A building safety performance evaluation method based on big data analysis

Abstract

The application discloses a kind of based on big data analysis's building safety performance evaluation method, it is related to building safety detection technical field, including the following steps: using five-in-one probe to building to be evaluated area is grid scanning, generates multimodal detection data frame;Terahertz echo is handled and combined millimeter wave ranging result generates facing perspective positioning chart;Based on facing perspective positioning chart extraction hyperspectral response characteristics, generates interface constraint spectral sample;Through interface constraint SpectralFormer model, generates base layer strength inversion chart;To suspected weak area carries out three-dimensional sound school test, generates sound school test chart;To multi-source result executes environment self-adapting dynamic weight fusion, generates component safety state table;Component safety state table is mapped to building model or drawing, generates safety evaluation heat map and outputs evaluation report.The application improves the comprehensiveness, accuracy and traceability of building safety evaluation.

Description

Technical Field

[0001] This invention relates to the field of building safety inspection technology, and in particular to a method for evaluating building safety performance based on big data analysis. Background Technology

[0002] With the increasing demand for safety assessments of existing buildings, urban renewal, preservation of historical buildings, and post-disaster investigations, non-destructive testing and safety performance assessment technologies for building walls, beams, slabs, columns, finishes, and base structures have received widespread attention. Existing technologies, such as millimeter-wave, infrared, ultrasonic, terahertz, and hyperspectral imaging, have been used for building surface condition identification, detection of hollow areas and debonding, detection of internal concrete defects, and material degradation analysis. For example, existing solutions utilize millimeter-wave or infrared methods to assist in identifying abnormal surface areas, terahertz imaging to identify interfaces or hollow features of non-metallic materials, and hyperspectral data to analyze concrete corrosion products or material degradation characteristics. There are also technologies that utilize ultrasonic arrays, full-matrix acquisition, and full-focus imaging to detect internal cracks, voids, or interface defects in concrete.

[0003] However, in existing building inspection services, there remains a technological gap where it is difficult to accurately infer the strength of the substrate through the finishing layer. For walls and columns with tiled, stone, plastered, or decorative finishes, traditional rebound, penetration, or core drilling methods usually require damaging the finish before accessing the substrate, which leads to significant resistance from homeowners and high repair costs. When using terahertz or millimeter-wave testing alone, although it can identify the thickness of the finishing layer, hollow areas, or debonding, it is difficult to directly determine the grade of the substrate mortar or the strength of the concrete. When using hyperspectral testing alone, the spectral response is easily affected by the finishing material, surface reflection, moisture, hollow areas, and interface debonding, making it impossible to reliably distinguish between changes in the surface finish and insufficient substrate strength.

[0004] Meanwhile, existing multi-source detection schemes mostly remain at the stage of displaying single results or summarizing with fixed weights, lacking a mechanism to embed terahertz interface parameters as constraints into the hyperspectral intensity inversion process, and also lacking a dynamic fusion mechanism to use three-dimensional acoustic verification results to reverse correct for strength risks and hollowing / debonding risks. Therefore, existing technologies cannot simultaneously obtain surface perspective positioning, base layer strength inversion, deep defect verification, and component safety scoring results without chiseling or drilling, resulting in problems such as high missed detection rates, large verification workload, insufficient traceability of results, and difficulty in quantifying maintenance priorities in the safety assessment of existing buildings.

[0005] Therefore, how to provide a building safety performance assessment method based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a building safety performance assessment method based on big data analysis. This invention fully utilizes terahertz detection, hyperspectral inversion, three-dimensional acoustic verification, and multimodal dynamic fusion technologies, and describes in detail the intelligent assessment process of building facade perspective positioning, base layer strength inversion, internal defect verification, and component safety scoring. It has the advantages of comprehensive detection dimensions, accurate defect identification, intuitive assessment results, and continuous model updates.

[0007] A building safety performance assessment method based on big data analysis according to an embodiment of the present invention includes the following steps: Step 1: Use a five-in-one probe to perform a gridded scan of the area of ​​the building to be evaluated. After spatial calibration of heterogeneous sensors, time synchronization, anomaly sampling rejection and confidence marking, multimodal detection data frames are generated. Step 2: Filter, extract the envelope, and identify the interface peaks in the terahertz echoes in the multimodal detection data frames. Combine the millimeter-wave ranging results to correct the interface depth and generate a surface perspective positioning map. Step 3: Determine the effective detection area based on the surface perspective positioning map, perform radiometric calibration, normalization and spectral feature extraction on the hyperspectral response, and embed the terahertz interface parameters into the spectral features to generate interface-constrained spectral samples; Step 4: Input the interface-constrained spectral samples into the interface-constrained SpectralFormer model, and generate the base layer intensity inversion map through spectral embedding, band coding, spectral feature extraction and interface constraint modulation; Step 5: Based on the base strength inversion map and the surface perspective positioning map, identify suspected weak areas, and use a three-dimensional acoustic phased array to perform full matrix acquisition and full-focus reconstruction to generate an acoustic verification map; Step 6: Perform environmental adaptive dynamic weight fusion based on the surface perspective positioning map, the base strength inversion map, and the acoustic verification map to calculate the component safety score and generate the component safety status table; Step 7: Map the component safety status table to the building model or drawings, generate a safety assessment heatmap, output a building safety performance assessment report, and add the current test results and review results to the historical test database, and update the model parameters based on the historical test samples.

[0008] Optionally, step one specifically includes: Establish a building inspection coordinate system based on the CAD drawings, BIM model or on-site measured outline of the building to be evaluated, divide the area to be evaluated into wall, beam, slab, column, finishing unit or masonry unit, and write the component number, floor number, inspection surface number and scanning path. A five-in-one probe is used to perform gridded scanning of the area to be evaluated. The five-in-one probe includes a terahertz radar, a hyperspectral camera, a three-dimensional acoustic phased array, a millimeter-wave radar, and a visible-infrared composite imaging unit. Spatial calibration and time synchronization of the five-in-one probe were performed to obtain the external parameter matrices of each sensor and collect the timestamps uniformly. Set the gridded scanning parameters according to the surface size, structural importance, and on-site risk area distribution of the component being inspected, and record the probe's spatial coordinates, attitude angle, and inspection point number; Anomaly identification and credibility marking were performed on the collected data to identify terahertz echo saturated data, hyperspectral anomalous bands, insufficient acoustic coupling data, visible light overexposure and underexposure data, and infrared temperature saturation data. The missed detection points are filled by interpolation, and the detection point number, component number, spatial coordinates, acquisition timestamp, acquisition results of each mode, anomaly marker, interpolation marker and confidence level of each mode are encapsulated to generate a multimodal detection data frame.

[0009] Optionally, step two specifically includes: The terahertz time-domain spectral signal in the multimodal detection data frame is subjected to DC component removal, baseline correction, bandpass filtering and amplitude normalization to obtain the terahertz echo signal; Envelope extraction is performed on the terahertz echo signal, and the reflection peaks in the echo envelope are obtained by peak detection. An interface peak sequence is formed according to the arrival time of the reflection peaks. The interface positions corresponding to different reflection peaks are calculated using the time-of-flight method. Combined with the propagation velocity parameters of the finishing material, adhesive material, and base material, the finishing surface, finishing back, adhesive layer interface, and base layer interface are determined, and the finishing layer thickness and base layer interface depth are obtained. When there are overlapping reflection peaks, missing peaks at the base interface, or insufficient signal-to-noise ratio in the terahertz echo, the interface depth is corrected by combining the millimeter-wave ranging results, probe attitude angle, and distance to the detection surface. Based on the abnormal reflection peaks, echo amplitude ratios, and echo energy attenuation between the back of the finish and the interface with the base layer, suspected hollow points are identified; based on the changes in the depth of the base layer interface between adjacent detection points, suspected debonding areas are identified. Spatial mapping was performed on the thickness of the finishing layer and the depth of the base layer interface. Suspected hollow points and suspected delamination areas were merged to generate a perspective positioning map of the finishing layer.

[0010] Optionally, step three specifically includes: Based on the thickness of the finishing layer, the location of the base layer interface, the suspected hollow area, the suspected debonding area, and the terahertz confidence level in the finishing layer perspective positioning diagram, the effective detection area and the interference detection area are determined, and interference markers are written into the interference detection area. Dark current correction and white board correction are performed on the hyperspectral data corresponding to the effective detection area and the interference detection area, and the sensor output value is converted into reflectance data; The reflectance data were smoothed and normalized, and the spectral features of the normalized hyperspectral curves were extracted to obtain the hyperspectral response features. Absorption feature enhancement is performed on the material's sensitive wavelength bands to extract spectral response features related to mortar aging, concrete moisture content, surface carbonization, and material degradation. The terahertz interface parameters at the same detection point are numerically normalized and encoded, and then combined with the hyperspectral response features to form a spectral feature vector with interface constraint information. The spectral feature vector, terahertz interface parameters, base material type, historical strength label, sample confidence label, and interference marker are encapsulated to generate interface constraint spectral samples, so that the terahertz interface parameters can be used as interface constraint information in subsequent base strength inversion.

