to limit the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A building structure anomaly detection method based on internal and external defect synergy, characterized by, The method comprises the following steps: S1, acquiring an image of an external surface of a building in real time; and acquiring a plurality of internal sensing data of a plurality of key building nodes in real time based on a built digital twin of the building; S2, inputting the data acquired in S1 into a trained building internal and external defect identification model, the model comprising a dual-path building defect identification module, an external structural defect identification path inputting the external surface image, adopting a multi-scale external defect sensing path, and outputting a detection result of external defects of the building through multi-resolution feature fusion; and an internal structural defect identification path inputting the internal sensing data, adopting a graph data calculation and fusion unit for component topology correlation characteristics, mining hidden internal defects based on component connection relationships, and outputting an identification result of internal defects of the building; S3, calculating a plurality of structural anomaly probabilities based on a built joint sensitivity table and combining the two identification results in S2, obtaining an overall early warning factor through adaptive weights and a synergistic amplification factor, and obtaining a multi-level early warning based on a preset threshold; The specific process is as follows: The identified external defect recognition result and the internal defect recognition result are integrated to obtain a complete defect recognition result ; According to the actual defect type and location, the corresponding sensitivity value is extracted from the joint sensitivity table , and a weighted combination is made to construct the risk response score of each type of structural anomaly , respectively corresponding to local buckling, member load capacity reduction, material performance degradation, structural deviation, uneven interlayer stiffness and local high stress concentration; The above score is converted into a probability of structural abnormality occurrence by a soft classification mechanism ; Based on structural anomaly probability Adaptive probability weights are assigned to each structural anomaly probability. , The base severity is then obtained by weighted summation. ; Statistical probability Set probability Number of anomalies with median value > 0.5 , synergistic amplification factor = 1 + 0.2 x ( - 1); Subsequently, the base severity is multiplied by the synergy amplification factor to obtain the final overall building structure early warning factor for use ; and the synergy amplification factor is multiplied by the base severity to obtain the final overall building structure early warning factor for use ; Based on early warning factors And set the safety threshold to determine, get multi-level real-time warning; The joint sensitivity table is obtained by building a building defect risk degree speculation dataset in a digital twin manner, obtaining type-structure mapping distribution and location-structure mapping distribution, combining SHAP values to obtain a type-structure anomaly sensitivity table and a location-structure anomaly sensitivity table, and jointly building a joint sensitivity table; specifically: Firstly, define several structural anomaly types , based on the constructed building digital twin, simulate the building operating conditions when several structural anomaly types occur, and record the following data respectively: Record whether each type of external defect occurs and obtain external defect type data ; similarly, obtain internal defect type data , and splice with to obtain complete defect type status data ; Record the position of the occurrence of external defects and the position of the occurrence of internal defects , and with splicing to obtain complete defect position data ; Based on the above manner, the dataset is constructed, the building defect risk degree mapping model is trained, and model parameters for type-structure after training are obtained and model parameters for position-structure ; Based on parameters The SHAP value of each defect type in the structural anomaly type discrimination process is calculated by using the SHAP algorithm, and the original SHAP sensitivity value is calculated to obtain the sensitivity value through the minimum-maximum normalization formula. The values of sensitivity of all types to each structural anomaly constitute a type-structural anomaly sensitivity table ; Similarly, based on parameters , the SHAP values of each defect position in the structural anomaly type discrimination process are calculated by using the SHAP algorithm, and the original SHAP sensitivity values are calculated to obtain the sensitivity values by using the minimum-maximum normalization formula. The position-structure anomaly sensitivity table is formed by the sensitivity values of all positions to each structure anomaly .

2. The building structure anomaly detection method based on the synergy of internal and external defects according to claim 1, characterized in that: The collection method of external structural data for training the building internal and external defect identification model is as follows: For the selected target building, the parameter information related to the building external contour and facade structure in the BIM model is called, including building height, external facade material, and component distribution data information; Then, a drone is preset in the built digital twin, and image external surface pictures are collected; During data calibration, a rectangular box is used to calibrate the defect area on the outer surface of the building. On this basis, structured label information is added to each calibrated defect, including defect type Eer, specifically wall surface cracks , external wall leakage , external component peeling , crenellation cracking , door and window frame deformation , and defect location Eew, with location coordinates (x, y, z).

3. The method of claim 1, wherein the method is characterized by: The collection method of internal structural data for training the building internal and external defect identification model is as follows: Deploy piezoelectric sensors to collect stress data at M key building nodes selected from the ends of concrete columns of the target building, the ends of concrete beams, the connecting nodes of steel structures, beam-column joints, main and secondary beam intersections, floor supports and mid-span locations, and the roots of building cantilever members Deploy vibration pickups to collect vibration amplitude data and vibration frequency data Deploy electrochemical sensors to collect corrosion current density data of steel bars Deploy ultrasonic sensors to collect reflected wave signal characteristics of building materials at the locations Deploy laser displacement meters to collect static displacement at the locations ; and combine to form building comprehensive perception data at a certain node of the building ; Defect type of defect in internal structure of building during data calibration Specifically including internal honeycomb of concrete head Structural crack in internal structure of component Internal stress concentration of steel structure Prestress relaxation , displacement of foundation pile body , corrosion of steel bar , and loose grouting , and defect occurrence position , and defect occurrence position coordinates (x, y, z).

