External wall insulation falling risk prediction method and system based on deep learning model

Abstract

The embodiment of the invention discloses an external wall thermal insulation falling risk prediction method and system based on a deep learning model, and the method comprises the steps: firstly collecting the multi-dimensional state data of a target building external wall, covering the thermal insulation layer material attribute, environment exposure history, structure connection strength, and surface deformation monitoring and construction technology record data; extracting a material aging feature vector from the thermal insulation layer material attribute data, and extracting a space-time distribution feature from the environmental exposure historical data to generate an environmental action feature tensor; fusing structural connection strength and surface deformation monitoring data into a structural deformation association map, converting construction process record data into process defect probability distribution, inputting the data into a pre-trained multi-modal risk prediction network, outputting an outer wall thermal insulation layer falling risk level and risk area thermodynamic diagram, and predicting the falling risk level of the outer wall thermal insulation layer. And finally, generating a dynamic monitoring strategy containing a sensor deployment optimization scheme and a detection period adjustment parameter according to a result, and realizing accurate risk prediction and efficient monitoring.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and system for predicting the risk of exterior wall insulation falling off based on a deep learning model. Background Art

[0002] With the continuous development of the construction industry, the exterior wall insulation system plays a vital role in building energy conservation and insulation. However, the problem of exterior wall insulation falling off frequently occurs, which not only affects the beauty and insulation performance of the building, but also poses a serious threat to the safety of personnel and property. Therefore, accurately predicting the risk of exterior wall insulation falling off and taking effective preventive measures have become important issues that need to be solved in the construction field.

[0003] In the existing technology, there are many limitations in the assessment methods for the risk of exterior wall insulation falling off. Some traditional methods mainly rely on manual experience and simple detection methods, and obtain limited information through on-site observation or basic physical testing, which makes it difficult to fully and accurately grasp the overall condition of the exterior wall insulation system. This method is not only highly subjective, but also unable to conduct in-depth analysis of potential risk factors, omitting a lot of key information, resulting in inaccurate risk assessment results and difficulty in providing reliable risk warnings.

[0004] Although some data-based methods have introduced certain data analysis technologies, most of the existing risk assessment methods are static assessments, lacking the ability to dynamically monitor and evaluate building changes over time. During the long-term use of a building, its exterior wall insulation system will be affected by the continuous effects of various environmental factors and the slow changes in its own structure. Static assessment methods cannot reflect these dynamic changes in a timely manner, making it difficult to detect new risk points in a timely manner, and thus unable to adjust monitoring and prevention measures in a timely manner, resulting in a significant reduction in the timeliness and effectiveness of risk prevention and control. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a method and system for predicting the risk of exterior wall insulation falling off based on a deep learning model.

[0006] In a first aspect, an embodiment of the present invention provides a method for predicting the risk of external wall thermal insulation falling off based on a deep learning model, which is applied to a system for predicting the risk of external wall thermal insulation falling off based on a deep learning model, comprising: Collecting a multi-dimensional state data set of the target building exterior wall, the multi-dimensional state data set including insulation layer material property data, environmental exposure history data, structural connection strength data, surface deformation monitoring data and construction process record data; Extracting material degradation characteristics from the insulation layer material property data to generate a material aging characteristic vector, and extracting spatiotemporal distribution characteristics from the environmental exposure history data to generate an environmental action characteristic tensor; Spatially align and fuse the structural connection strength data and the surface deformation monitoring data to generate a structural deformation correlation map, and convert the construction process record data into a process defect probability distribution; Input the material aging feature vector, the environmental action feature tensor, the structural deformation correlation map, and the process defect probability distribution into a pre-trained multi-modal risk prediction network to output the risk level of the external wall thermal insulation layer shedding and the heat map of the risk area; Generate a dynamic monitoring strategy based on the risk level and the heat map of the risk area, where the dynamic monitoring strategy includes an optimized sensor deployment plan and detection cycle adjustment parameters.

[0007] In a second aspect, an embodiment of the present invention provides an external wall thermal insulation shedding risk prediction system based on a deep learning model, including: A processor; A memory, in which a computer program is stored, and when the computer program is executed, it implements the external wall thermal insulation shedding risk prediction method described in the first aspect based on a deep learning model.

[0008] As described above, in the embodiment of the present invention, in the data acquisition layer, a multi-dimensional state data set of the external wall of the target building is comprehensively collected, covering the thermal insulation layer material attribute data, the environmental exposure history data, the structural connection strength data, the surface deformation monitoring data, and the construction process record data, which can comprehensively depict the actual situation of the external wall thermal insulation system from multiple key perspectives. Secondly, unique data feature processing is performed on different types of data. The material degradation characteristics of the thermal insulation layer material attribute data are extracted to generate a material aging feature vector, which accurately captures the aging trend of the material with time and environmental factors; the spatio-temporal distribution characteristics of the environmental exposure history data are extracted to generate an environmental action feature tensor, which comprehensively considers the environmental factors in the time and space dimensions, comprehensively shows the dynamic impact of environmental factors on the external wall thermal insulation system, effectively excavates the potential laws behind the data, and provides more representative and valuable feature information for subsequent risk analysis.

[0009] Furthermore, in the data fusion link, the structural connection strength data and the surface deformation monitoring data are spatially aligned and fused to generate a structural deformation correlation map, and the construction process record data is converted into a process defect probability distribution. By deeply exploring the internal relationship between different data, a map and a probability distribution that can accurately reflect the complex relationship between the structure and performance of the external wall thermal insulation system are constructed, providing a more intuitive and comprehensive perspective for risk prediction and greatly improving the accuracy of risk assessment.