[0011] Optionally, step four specifically includes: The hyperspectral reflectance sequence in the interface-constrained spectral sample is divided into several continuous band groups according to the band order. During the model training phase, the SpectralFormer model is constrained by inputting continuous band groups, terahertz interface parameters, base material type, historical intensity labels, sample confidence labels, and interference markers into the interface. During the model inference phase, the continuous band group, terahertz interface parameters, and base material type of the current detection point are input into the trained interface constraint SpectralFormer model. The interface-constrained SpectralFormer model includes a band group embedding layer, an interface-aware location encoding layer, an interface-constrained multi-head self-attention layer, a cross-layer adaptive fusion layer, an interface gating layer, a sample weight calculation module, and a multi-task intensity output layer. A spectral group token is generated through a band group embedding layer, an interface-aware spectral token is generated through an interface-aware position coding layer, spectral features are extracted through an interface-constrained multi-head self-attention layer, cross-layer fused spectral features are generated through a cross-layer adaptive fusion layer, and interface-constrained modulation is performed on the cross-layer fused spectral features through an interface gating layer to generate interface-constrained spectral features. Based on the interface-constrained spectral characteristics, the base layer strength grade, continuous strength value, strength confidence level, and suspected probability of insufficient strength are output, and bound to the detection point coordinates, component number, and base layer material type to generate a base layer strength inversion map.

[0012] Optionally, the interface constraint modulation specifically includes: The terahertz interface parameters are encoded to obtain an interface parameter vector. The terahertz interface parameters include the finishing layer thickness, the base layer interface depth, the echo energy attenuation, the hollow mark, the debonding mark, and the terahertz confidence level. The band position code and the interface parameter vector are concatenated through the interface-aware position coding layer and mapped to the interface-aware position code. The interface-aware position code is then superimposed on the spectral group token to generate the interface-aware spectral token. The interface-constrained multi-head self-attention layer maps the query vector, key vector, and value vector of the interface-aware spectral token, and maps the interface parameter vector to the attention bias matrix. The attention bias matrix is ​​added to the dot product of the query vector and the key vector to obtain the interface-constrained attention weight. By fusing shallow absorption peak features, mid-layer spectral band combination features, deep material degradation features, and interface parameter vectors through a cross-layer adaptive fusion layer, the spectral features after cross-layer fusion are obtained. The interface parameter vector is input into the gated network through the interface gated layer to generate the interface gated vector. The interface gated vector is then used to multiply the spectral features after cross-layer fusion element by element to generate the interface constrained spectral features. The sample weight calculation module generates training sample weights based on the historical intensity label source, terahertz confidence, hyperspectral anomaly label, hollow label, and debonding label. The multi-task strength output layer is configured with mortar grade classification branches, concrete strength grade classification branches, continuous strength regression branches, strength confidence branches, and strength deficiency probability branches. During the model training phase, the total training loss is generated based on the mortar grade classification loss, concrete strength grade classification loss, continuous strength regression loss, suspected insufficient strength probability loss, strength confidence loss, and gating constraint loss, and the model parameters are updated in combination with the training sample weights.

[0013] Optionally, step five specifically includes: Based on the strength confidence and suspected strength insufficiency probability in the base strength inversion map, and the suspected hollow area, suspected debonding area and terahertz confidence in the surface perspective positioning map, the candidate range for three-dimensional acoustic testing is determined. When the intensity confidence level is lower than the intensity confidence level threshold, the corresponding detection point is marked as a low confidence inversion area; when the suspected intensity deficiency probability is not lower than the suspected intensity deficiency probability threshold, the corresponding detection point is marked as a suspected weak area. Both the low confidence inversion area and the suspected weak area are included in the candidate range of 3D acoustic verification. Acoustic verification points are set up for the candidate range of three-dimensional acoustic verification, and the coupling status of each acoustic verification point is checked. A three-dimensional acoustic phased array is used for full matrix acquisition. Each transmitting element in the array is excited sequentially, and all receiving elements synchronously receive the echo signal to obtain full matrix acoustic data. The full matrix acoustic data is filtered, time-corrected, time-gain compensated, and amplitude-normalized, and a three-dimensional reconstruction mesh is established within the verification range. The three-dimensional reconstructed mesh was imaged using the full-focusing method to obtain the acoustic reflection intensity of each voxel. Based on the three-dimensional acoustic reflection intensity data, threshold segmentation, connected component analysis and defect type labeling are performed to identify suspected areas of deep cracks, holes, loose base layers and abnormal reinforcement. The three-dimensional acoustic imaging results are spatially overlaid with the base strength inversion map and the surface perspective positioning map to generate acoustic verification confidence and verification matching marks, and form an acoustic verification map.

[0014] Optionally, step six specifically includes: The terahertz results, hyperspectral intensity inversion results, three-dimensional acoustic calibration results, millimeter-wave ranging results, visible light surface condition and infrared temperature condition are aligned according to the detection point number. For each detection point, configure terahertz initial weight, hyperspectral initial weight, three-dimensional acoustic initial weight, millimeter wave initial weight, visible light initial weight and infrared initial weight, and the sum of the six types of initial weights is 1; The confidence level of each mode is determined based on terahertz confidence level, hyperspectral confidence level, three-dimensional acoustic confidence level, millimeter wave confidence level, visible light confidence level and infrared confidence level, wherein the three-dimensional acoustic confidence level is taken from the acoustic verification confidence level in the acoustic verification diagram. An environmental correction coefficient is generated based on the environmental state, and an environmental adaptive dynamic weight fusion is performed based on the initial weight, confidence level, and environmental correction coefficient of the same modality to obtain the modality fusion weight. The risk values ​​of the test points are calculated based on the surface perspective positioning diagram, the base strength inversion diagram and the acoustic verification diagram. The risk values ​​of the test points include strength risk, hollow debonding risk, deep crack risk, abnormal steel reinforcement risk and loose and porous base layer risk. Calculate the strength consistency coefficient and the hollow debonding consistency coefficient based on the verification matching mark; Based on the modal fusion weight, the risk value of the detection point, the strength consistency coefficient, and the hollow debonding consistency coefficient, the comprehensive risk of the detection point strength, the comprehensive risk of the detection point hollow debonding, the comprehensive risk of the detection point deep cracks, the comprehensive risk of the detection point abnormal reinforcement, and the low confidence detection risk of the detection point are calculated. Among them, the low confidence detection risk of the detection point is calculated based on the confidence of each modality and the interpolation mark. The risks of the detection points are aggregated according to the component number to obtain the comprehensive risk of component strength, the comprehensive risk of component hollowing and debonding, the comprehensive risk of component deep cracks, the comprehensive risk of component abnormal reinforcement, and the risk of component low-reliability detection. Configure component risk weights according to the type of the component being detected, calculate component safety scores based on component risk weights and the risks of each component, generate safety levels based on component safety scores, and encapsulate component safety scores and safety levels into a component safety status table.

[0015] Optionally, the environment-adaptive dynamic weight fusion specifically includes: The initial weights for each modality are determined by normalizing the single-modal recognition accuracy of each modality for the corresponding risk item in the historical review samples; When the area ratio of overexposed, underexposed, or specularly reflected regions in a visible light image reaches the preset interference judgment condition, the visible light environment correction coefficient is a value greater than 0 and less than 1, and the hyperspectral environment correction coefficient is a value greater than 0 and less than 1. When the infrared thermal image is subject to temperature saturation, surface dampness, or rain interference, the infrared environment correction factor is taken as a value greater than 0 and less than 1, and the terahertz environment correction factor is taken as a value greater than 0 and less than 1. When millimeter-wave ranging results show abnormal reflections from metal mesh, water pipes, or embedded parts, the terahertz environment correction factor takes a value greater than 0 and less than 1, while the millimeter-wave environment correction factor and the three-dimensional acoustic environment correction factor take a value greater than 1. Multiply the initial weights, confidence levels, and environmental correction coefficients of the same modality to obtain the unnormalized weights of that modality. Then, divide the unnormalized weights by the sum of the unnormalized weights of the six modalities at the same detection point to obtain the fusion weights of that modality. Strength risk is obtained from the probability of suspected strength deficiency; hollow and debonding risk is obtained from the normalized risk value of suspected hollow and debonding areas; deep crack risk, steel reinforcement abnormality risk and base layer looseness and void risk are obtained by normalizing the defect type marking, acoustic reflection intensity, abnormal body volume and abnormal depth in the acoustic verification diagram. When the strength risk at the same test point is not lower than the preset risk threshold, the risk of loose soil and pores in the base layer is not lower than the preset risk threshold, and the verification matching mark shows that there is a spatial intersection between the two, the strength consistency coefficient takes a value greater than 1; when the strength risk is not lower than the preset risk threshold but the risk of loose soil and pores in the base layer is lower than the preset risk threshold, the strength consistency coefficient takes a value greater than 0 and less than 1. When the risk of delamination and detachment at the same detection point is not lower than the preset risk threshold, the risk of loose pores in the base layer is not lower than the preset risk threshold, and the verification matching mark shows that there is a spatial intersection between the two, the consistency coefficient of delamination and detachment is greater than 1; when the risk of delamination and detachment is not lower than the preset risk threshold but the risk of loose pores in the base layer is lower than the preset risk threshold, the consistency coefficient of delamination and detachment is greater than 0 and less than 1. The comprehensive risk of the intensity of the detection point is determined by the hyperspectral fusion weight, intensity risk, three-dimensional acoustic fusion weight, risk of loose and porous base layer, and intensity consistency coefficient. The comprehensive risk of delamination and debonding at the detection point is jointly determined by the terahertz fusion weight, the risk of delamination and debonding, the three-dimensional acoustic fusion weight, the risk of loose and porous base layer, and the consistency coefficient of delamination and debonding.