4. The building structure anomaly detection method based on the synergy of internal and external defects according to claim 1, characterized in that: The external structural defect identification path specifically comprises: Input image data First, the image of the building exterior surface is preliminarily feature-extracted by an external defect detail perception unit to obtain features Then, the Inception feature extraction module is passed through, different size features are captured in parallel through multi-scale convolution kernels, and output feature maps are calculated. Subsequently, the features The image attention segmentation unit is fed in: first, the features The local gradient amplitude is calculated by the Sobel operator to generate an initial spatial attention mask, and the area with a higher gradient has a stronger corresponding relationship with the defect edge; subsequently, the bidirectional gating mechanism is used to guide the features to tilt to the corresponding feature area of the high gradient area through the spatial attention mask, and finally the enhanced defect feature map is output ; Subsequently, the image is sent into the external defect detail parallel perception unit for external defect identification processing to obtain , respectively , the external defect detail parallel unit includes four parallelly arranged external defect detail perception units; and the adaptive integration of the features is completed by using the image channel attention unit to obtain the fused features ; Fused features The features are calculated again sequentially through 3x3 convolution layers , and then processed through a Dropout layer and two KNN classification layers in cascade, to finally output the detection result of the building exterior defects .

5. The method of claim 1, wherein the method is characterized by: The internal structural defect identification path specifically comprises: First, collect the internal defect data of the building's abnormal structure. The data is input into the graph data construction unit. In this unit, the two ends of each concrete column, the two ends of each concrete beam, the connection nodes of the steel structure, the beam-column nodes, the intersections of primary and secondary beams, the base of the floor slab, the mid-span of the floor slab, and the root of the cantilevered structural members are considered as nodes of the graph data. The node features are all the comprehensive building perception data collected by each set of sensors. , The physical connections between sensor deployment locations are used to construct edges, thereby forming a building representation map data with prior physical topology. ; Subsequently, building representation map data At the same time, input three groups of structural differentiated graph convolution channels, each group of channels contains three layers of graph convolution operations, respectively, the graph convolution path based on Chebyshev polynomial is extracted to obtain high-order features , the graph convolution path based on neighbor aggregation mechanism is extracted to obtain heterogeneous neighborhood features , and the standard graph convolution path is extracted to obtain basic structure features ; Features Multi-channel fusion is performed in the feature stitching layer, and non-linear mapping is performed through the Elu activation layer to obtain intermediate features ; then the features are sent to a deep graph convolution fusion module to perform two-layer Chebyshev graph convolution calculation and processing to obtain features , perform two-layer aggregated graph convolution calculation and processing to obtain features , and perform two-layer standard graph convolution calculation and processing to obtain features ; On this basis, by introducing the internal and external defect interaction unit, the features calculated in the building external structure defect recognition channel are fused , and the features of the internal and external defect interaction unit are calculated to realize the correlation modeling between the internal state and the external performance of the structure. In the internal and external defect interaction unit, two interconnection channel attention layers are included, in the first three-channel attention layer, the features are adaptively fused to obtain , then the internal and external fusion is performed on the features and the features by using a double-channel attention layer, and the interaction output features are obtained. Interaction output features After the ReLU activation function and the Dropout layer processing, the final output layer gives the identification result of the building interior defects .

6. The building structure anomaly detection method based on the synergy of internal and external defects according to claim 1, characterized in that: The joint sensitivity table is built by an unsupervised optimization method: The optimization objectives include: The structural anomaly sensitivity minimum fluctuation objective: the type sensitivity and the location sensitivity are fused by weighting, for each type of structural anomaly, the average fluctuation value of the joint sensitivity output of the anomaly under different data inputs is calculated, and the objective term is the average fluctuation degree of all structural anomaly sensitivity outputs; The structural anomaly sensitivity maximum discrimination objective: in the constructed joint sensitivity table, for each defect type-location combination, the sensitivity value distribution of each type of structural anomaly is observed; the discrimination degree of the distribution is calculated, and the discrimination degree objective is the weighted average value of the difference of the distribution, and the objective is to make it as large as possible in the optimization to ensure that the model can distinguish different structural anomaly risk characteristics; Based on the above two optimization objectives, the particle swarm optimization algorithm is selected for unsupervised solution, and finally the joint sensitivity table capable of fully representing the two risk sensitivity distributions is obtained .

7. The building structural anomaly detection method based on internal and external defect synergy according to claim 1, characterized in that: A three-level building real-time early warning mechanism is constructed, and safety threshold, alarm threshold and emergency alarm threshold are set ;​​ According to the real-time calculation of the building structure overall early warning factor and the threshold value, when is less than , the building is in a safe state, and no early warning signal is sent; when is greater than or equal to and less than , a slight early warning signal is sent; when is greater than or equal to and less than , a medium early warning signal is sent; and when is greater than or equal to , an emergency early warning signal is sent.

Images