[0016] Optionally, step seven specifically includes: Map the component safety status table, inspection point coordinates, component number, surface perspective positioning diagram, base strength inversion diagram, and acoustic verification diagram to the building BIM model, CAD drawings, floor plan, or building elevation. Establish the transformation relationship between the detection coordinate system and the building model coordinate system based on the on-site control points, component boundary points and detection start points, and map the detection point coordinates to the corresponding component surface; A safety assessment heatmap is generated by overlaying strength display layers, hollow debonding display layers, internal defect display layers, and low-confidence detection layers according to component numbers. Maintenance priorities are generated based on component safety scores and safety levels; Generate a building safety performance assessment report according to the structured report template. The report content includes the inspection object, inspection scope, number of inspection points, component safety status table, safety assessment heat map, base layer strength distribution, hollow and debonded distribution, internal defect distribution, safety level, maintenance priority, and a list of review points. The results of manual review, core drilling verification, partial chipping verification, repair and re-inspection are added to the historical inspection database according to the inspection batch, inspection point number and component number to form a traceable historical inspection sample sequence. When the number of newly added verification samples reaches the preset update threshold, the training sample set, sample confidence labels, dynamic weight parameters, risk judgment threshold, and interface constraint SpectralFormer model parameters of the interface constraint model are updated based on the historical detection sample sequence.

[0017] The beneficial effects of this invention are: (1) This invention uses a five-in-one probe to perform gridded scanning of the area to be evaluated, and performs spatial calibration, time synchronization, anomaly removal and credibility marking on terahertz, hyperspectral, three-dimensional acoustic, millimeter wave and visible-infrared data. This solves the problems of inconsistent data coordinates, asynchronous sampling and difficulty in distinguishing anomalies in existing building inspections. It enables the detection results of different modes to be aligned at the same detection point and under the same component number, providing a consistent data basis for subsequent defect identification, strength inversion and component scoring, and improving the traceability of on-site inspection results and the convenience of engineering review. (2) This invention embeds the thickness of the decorative layer, the depth of the base layer interface, the hollow mark, the debonding mark and the terahertz confidence generated by the terahertz echo into the hyperspectral features, and participates in the base layer strength inversion through the interface-constrained SpectralFormer model, the interface-aware position encoding, the attention bias matrix and the interface gating vector, to solve the problem that existing hyperspectral detection is easily affected by the decorative layer occlusion, interface debonding and invisible factors of the base layer, so that the strength inversion no longer depends only on the surface spectral response, and improves the reliability of the determination of the base layer strength level, continuous strength value and the probability of suspected insufficient strength. (3) The present invention determines the suspected weak areas based on the base strength inversion map and the surface perspective positioning map, and generates an acoustic verification map by using a three-dimensional acoustic phased array full matrix acquisition and full-focus reconstruction. Then, the acoustic verification credibility and verification matching mark are used in reverse to participate in the environmental adaptive dynamic weight fusion, which solves the problem of fixed weighting of existing multi-source detection results and difficulty in correcting single-modal misjudgment. This enables the strength risk, hollow debonding risk and deep defect risk to be dynamically adjusted according to environmental interference and verification consistency, and finally forms a component safety status table and safety assessment heat map, providing a quantitative basis for determining maintenance priorities and building safety management. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a building safety performance assessment method based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of a building safety performance assessment method based on big data analysis proposed in this invention; Figure 3 This is a framework diagram of the SpectralFormer model for interface constraints in a building safety performance evaluation method based on big data analysis proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A method for assessing building safety performance based on big data analysis includes the following steps: Step 1: Use a five-in-one probe to perform a gridded scan of the area of ​​the building to be evaluated. After spatial calibration of heterogeneous sensors, time synchronization, anomaly sampling rejection and confidence marking, multimodal detection data frames are generated. Step 2: Filter, extract the envelope, and identify the interface peaks in the terahertz echoes in the multimodal detection data frames. Combine the millimeter-wave ranging results to correct the interface depth and generate a surface perspective positioning map. Step 3: Determine the effective detection area based on the surface perspective positioning map, perform radiometric calibration, normalization and spectral feature extraction on the hyperspectral response, and embed the terahertz interface parameters into the spectral features to generate interface-constrained spectral samples; Step 4: Input the interface-constrained spectral samples into the interface-constrained SpectralFormer model, and generate the base layer intensity inversion map through spectral embedding, band coding, spectral feature extraction and interface constraint modulation; Step 5: Based on the base strength inversion map and the surface perspective positioning map, identify suspected weak areas, and use a three-dimensional acoustic phased array to perform full matrix acquisition and full-focus reconstruction to generate an acoustic verification map; Step 6: Perform environmental adaptive dynamic weight fusion based on the surface perspective positioning map, the base strength inversion map, and the acoustic verification map to calculate the component safety score and generate the component safety status table; Step 7: Map the component safety status table to the building model or drawings, generate a safety assessment heatmap, output a building safety performance assessment report, and add the current test results and review results to the historical test database, and update the model parameters based on the historical test samples.

[0021] In this embodiment, step one specifically includes: Establish a building inspection coordinate system based on the CAD drawings, BIM model or on-site measured outline of the building to be evaluated, divide the area to be evaluated into wall, beam, slab, column, finishing unit or masonry unit, and write the component number, floor number, inspection surface number and scanning path. A five-in-one probe was used to perform gridded scanning of the area to be evaluated. The five-in-one probe includes a terahertz radar, a hyperspectral camera, a three-dimensional acoustic phased array, a millimeter-wave radar, and a visible-infrared composite imaging unit. The terahertz radar uses a terahertz time-domain spectral acquisition method, the hyperspectral camera uses a line-scan hyperspectral imaging method, the three-dimensional acoustic phased array uses a multi-element pulse echo acquisition method, the millimeter-wave radar uses a frequency-modulated continuous wave ranging method, and the visible-infrared composite imaging unit is used to acquire visible light images and infrared thermal images. Spatial calibration and time synchronization of the five-in-one probe were performed. The external parameter matrix of each sensor was obtained by using a checkerboard calibration plate, a stepped thickness reflective target, a metal ball reflective target and an acoustic reflective target. The timestamp was uniformly collected by hardware trigger pulse, IEEE1588 precise time protocol and second pulse signal. Gridded scanning parameters are set according to the surface size, structural importance, and on-site risk area distribution of the components to be inspected, and the probe spatial coordinates, attitude angles, and inspection point numbers are recorded by a laser rangefinder, wheel encoder, and inertial measurement unit. Anomaly identification and credibility marking were performed on the collected data. Analog-to-digital conversion saturation detection was used to identify terahertz echo saturated data. Dark current correction and whiteboard correction were used to identify hyperspectral anomalous bands. A-scan signal variance and first wave energy threshold were used to identify acoustically insufficient coupling data. Image histogram statistics were used to identify visible light overexposure, underexposure and infrared temperature saturation data. Missed detection points are filled by inverse distance weighted interpolation, and the detection point number, component number, spatial coordinates, acquisition timestamp, acquisition results of each mode, anomaly marker, interpolation marker, confidence level of each mode and overall confidence level are encapsulated to generate a multimodal detection data frame; In this embodiment, step two specifically includes: Terahertz time-domain spectral signals are read from multimodal detection data frames, and DC component removal, baseline correction, Butterworth bandpass filtering, and amplitude normalization are performed on the signals to obtain terahertz echo signals for interface identification. The terahertz echo signal is subjected to Hilbert transform to extract the echo envelope, and the reflection peaks in the echo envelope are obtained through a peak detection algorithm. An interface peak sequence is formed according to the arrival time of the reflection peaks. The interface positions corresponding to different reflection peaks are calculated using the time-of-flight method. Combined with the propagation velocity parameters of the finishing material, adhesive material, and base material, the finishing surface, finishing back, adhesive layer interface, and base layer interface are determined, and the finishing layer thickness and base layer interface depth are obtained. When there are overlapping reflection peaks, missing peaks at the base interface, or insufficient signal-to-noise ratio in the terahertz echo, the frequency-modulated continuous wave ranging results of the millimeter-wave radar are read, and the interface depth is corrected by combining the probe attitude angle and the distance to the detection surface. Based on the abnormal reflection peaks, echo amplitude ratios, and echo energy attenuation between the back of the finish and the base layer interface, the hollow area is determined. When the detection point shows air layer reflection characteristics, or when the echo energy attenuation of the base layer interface relative to the normal detection point of the same detection component exceeds the preset attenuation threshold, the detection point is marked as a suspected hollow point. Debonding areas are determined by the changes in the interface depth of the base layer at adjacent test points. When multiple adjacent test points in the same test component have continuous abrupt changes in interface depth and form a connected domain, suspected debonding areas are determined by connected domain analysis. Spatial mapping is performed on the thickness of the finishing layer and the depth of the base layer interface using inverse distance weighted interpolation or ordinary kriging interpolation. The suspected hollow points and suspected debonding areas are merged using a connected component labeling algorithm to generate a finishing layer perspective positioning map. The finishing layer perspective positioning map includes the distribution of finishing layer thickness, the location of the base layer interface, the suspected hollow areas, the suspected debonding areas, and the terahertz confidence level. In this embodiment, step three specifically includes: Read the perspective positioning map of the finish layer, and determine the effective detection area and interference detection area based on the thickness of the finish layer, the location of the base interface, the suspected hollow area, the suspected debonding area, and the terahertz confidence level. Extract the hyperspectral response features of the effective detection area and write interference markers into the interference detection area. Read the hyperspectral data corresponding to the effective detection area and the interference detection area from the multimodal detection data frame, perform dark current correction and white board correction on the raw hyperspectral data, and convert the sensor output value into reflectance data; Savitzky-Golay smoothing filter was applied to the reflectivity data, and standard normal variable transformation or L2 normalization was used to reduce the influence of changes in light intensity, probe distance and surface reflection on the spectral curve. Spectral features were extracted from the normalized hyperspectral curve, including full-band reflectance, characteristic spectral ratio, absorption peak position, absorption peak depth, half-maximum width, spectral slope, first derivative features, and second derivative features. The continuum removal method was used to enhance the absorption characteristics of the material's sensitive bands and extract the spectral response characteristics related to mortar aging, concrete moisture content, surface carbonization, and material degradation. Read the terahertz interface parameters of the same detection point in the surface perspective positioning map. The terahertz interface parameters include the surface layer thickness, the base layer interface depth, the echo energy attenuation, the hollow mark, the debonding mark, and the terahertz confidence level. Numerical normalization and one-heat encoding are performed on the terahertz interface parameters to convert continuous parameters into normalized values, and the hollow and debonding marks are converted into binary features. These features are then combined with the hyperspectral response features to form a spectral feature vector with interface constraint information. Read the rebound test value, penetration test value, core strength value, mortar grade, concrete strength grade, building age, structural form, finishing material type and base material type from the historical test database. The base material type is determined by design data, historical test records or on-site recording results. Bind the above data to the spectral feature vector of the corresponding test point. Sample confidence labels are generated based on the interference markers at the detection points, historical detection sources, and confidence levels of each modality. The interference markers include hollow markers, debonding markers, hyperspectral anomaly markers, and terahertz low confidence markers. The historical strength labels include core strength, penetration test value, rebound test value, mortar grade, and concrete strength grade. The spectral feature vector, terahertz interface parameters, base material type, historical strength label, sample confidence label and interference marker are encapsulated to generate interface constraint spectral samples, so that the terahertz interface parameters can be used as interface constraint information in the subsequent base strength inversion, rather than just as an independent detection result in the result summary. In this embodiment, step four specifically includes: The interface constraint spectral samples are read. The model training phase is pre-executed in the historical detection database. The on-site evaluation phase calls the trained interface constraint SpectralFormer model for inference. During the model training phase, the hyperspectral reflectance sequence of each detection point is divided into several continuous band groups according to the band order, and the continuous band groups, terahertz interface parameters, base material type, historical intensity labels, sample confidence labels, and interference markers are input into the interface-constrained SpectralFormer model; during the model inference phase, the continuous band groups, terahertz interface parameters, and base material type of the current detection point are input into the trained interface-constrained SpectralFormer model. The interface-constrained SpectralFormer model includes a band group embedding layer, an interface-aware location encoding layer, an interface-constrained multi-head self-attention layer, a cross-layer adaptive fusion layer, an interface gating layer, a sample weight calculation module, and a multi-task intensity output layer. Linear projection of continuous band groups is performed through the band group embedding layer to generate spectral group tokens, and the band position codes are written according to the order of the spectral group tokens in the hyperspectral reflectance sequence. The terahertz interface parameters are encoded, with the finishing layer thickness, base interface depth, echo energy attenuation and terahertz confidence being processed by maximum and minimum normalization, and the hollow mark and debonding mark being processed by binary encoding to obtain the interface parameter vector. The band position code and the interface parameter vector are concatenated through the interface-aware position coding layer and mapped to the interface-aware position code through the fully connected layer. The interface-aware position code is then superimposed on the spectral group token to generate the interface-aware spectral token. The interface-constrained multi-head self-attention layer maps the query vector, key vector, and value vector of the interface-aware spectral token, and maps the interface parameter vector to the attention bias matrix. The attention bias matrix is ​​added to the dot product of the query vector and the key vector to obtain the interface-constrained attention weight. The interface-constrained multi-head self-attention layer extracts local absorption peak features in adjacent bands, cross-band combination features, and material degradation response features, and writes the spectral features output by each coding layer into the cross-layer feature cache. The shallow absorption peak features, mid-layer spectral band combination features, and deep material degradation features of the outputs of different coding layers are read through the cross-layer adaptive fusion layer. The spectral features of each layer and the interface parameter vector are input into the Softmax weight layer to obtain the cross-layer fusion weight. The outputs of different coding layers are then weighted and fused according to the cross-layer fusion weight. The interface parameter vector is input into a gated network consisting of a fully connected layer, a ReLU activation function, and a Sigmoid activation function through an interface gating layer. This generates an interface gating vector with values ​​ranging from 0 to 1. The interface gating vector is then used to multiply the spectral features after cross-layer fusion element by element to generate interface-constrained spectral features. Through the above processing, the terahertz interface parameters are simultaneously involved in the interface sensing position encoding, attention bias matrix and interface gating vector generation, thereby constraining the hyperspectral intensity inversion process. The sample weight calculation module generates training sample weights based on the source of historical strength labels, terahertz confidence level, hyperspectral anomaly markers, hollowness markers, and debonding markers. The weight of core strength label samples is set to 1.00, penetration test label samples to 0.85, rebound test label samples to 0.70, and samples without measured strength labels to 0.30. Samples with hollowness markers, debonding markers, hyperspectral anomaly markers, or low-confidence terahertz markers have their strength supervision weights reduced according to the corresponding interference markers. These sample weights are determined by ranking the samples based on the average deviation of each type of historical strength label relative to the core strength results. The multi-task strength output layer outputs the interface constraint spectral features. This layer includes branches for mortar grade classification, concrete strength grade classification, continuous strength regression, strength confidence, and suspected insufficient strength probability. These branches are set up in parallel, each using the interface constraint spectral features corresponding to the same detection point as input, without redundant division of the interface constraint spectral features. Each output branch is configured with independent fully connected layer parameters, output dimensions, and activation functions to generate different types of strength assessment results.

[0022] The mortar grade classification branch takes interface-constrained spectral features as input, maps them to the mortar grade category space through a fully connected layer, and outputs the mortar grade category probability through the Softmax function. The mortar grade categories include those defined in the historical testing database. The concrete strength grade classification branch takes interface-constrained spectral features as input, maps them to the concrete strength grade category space through another set of fully connected layers, and outputs the concrete strength grade category probability through the Softmax function. The concrete strength grade categories include those defined in the historical testing database. The continuous strength regression branch takes interface-constrained spectral features as input, generates continuous strength values ​​through a fully connected layer and a linear output layer. The strength confidence branch takes interface-constrained spectral features as input, generates strength confidence through a fully connected layer and a Sigmoid function. The suspected strength deficiency probability branch takes interface-constrained spectral features as input, generates the suspected strength deficiency probability through a fully connected layer and a Sigmoid function. During the model training phase, the output branches participating in supervision are determined based on the type of base material and historical strength labels. When the historical strength label corresponds to the mortar grade, the mortar grade classification branch participates in classification supervision; when the historical strength label corresponds to the concrete strength grade, the concrete strength grade classification branch participates in classification supervision; when the historical strength label includes core drilling strength values, penetration test values, or rebound test values, the continuous strength regression branch participates in regression supervision; the strength confidence branch performs confidence supervision based on the source of the historical strength label, sample confidence labels, and interference markers; the strength deficiency suspected probability branch performs binary classification supervision based on the comparison results between the historical strength label and the preset strength qualification conditions. The mortar grade classification loss and concrete strength grade classification loss are calculated using weighted cross-entropy loss, the continuous strength regression loss is calculated using mean square error loss, the strength confidence loss is calculated using confidence calibration loss, the suspected strength deficiency probability loss is calculated using binary cross-entropy loss, and the gate constraint loss is calculated using interface gating regularization. The mortar grade classification loss, concrete strength grade classification loss, continuous strength regression loss, strength confidence loss, suspected strength deficiency probability loss, and gate constraint loss are weighted and summed according to their respective weights to obtain the total training loss. The total training loss is then multiplied by the training sample weights and used to update the parameters of the interface constraint SpectralFormer model. The interface constraint spectral characteristics of the same detection point are synchronously input into each output branch; when the base material type is mortar base, the output result of the mortar grade classification branch is used as the main result of the base strength level; when the base material type is concrete base, the output result of the concrete strength grade classification branch is used as the main result of the base strength level; the continuous strength regression branch outputs continuous strength values, the strength confidence branch outputs strength confidence, and the strength deficiency suspected probability branch outputs strength deficiency suspected probability; the base strength level, continuous strength values, strength confidence, and strength deficiency suspected probability are bound to the detection point coordinates, component number, and base material type, and spatial smoothing is used to generate a base strength inversion map.

[0023] In this embodiment, step five specifically includes: Read the base layer strength inversion map and the surface perspective positioning map, and delineate the suspected weak areas according to the strength confidence, the probability of insufficient strength, the suspected hollow area, the suspected debonding area, and the terahertz confidence. The strength confidence threshold and the probability of insufficient strength threshold are determined according to the receiver operating characteristic curve of historical verification samples. Using core drilling verification results or local chipping verification results as true labels, the false alarm rate and false negative rate under different thresholds are calculated, and the threshold with the false negative rate that meets the detection safety requirements and has the smallest overall error is selected. In this embodiment, detection points with an intensity confidence level lower than 0.60 are marked as low-confidence inversion areas, and detection points with an intensity deficiency probability of not less than 0.70 are marked as suspected weak areas. Both low-confidence inversion areas and suspected weak areas are included in the candidate range of three-dimensional acoustic verification. 0.60 and 0.70 are a set of engineering values ​​determined based on historical verification samples. Acoustic verification points are set up for the candidate range of three-dimensional acoustic verification. The verification range and the spacing between scanning points are determined based on the area of ​​the suspected weak area, the area of ​​the low confidence inversion area, the component size and the spacing between the acoustic phased array elements. At each acoustic verification point, the coupling status is checked. Coupling agent or dry coupling pad is used to make the three-dimensional acoustic phased array fit the detection surface. The first wave amplitude and A-scan signal variance are read. When the first wave amplitude is lower than the preset amplitude threshold or the A-scan signal variance is lower than the preset variance threshold, re-coupling and acquisition are performed again. A three-dimensional acoustic phased array is used for full matrix acquisition. Each transmitting element in the array is excited sequentially, and all receiving elements synchronously receive the echo signal to obtain full matrix acoustic data composed of the transmitting element number, receiving element number, sampling time, and echo amplitude. The full matrix acoustic data is processed by bandpass filtering, time zero-point correction, time gain compensation and amplitude normalization. The initial sound velocity is selected according to the material type of the detected component. The longitudinal wave sound velocity of concrete is used for concrete components, and the corresponding sound velocity of masonry or mortar base is used for masonry or mortar base. A three-dimensional reconstruction mesh is established within the verification range, voxel points are divided according to the detection resolution requirements, and the round-trip propagation time of the sound wave is calculated based on the propagation path from each voxel point to the transmitting and receiving array elements. The three-dimensional reconstructed mesh is imaged using the full-focusing method. The amplitude is extracted from the corresponding A-scan signal according to the round-trip propagation time of the sound wave. The amplitudes corresponding to all transmitting and receiving array elements are coherently superimposed to obtain the acoustic reflection intensity of each voxel. Threshold segmentation and connected component analysis were performed on the three-dimensional acoustic reflection intensity data. High reflection anomalies and low amplitude attenuation regions were extracted using the Otsu thresholding method or a fixed amplitude threshold. Morphological opening and closing operations were used to remove isolated noise points and connect broken regions. Defect types are marked according to the geometric shape of the abnormal body. When the abnormal body is distributed continuously in a linear or band-like manner, it is marked as a suspected area of ​​deep cracks. When the abnormal body is distributed in a sheet-like manner with low amplitude attenuation, it is marked as a suspected area of ​​loose base layer. When the abnormal body is a locally closed high reflection area, it is marked as a suspected area of ​​voids. When the abnormal body is in the same direction as the reinforcement of the component and the echo amplitude is concentrated, it is marked as a suspected area of ​​reinforcement abnormality. The three-dimensional acoustic imaging results are registered with the coordinates of the detection points and spatially superimposed with the base strength inversion map and the surface perspective positioning map. When there is a spatial intersection between the acoustic abnormal area and the suspected area of ​​insufficient strength, low confidence inversion area, suspected area of ​​hollowness or suspected area of ​​debonding, a verification matching mark is written. An acoustic verification map is generated based on the location of acoustic anomalies, defect type markers, maximum reflection amplitude, anomaly volume, anomaly depth, acoustic verification confidence level, and verification matching markers within each verification range. The acoustic verification map includes suspected areas of deep cracks, suspected areas of voids, suspected areas of loose base layers, suspected areas of steel reinforcement anomalies, acoustic verification confidence level, and verification matching markers. In this embodiment, step six specifically includes: Read multimodal detection data frames, surface perspective positioning maps, base layer strength inversion maps, and acoustic verification maps. Align the terahertz results, hyperspectral intensity inversion results, three-dimensional acoustic verification results, millimeter-wave ranging results, visible light surface conditions, and infrared temperature conditions according to the detection point numbers. For each detection point, an initial weight is configured for terahertz, hyperspectral, three-dimensional acoustic, millimeter wave, visible light, and infrared, with the sum of the six initial weights being 1. The initial weight for each modality is determined by normalizing the single-modal recognition accuracy of each modality for the corresponding risk item in the historical review samples. The terahertz confidence level, hyperspectral confidence level, three-dimensional acoustic confidence level, millimeter wave confidence level, visible light confidence level, and infrared confidence level are read for each detection point. The confidence level ranges from 0 to 1. The three-dimensional acoustic confidence level is taken from the acoustic verification confidence level in the acoustic verification map. The interpolation marker is used in the calculation of low confidence detection risk. Environmental correction coefficients are generated based on environmental conditions. When the area ratio of overexposed, underexposed, or specularly reflected regions in the visible light image reaches the preset interference judgment condition, the visible light environmental correction coefficient is 0.50, and the hyperspectral environmental correction coefficient is 0.70. When the infrared thermal image has temperature saturation, surface dampness, or rain interference, the infrared environmental correction coefficient is 0.50, and the terahertz environmental correction coefficient is 0.75. When the millimeter wave ranging result has abnormal reflections from metal mesh, water pipes, or embedded parts, the terahertz environmental correction coefficient is 0.60, the millimeter wave environmental correction coefficient is 1.20, and the three-dimensional acoustic environmental correction coefficient is 1.20. When the corresponding environmental condition is not triggered, the corresponding environmental correction coefficient is 1.00. The above values ​​are determined based on the consistency rate between each mode and the manual verification result under different environmental interference conditions. Modes with decreased consistency rate use a correction coefficient less than 1, while modes with increased consistency rate or those that can compensate for interference use a correction coefficient greater than 1. Calculate the modal fusion weight for each detection point, multiply the initial weight, confidence level, and environmental correction coefficient of the same modality to obtain the unnormalized weight of that modality, and then divide the unnormalized weight by the sum of the unnormalized weights of the six modalities at the same detection point to obtain the fusion weight of that modality. The risk values ​​of the test points are calculated based on the surface perspective positioning map, the base strength inversion map, and the acoustic verification map. The strength risk is obtained from the probability of insufficient strength, the hollow and debonding risk is obtained from the normalized risk values ​​of the hollow and debonding suspected areas, and the deep crack risk, the steel reinforcement abnormal risk, and the base loose and pore risk are obtained from the defect type marking, acoustic reflection intensity, abnormal body volume, and abnormal depth in the acoustic verification map. Each risk value is normalized to 0 to 1. The strength consistency coefficient and the hollow debonding consistency coefficient are calculated based on the verification matching marks. When the strength risk at the same test point is not less than 0.70 and the risk of loose porosity in the base layer is not less than 0.70, and the verification matching marks show that there is a spatial intersection between the two, the strength consistency coefficient is 1.20; when the strength risk is not less than 0.70 but the risk of loose porosity in the base layer is less than 0.70, the strength consistency coefficient is 0.80; in other cases, the strength consistency coefficient is 1.00. When the risk of debonding and hollowing at the same test point is not less than 0.70 and the risk of loose pores in the base layer is not less than 0.70, and the verification matching mark shows that there is a spatial intersection between the two, the consistency coefficient of debonding and hollowing is 1.15; when the risk of debonding and hollowing is not less than 0.70 but the risk of loose pores in the base layer is less than 0.70, the consistency coefficient of debonding and hollowing is 0.85; in other cases, the consistency coefficient of debonding and hollowing is 1.00; the above strength consistency coefficient and debonding and hollowing consistency coefficient are determined based on the actual defect confirmation rate of common anomalies and single-modal anomalies in the multimodal cross-validation samples; Therefore, the acoustic verification confidence and verification matching mark in the acoustic verification diagram are used in reverse to participate in the calculation of the comprehensive risk of intensity and the comprehensive risk of delamination and debonding, avoiding the fixed weighting and summarization of only the terahertz detection results, hyperspectral inversion results and acoustic verification results. To calculate the comprehensive risk of the intensity at the detection point, the hyperspectral fusion weight is multiplied by the intensity risk, and the three-dimensional acoustic fusion weight is multiplied by the risk of loose and porous base layer. The two products are then added together and multiplied by the intensity consistency coefficient. To calculate the comprehensive risk of hollowing and debonding at the detection point, multiply the terahertz fusion weight by the risk of hollowing and debonding, multiply the three-dimensional acoustic fusion weight by the risk of loose and porous base layer, add the two products together, and then multiply by the hollowing and debonding consistency coefficient. To calculate the comprehensive risk of deep cracks at the detection point, multiply the 3D acoustic fusion weights by the deep crack risk; to calculate the comprehensive risk of rebar anomalies at the detection point, multiply the 3D acoustic fusion weights by the rebar anomaly risk; to calculate the low-confidence detection risk at the detection point, multiply the six modal fusion weights by their corresponding confidence levels and sum them, then subtract the sum from 1. When interpolation markers exist at the detection point, add an interpolation penalty term to the low-confidence detection risk. The risks of the test points are aggregated according to the component number. The comprehensive strength risk, comprehensive hollow and debonding risk, comprehensive deep crack risk, comprehensive steel reinforcement abnormality risk and low confidence test risk of each test point in the same test component are respectively averaged by area weighting to obtain the comprehensive strength risk, comprehensive hollow and debonding risk, comprehensive deep crack risk, comprehensive steel reinforcement abnormality risk and low confidence test risk of the component. The risk weights for each component are configured according to its type. When the component is a load-bearing wall, frame column, or main beam, the strength risk weight is 0.35, the hollow and debonded risk weight is 0.10, the deep crack risk weight is 0.30, the abnormal reinforcement risk weight is 0.15, and the low-confidence detection risk weight is 0.10. When the component is a non-load-bearing partition wall, exterior wall cladding, or decorative panel, the strength risk weight is 0.15, the hollow and debonded risk weight is 0.45, the deep crack risk weight is 0.15, the abnormal reinforcement risk weight is 0.05, and the low-confidence detection risk weight is 0.20. When the component is a floor slab or general wall, the strength risk weight is 0.25, the hollow and debonded risk weight is 0.20, the deep crack risk weight is 0.25, the abnormal reinforcement risk weight is 0.15, and the low-confidence detection risk weight is 0.15. The above values ​​are determined based on the component type, load-bearing capacity, defect consequences, and historical maintenance records, and the sum of all risk weights is 1. To calculate the component safety score, multiply the component's overall strength risk by its strength risk weight, multiply the component's overall hollow and debonding risk by its hollow and debonding risk weight, multiply the component's overall deep crack risk by its deep crack risk weight, multiply the component's overall steel reinforcement abnormality risk by its steel reinforcement abnormality risk weight, and multiply the component's low-confidence detection risk by its low-confidence detection risk weight. Sum the five products and multiply by 100 to get the overall deduction value. Then, subtract the overall deduction value from 100 to get the component safety score. When the safety score of a component is not lower than 85, the component will be marked as Grade A; when the safety score of a component is not lower than 70 but lower than 85, the component will be marked as Grade B; when the safety score of a component is not lower than 60 but lower than 70, the component will be marked as Grade C; when the safety score of a component is lower than 60, the component will be marked as Grade D. The component number, component type, comprehensive risk of component strength, comprehensive risk of component hollowness and debonding, comprehensive risk of component deep cracks, comprehensive risk of component abnormal reinforcement, risk of component low reliability detection, component safety score and safety level are encapsulated to generate a component safety status table; In this embodiment, step seven specifically includes: Read the component safety status table, inspection point coordinates, component number, surface perspective positioning diagram, base strength inversion diagram and acoustic verification diagram, and import the building BIM model, CAD drawings, floor plan or building elevation; Based on the on-site control points, component boundary points, and the starting point of the inspection, establish the transformation relationship between the inspection coordinate system and the building model coordinate system. Rigid body transformation is used for the three-dimensional model, and affine transformation is used for the two-dimensional drawings. Map the coordinates of the detection points to the surface of the corresponding components, and overlay the strength display layer, the hollow debonding display layer, the internal defect display layer and the low confidence detection layer according to the component number to generate a safety assessment heat map. In the safety assessment heat map, the base strength grade, continuous strength value and strength confidence are mapped to the strength risk color scale, the suspected hollow area and the suspected debonding area are mapped to the area mark, and the suspected deep crack area, suspected hole area, suspected loose base area and suspected abnormal steel reinforcement area are mapped to the defect mark. The low confidence inversion area and low confidence detection point are mapped to the gray mark. Maintenance priorities are generated based on component safety scores and safety levels. Grade D components are set as Level 1, Grade C components with suspected deep cracks or suspected abnormal rebar areas are set as Level 2, Grade C components with only suspected hollow areas or suspected debonding areas are set as Level 3, and Grade A and B components are set as routine inspections. Generate a building safety performance assessment report according to the structured report template, and write in the inspection object, inspection scope, number of inspection points, component safety status table, safety assessment heat map, base layer strength distribution, hollow and debonded distribution, internal defect distribution, safety level, maintenance priority and review point list; The results of manual review, core drilling verification, partial chipping verification, repair and re-inspection are added to the historical inspection database according to the inspection batch, inspection point number and component number to form a traceable historical inspection sample sequence. The model parameters are iteratively updated based on this sample sequence without overwriting the existing historical inspection data. When the number of newly added verification samples reaches the preset update threshold, the training sample set, sample confidence labels, dynamic weight parameters, risk judgment threshold, and interface constraint SpectralFormer model parameters of the interface constraint model are updated based on the historical test sample sequence. The model version number, update date, and applicable building type are also written. The preset update threshold is determined based on the minimum number of training samples after sample grouping. The judgment is based on whether the newly added verification samples in the same group can cover the mortar grade, concrete strength grade, hollow, debonding, crack, hole, loose base layer, and abnormal steel reinforcement categories defined in the historical test database. Safety assessment heatmaps, building safety performance assessment reports, historical inspection database updates, and model parameter updates are used for result expression, record verification, and sample accumulation. The safety status of components is mainly determined based on the component safety score and safety level generated in step six.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to a safety inspection scenario prior to the renovation of an existing office building. The building is a six-story frame structure. The inspection scope included the walls of the corridors from the first to the third floor, the exterior wall finishes, columns, and parts of the floor slabs, covering an area of ​​approximately 1260 square meters. This area was divided into 312 inspection grids, forming 1846 inspection points. A five-in-one probe was used for grid scanning on-site, with a grid spacing of 0.40 meters and a verification spacing of 0.20 meters for suspected weak areas. The strength confidence threshold was set at 0.60, the suspected insufficient strength probability threshold at 0.70, and the risk consistency judgment threshold at 0.70. After the inspection, 38 core drilling or rebound verification points, 42 partial chipping verification points, and 96 three-dimensional acoustic verification points were selected as verification samples. The results of manual verification, core drilling verification, partial chipping verification, and post-renovation inspection were used as the comparison basis.

[0025] In this embodiment, Method A is a terahertz and millimeter-wave interface detection method, mainly used to identify the thickness of the finishing layer, the depth of the base layer interface, suspected areas of hollowness, and suspected areas of debonding; Method B is a base layer strength estimation method combining hyperspectral imaging and a traditional regression model, mainly predicting the base layer strength based on the surface spectral response; Method C is a three-dimensional acoustic phased array detection method, mainly identifying deep cracks, pores, and loose base layers through full matrix acquisition and full-focus reconstruction; Method D is a method combining a common SpectralFormer model with fixed weight fusion, which does not introduce interface constraint modulation of terahertz interface parameters, nor dynamically adjust the fusion weights based on acoustic verification credibility and verification matching markers. The method of this invention first generates multimodal detection data frames and a finishing layer perspective positioning map, then embeds terahertz interface parameters into hyperspectral response features to generate interface constraint spectral samples, and generates a base layer strength inversion map through an interface constraint SpectralFormer model; subsequently, it performs three-dimensional acoustic verification on suspected weak areas, and generates a component safety status table and a safety assessment heatmap through environmentally adaptive dynamic weight fusion.

[0026] Table 1. Comparison of different detection methods in the safety assessment of existing office buildings.

[0027] Table 1 shows that Method A has advantages in identifying hollow and debonded surfaces, with a recall rate of 91.8%, indicating that terahertz echo combined with millimeter-wave ranging is suitable for identifying anomalies at the interface between the finishing layer and the base layer. However, this method is insufficient in reflecting the base layer strength and deep defects, with an average absolute error of 6.4 MPa for base layer strength and a recall rate of only 42.6% for deep defects. Method B is better than simple interface detection in estimating base layer strength, with an average absolute error of 4.9 MPa. However, since it does not introduce terahertz interface parameters, the surface spectral response is prone to deviating from the true state of the base layer when encountering changes in finishing layer thickness, hollow surfaces, and debonding, resulting in a low-confidence detection point ratio of 18.5%. Method C is more sensitive to deep cracks, pores, and base layer looseness, with a deep defect recognition recall rate of 93.2%. However, its average detection time per component is 12.7 minutes, and its ability to directly judge the base layer strength level is insufficient.

[0028] Compared with Method D, the average absolute error of the base layer strength of the present invention is reduced from 3.8 MPa to 2.6 MPa, and the consistency rate of strength level is increased from 83.6% to 91.7%. This indicates that after the terahertz interface parameters participate in the strength inversion through interface-aware position encoding, attention bias matrix, and interface gating vector, the influence of surface occlusion and interface debonding on hyperspectral intensity judgment can be weakened. The overall safety level consistency rate is increased from 85.1% to 93.4%, and the false alarm rate is reduced from 11.8% to 6.5%. This indicates that after the acoustic verification confidence and verification matching mark participate in the environmental adaptive dynamic weight fusion, it can correct the misjudgment of a single mode. The proportion of low confidence detection points is reduced from 9.6% to 5.1%, indicating that the present invention can adjust the contribution of each mode through confidence and environmental correction coefficients in the field environment with interpolation points, surface reflection, local moisture, and interference from embedded parts.

[0029] Based on the results of the embodiments, this invention can generate a surface perspective positioning map, a base strength inversion map, an acoustic verification map, a component safety status table, and a safety assessment heat map in the same testing process, enabling engineers to directly locate components from Grade A to Grade D and their corresponding risk types. For existing building renovation, maintenance inspections, and safety verification scenarios, this invention can reduce missed detections and false alarms caused by single testing methods, improve the stability of component safety scores, and provide quantitative basis for subsequent maintenance priority determination, retesting point layout, and historical testing database updates.

[0030] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing building safety performance based on big data analysis, characterized in that, Includes the following steps: Step 1: Use a five-in-one probe to perform a gridded scan of the area of ​​the building to be evaluated. After spatial calibration of heterogeneous sensors, time synchronization, anomaly sampling rejection and confidence marking, multimodal detection data frames are generated. Step 2: Filter, extract the envelope, and identify the interface peaks in the terahertz echoes in the multimodal detection data frames. Combine the millimeter-wave ranging results to correct the interface depth and generate a surface perspective positioning map. Step 3: Determine the effective detection area based on the surface perspective positioning map, perform radiometric calibration, normalization and spectral feature extraction on the hyperspectral response, and embed the terahertz interface parameters into the spectral features to generate interface-constrained spectral samples; Step 4: Input the interface-constrained spectral samples into the interface-constrained SpectralFormer model, and generate the base layer intensity inversion map through spectral embedding, band coding, spectral feature extraction and interface constraint modulation; Step 5: Based on the base strength inversion map and the surface perspective positioning map, identify suspected weak areas, and use a three-dimensional acoustic phased array to perform full matrix acquisition and full-focus reconstruction to generate an acoustic verification map; Step 6: Perform environmental adaptive dynamic weight fusion based on the surface perspective positioning map, the base strength inversion map, and the acoustic verification map to calculate the component safety score and generate the component safety status table; Step 7: Map the component safety status table to the building model or drawings, generate a safety assessment heatmap, output a building safety performance assessment report, and add the current test results and review results to the historical test database, and update the model parameters based on the historical test samples.

2. The building safety performance assessment method based on big data analysis according to claim 1, characterized in that, Step one specifically includes: Establish a building inspection coordinate system based on the CAD drawings, BIM model or on-site measured outline of the building to be evaluated, divide the area to be evaluated into wall, beam, slab, column, finishing unit or masonry unit, and write the component number, floor number, inspection surface number and scanning path. A five-in-one probe is used to perform gridded scanning of the area to be evaluated. The five-in-one probe includes a terahertz radar, a hyperspectral camera, a three-dimensional acoustic phased array, a millimeter-wave radar, and a visible-infrared composite imaging unit. Spatial calibration and time synchronization were performed on the five-in-one probe to obtain the external parameter matrix of each sensor and collect the timestamp uniformly. Set the gridded scanning parameters according to the surface size, structural importance, and on-site risk area distribution of the component being inspected, and record the probe's spatial coordinates, attitude angle, and inspection point number; Anomaly identification and credibility marking were performed on the collected data to identify terahertz echo saturated data, hyperspectral anomalous bands, insufficient acoustic coupling data, visible light overexposure and underexposure data, and infrared temperature saturation data. The missed detection points are filled by interpolation, and the detection point number, component number, spatial coordinates, acquisition timestamp, acquisition results of each mode, anomaly marker, interpolation marker and confidence level of each mode are encapsulated to generate a multimodal detection data frame.

3. The method for assessing building safety performance based on big data analysis according to claim 1, characterized in that, Step two specifically includes: The terahertz time-domain spectral signal in the multimodal detection data frame is subjected to DC component removal, baseline correction, bandpass filtering and amplitude normalization to obtain the terahertz echo signal; Envelope extraction is performed on the terahertz echo signal, and the reflection peaks in the echo envelope are obtained by peak detection. An interface peak sequence is formed according to the arrival time of the reflection peaks. The interface positions corresponding to different reflection peaks are calculated based on the time-of-flight method. Combined with the propagation velocity parameters of the finishing material, adhesive material and base material, the finishing surface, finishing back, adhesive layer interface and base layer interface are determined, and the finishing layer thickness and base layer interface depth are obtained. When there are overlapping reflection peaks, missing peaks at the base interface, or insufficient signal-to-noise ratio in the terahertz echo, the interface depth is corrected by combining the millimeter-wave ranging results, probe attitude angle, and distance to the detection surface. Based on the abnormal reflection peaks, echo amplitude ratios, and echo energy attenuation between the back of the finish and the interface with the base layer, suspected hollow points are identified; based on the changes in the depth of the base layer interface between adjacent detection points, suspected debonding areas are identified. Spatial mapping was performed on the thickness of the finishing layer and the depth of the base layer interface. Suspected hollow points and suspected delamination areas were merged to generate a perspective positioning map of the finishing layer.

4. The building safety performance assessment method based on big data analysis according to claim 1, characterized in that, Step three specifically includes: Based on the thickness of the finishing layer, the location of the base layer interface, the suspected hollow area, the suspected debonding area, and the terahertz confidence level in the finishing layer perspective positioning diagram, the effective detection area and the interference detection area are determined, and interference markers are written into the interference detection area. Dark current correction and white board correction are performed on the hyperspectral data corresponding to the effective detection area and the interference detection area, and the sensor output value is converted into reflectance data; The reflectance data were smoothed and normalized, and the spectral features of the normalized hyperspectral curves were extracted to obtain the hyperspectral response features. Absorption feature enhancement is performed on the material's sensitive wavelength bands to extract spectral response features related to mortar aging, concrete moisture content, surface carbonization, and material degradation. The terahertz interface parameters at the same detection point are numerically normalized and encoded, and then combined with the hyperspectral response features to form a spectral feature vector with interface constraint information. The spectral feature vector, terahertz interface parameters, base material type, historical strength label, sample confidence label, and interference marker are encapsulated to generate interface constraint spectral samples, so that the terahertz interface parameters can be used as interface constraint information in subsequent base strength inversion.

5. The method for assessing building safety performance based on big data analysis according to claim 1, characterized in that, Step four specifically includes: The hyperspectral reflectance sequence in the interface-constrained spectral sample is divided into several continuous band groups according to the band order. During the model training phase, the SpectralFormer model is constrained by inputting continuous band groups, terahertz interface parameters, base material type, historical intensity labels, sample confidence labels, and interference markers into the interface. During the model inference phase, the continuous band group, terahertz interface parameters, and base material type of the current detection point are input into the trained interface constraint SpectralFormer model. The interface-constrained SpectralFormer model includes a band group embedding layer, an interface-aware location encoding layer, an interface-constrained multi-head self-attention layer, a cross-layer adaptive fusion layer, an interface gating layer, a sample weight calculation module, and a multi-task intensity output layer. A spectral group token is generated through a band group embedding layer, an interface-aware spectral token is generated through an interface-aware position coding layer, spectral features are extracted through an interface-constrained multi-head self-attention layer, cross-layer fused spectral features are generated through a cross-layer adaptive fusion layer, and interface-constrained modulation is performed on the cross-layer fused spectral features through an interface gating layer to generate interface-constrained spectral features. Based on the interface-constrained spectral characteristics, the base layer strength grade, continuous strength value, strength confidence level, and suspected probability of insufficient strength are output, and bound to the detection point coordinates, component number, and base layer material type to generate a base layer strength inversion map.

6. The method for assessing building safety performance based on big data analysis according to claim 5, characterized in that, The interface constraint modulation specifically includes: The terahertz interface parameters are encoded to obtain an interface parameter vector. The terahertz interface parameters include the finishing layer thickness, the base layer interface depth, the echo energy attenuation, the hollow mark, the debonding mark, and the terahertz confidence level. The band position code and the interface parameter vector are concatenated through the interface-aware position coding layer and mapped to the interface-aware position code. The interface-aware position code is then superimposed on the spectral group token to generate the interface-aware spectral token. The interface-constrained multi-head self-attention layer maps the query vector, key vector, and value vector of the interface-aware spectral token, and maps the interface parameter vector to the attention bias matrix. The attention bias matrix is ​​added to the dot product of the query vector and the key vector to obtain the interface-constrained attention weight. By fusing shallow absorption peak features, mid-layer spectral band combination features, deep material degradation features, and interface parameter vectors through a cross-layer adaptive fusion layer, the spectral features after cross-layer fusion are obtained. The interface parameter vector is input into the gated network through the interface gated layer to generate the interface gated vector. The interface gated vector is then used to multiply the spectral features after cross-layer fusion element by element to generate the interface constrained spectral features. The sample weight calculation module generates training sample weights based on the historical intensity label source, terahertz confidence, hyperspectral anomaly label, hollow label, and debonding label. The multi-task strength output layer is configured with mortar grade classification branches, concrete strength grade classification branches, continuous strength regression branches, strength confidence branches, and strength deficiency probability branches. During the model training phase, the total training loss is generated based on the mortar grade classification loss, concrete strength grade classification loss, continuous strength regression loss, suspected insufficient strength probability loss, strength confidence loss, and gating constraint loss, and the model parameters are updated in combination with the training sample weights.

7. The method for assessing building safety performance based on big data analysis according to claim 1, characterized in that, Step five specifically includes: Based on the strength confidence and suspected strength insufficiency probability in the base strength inversion map, and the suspected hollow area, suspected debonding area and terahertz confidence in the surface perspective positioning map, the candidate range for three-dimensional acoustic testing is determined. When the intensity confidence level is lower than the intensity confidence level threshold, the corresponding detection point is marked as a low confidence inversion area; when the suspected intensity deficiency probability is not lower than the suspected intensity deficiency probability threshold, the corresponding detection point is marked as a suspected weak area. Both the low confidence inversion area and the suspected weak area are included in the candidate range of 3D acoustic verification. Acoustic verification points are set up for the candidate range of three-dimensional acoustic verification, and the coupling status of each acoustic verification point is checked. A three-dimensional acoustic phased array is used for full matrix acquisition. Each transmitting element in the array is excited sequentially, and all receiving elements synchronously receive the echo signal to obtain full matrix acoustic data. The full matrix acoustic data is filtered, time-corrected, time-gain compensated, and amplitude-normalized, and a three-dimensional reconstruction mesh is established within the verification range. The three-dimensional reconstructed mesh was imaged using the full-focusing method to obtain the acoustic reflection intensity of each voxel. Based on the three-dimensional acoustic reflection intensity data, threshold segmentation, connected component analysis and defect type labeling are performed to identify suspected areas of deep cracks, holes, loose base layers and abnormal reinforcement. The three-dimensional acoustic imaging results are spatially overlaid with the base strength inversion map and the surface perspective positioning map to generate acoustic verification confidence and verification matching marks, and form an acoustic verification map.

8. The method for assessing building safety performance based on big data analysis according to claim 1, characterized in that, Step six specifically includes: The terahertz results, hyperspectral intensity inversion results, three-dimensional acoustic calibration results, millimeter-wave ranging results, visible light surface condition and infrared temperature condition are aligned according to the detection point number. For each detection point, configure terahertz initial weight, hyperspectral initial weight, three-dimensional acoustic initial weight, millimeter wave initial weight, visible light initial weight and infrared initial weight, and the sum of the six types of initial weights is 1; The confidence level of each mode is determined based on terahertz confidence level, hyperspectral confidence level, three-dimensional acoustic confidence level, millimeter wave confidence level, visible light confidence level and infrared confidence level, wherein the three-dimensional acoustic confidence level is taken from the acoustic verification confidence level in the acoustic verification diagram. An environmental correction coefficient is generated based on the environmental state, and an environmental adaptive dynamic weight fusion is performed based on the initial weight, confidence level, and environmental correction coefficient of the same modality to obtain the modality fusion weight. The risk values ​​of the test points are calculated based on the surface perspective positioning diagram, the base strength inversion diagram and the acoustic verification diagram. The risk values ​​of the test points include strength risk, hollow debonding risk, deep crack risk, abnormal steel reinforcement risk and loose and porous base layer risk. Calculate the strength consistency coefficient and the hollow debonding consistency coefficient based on the verification matching mark; Based on the modal fusion weight, the risk value of the detection point, the strength consistency coefficient, and the hollow debonding consistency coefficient, the comprehensive risk of the detection point strength, the comprehensive risk of the detection point hollow debonding, the comprehensive risk of the detection point deep cracks, the comprehensive risk of the detection point abnormal reinforcement, and the low confidence detection risk of the detection point are calculated. Among them, the low confidence detection risk of the detection point is calculated based on the confidence of each modality and the interpolation mark. The risks of the detection points are aggregated according to the component number to obtain the comprehensive risk of component strength, the comprehensive risk of component hollowing and debonding, the comprehensive risk of component deep cracks, the comprehensive risk of component abnormal reinforcement, and the risk of component low-reliability detection. Configure component risk weights according to the type of the component being detected, calculate component safety scores based on component risk weights and the risks of each component, generate safety levels based on component safety scores, and encapsulate component safety scores and safety levels into a component safety status table.

9. The method for assessing building safety performance based on big data analysis according to claim 8, characterized in that, The environmental adaptive dynamic weight fusion specifically includes: The initial weights for each modality are determined by normalizing the single-modal recognition accuracy of each modality for the corresponding risk item in the historical review samples; When the area ratio of overexposed, underexposed, or specularly reflected regions in a visible light image reaches the preset interference judgment condition, the visible light environment correction coefficient is a value greater than 0 and less than 1, and the hyperspectral environment correction coefficient is a value greater than 0 and less than 1. When the infrared thermal image is subject to temperature saturation, surface dampness, or rain interference, the infrared environment correction factor is taken as a value greater than 0 and less than 1, and the terahertz environment correction factor is taken as a value greater than 0 and less than 1. When millimeter-wave ranging results show abnormal reflections from metal mesh, water pipes, or embedded parts, the terahertz environment correction factor takes a value greater than 0 and less than 1, while the millimeter-wave environment correction factor and the three-dimensional acoustic environment correction factor take a value greater than 1. Multiply the initial weights, confidence levels, and environmental correction coefficients of the same modality to obtain the unnormalized weights of that modality. Then, divide the unnormalized weights by the sum of the unnormalized weights of the six modalities at the same detection point to obtain the fusion weights of that modality. Strength risk is obtained from the probability of suspected strength deficiency; hollow and debonding risk is obtained from the normalized risk value of suspected hollow and debonding areas; deep crack risk, steel reinforcement abnormality risk and base layer looseness and void risk are obtained by normalizing the defect type marking, acoustic reflection intensity, abnormal body volume and abnormal depth in the acoustic verification diagram. When the strength risk at the same test point is not lower than the preset risk threshold, the risk of loose soil and pores in the base layer is not lower than the preset risk threshold, and the verification matching mark shows that there is a spatial intersection between the two, the strength consistency coefficient takes a value greater than 1; when the strength risk is not lower than the preset risk threshold but the risk of loose soil and pores in the base layer is lower than the preset risk threshold, the strength consistency coefficient takes a value greater than 0 and less than 1. When the risk of delamination and detachment at the same detection point is not lower than the preset risk threshold, the risk of loose pores in the base layer is not lower than the preset risk threshold, and the verification matching mark shows that there is a spatial intersection between the two, the consistency coefficient of delamination and detachment is greater than 1; when the risk of delamination and detachment is not lower than the preset risk threshold but the risk of loose pores in the base layer is lower than the preset risk threshold, the consistency coefficient of delamination and detachment is greater than 0 and less than 1. The comprehensive risk of the intensity of the detection point is determined by the hyperspectral fusion weight, intensity risk, three-dimensional acoustic fusion weight, risk of loose and porous base layer, and intensity consistency coefficient. The comprehensive risk of delamination and debonding at the detection point is jointly determined by the terahertz fusion weight, the risk of delamination and debonding, the three-dimensional acoustic fusion weight, the risk of loose and porous base layer, and the consistency coefficient of delamination and debonding.

10. The method for assessing building safety performance based on big data analysis according to claim 1, characterized in that, Step seven specifically includes: Map the component safety status table, inspection point coordinates, component number, surface perspective positioning diagram, base strength inversion diagram, and acoustic verification diagram to the building BIM model, CAD drawings, floor plan, or building elevation. Establish the transformation relationship between the detection coordinate system and the building model coordinate system based on the on-site control points, component boundary points and detection start points, and map the detection point coordinates to the corresponding component surface; A safety assessment heatmap is generated by overlaying strength display layers, hollow debonding display layers, internal defect display layers, and low-confidence detection layers according to component numbers. Maintenance priorities are generated based on component safety scores and safety levels; Generate a building safety performance assessment report according to the structured report template. The report content includes the inspection object, inspection scope, number of inspection points, component safety status table, safety assessment heat map, base layer strength distribution, hollow and debonded distribution, internal defect distribution, safety level, maintenance priority, and a list of review points. The results of manual review, core drilling verification, partial chipping verification, repair and re-inspection are added to the historical inspection database according to the inspection batch, inspection point number and component number to form a traceable historical inspection sample sequence. When the number of newly added verification samples reaches the preset update threshold, the training sample set, sample confidence labels, dynamic weight parameters, risk judgment threshold, and interface constraint SpectralFormer model parameters of the interface constraint model are updated based on the historical detection sample sequence.

Images