[0010] Then, input the material aging feature vector, environmental action feature tensor, structural deformation correlation map, and process defect probability distribution obtained after processing into a pre-trained multi-modal risk prediction network to output the risk level of the external wall thermal insulation layer shedding and the thermal map of the risk area. This multi-modal risk prediction network integrates data features of multiple different modalities, fully leveraging the advantages of deep learning models in processing complex data and non-linear relationships, and can more accurately analyze the comprehensive impact of various factors on the risk of external wall thermal insulation shedding, thereby achieving high-precision prediction of the risk of external wall thermal insulation shedding, accurately identifying the risk level and the distribution of the risk area, and providing clear guidance for subsequent maintenance and management.

[0011] Finally, generate a dynamic monitoring strategy based on the risk level and the thermal map of the risk area, including an optimized sensor deployment plan and detection cycle adjustment parameters. By dynamically adjusting the sensor deployment and detection cycle, it is possible to ensure the effective monitoring of the external wall thermal insulation condition while reasonably allocating monitoring resources, avoiding waste of unnecessary monitoring costs, achieving the optimal balance between monitoring efficiency and resource utilization, and providing intelligent and efficient safeguard measures for the long-term stable operation of the building external wall thermal insulation system. Brief Description of the Drawings

[0012] Figure 1 It is a schematic flowchart of the steps of a method for predicting the risk of external wall thermal insulation shedding based on a deep learning model provided by an embodiment of the present invention; Figure 2 For the embodiment of the present invention to execute Figure 1 The structural schematic block diagram of a system for predicting the risk of external wall thermal insulation shedding based on a deep learning model for the method for predicting the risk of external wall thermal insulation shedding based on a deep learning model in Detailed Embodiments

[0013] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the protection scope of the present invention.

[0014] See Figure 1 As shown in Step S110, collect a multi-dimensional state data set of the external wall of the target building, and the multi-dimensional state data set includes thermal insulation layer material attribute data, environmental exposure history data, structural connection strength data, surface deformation monitoring data, and construction process record data.

[0015] In this embodiment, a high-rise commercial building located in the central area of the city can be considered. The building has been built for a certain number of years, and multi-dimensional status data of its exterior wall can be collected.

[0016] For the data of the insulation layer material properties, a material composition analyzer is used. For example, the exterior wall insulation layer of this building uses polystyrene foam board (EPS) material. Through the material composition analyzer, data such as the polymer degradation rate, the attenuation coefficient of the hydrophobic property, and the compressive strength change curve of the EPS insulation layer are obtained. In actual measurement, over time, it is found that the polymer degradation rate gradually increases from an extremely low initial level, indicating that the polymer molecular chains inside the insulation layer begin to break under the action of environmental factors. The attenuation coefficient of the hydrophobic property also shows a trend of increasing with time. The insulation layer that could originally effectively prevent water intrusion now allows more water to penetrate, affecting its insulation performance. The compressive strength change curve shows that within a few years after the building was completed, the compressive strength decreased slightly, which is an important manifestation of the aging of the insulation layer material.

[0017] Regarding the environmental exposure history data, data is obtained through a distributed temperature and humidity sensor array. The temperature gradient change sequence of the exterior wall of this building has obvious differences in different seasons. In summer, due to long-term direct sunlight on the sunny side of the exterior wall, the temperature can reach a very high value. The temperature difference fluctuation pattern between the sunny side and the shady side shows a large temperature difference during the day and a gradually shrinking temperature difference at night. For example, in the hot summer afternoon, the temperature on the sunny side may reach 50°C, while the shady side is only 30°C. The humidity penetration depth data shows that in the rainy season, due to the large rainfall and long duration, the humidity can penetrate to a certain depth of the insulation layer, which is related to the sealing performance of the exterior wall and the water absorption of the insulation layer material.

[0018] The structural connection strength data is measured by a stress wave detection device. For the residual value of the anchor bolt pull-out force, it is found that with the increase of the building's service life, due to the long-term influence of factors such as wind load and temperature change, the residual value of the pull-out force of some anchor bolts gradually decreases. For example, the residual value of the anchor bolt pull-out force on the top floor of the building has decreased by about 20% compared with the initial stage of construction, indicating that a certain degree of weakening has occurred in the structural connection part. The distribution of the interfacial bonding stress also shows unevenness. In some local areas, the bonding stress is significantly lower than other areas, which may be caused by the construction technology or the interaction between the insulation layer and the base wall.

[0019] The surface deformation monitoring data is collected by a laser scanner and an infrared thermal imager. The three-dimensional micro-displacement data collected by the laser scanner shows that there are small displacements in some parts of the building exterior wall, especially at the corners of the building and on the walls subjected to large wind loads. The crack propagation trend data indicates that some originally small cracks tend to gradually expand over time. Through the infrared thermal imager, it is found that there are hollow areas distributed between the insulation layer and the base wall. For example, in the middle area of the exterior wall of a certain floor, a hollow area with an area of about 1 square meter is found, which may be caused by uneven application of the adhesive during construction or unevenness of the surface of the base wall.

[0020] The construction process record data is extracted from the project archive. The index of the evenness of the adhesive application shows that there are significant differences in the thickness of the adhesive application in some wall areas, and the difference between the maximum thickness and the minimum thickness is about 0.5 mm. The data of the distribution of the density of the anchor bolts indicates that in some parts of the wall, the number of anchor bolts per square meter deviates from the preset standard density, and the number of anchor bolts is less in some areas. The deviation value data of the lap length of the fiberglass mesh shows that there are certain lap length deviations in both the vertical and horizontal directions of the wall. The maximum deviation in the vertical direction reaches 5 cm. These problems in the construction process may have a potential impact on the stability of the insulation layer.

[0021] Step S120, extract the material degradation characteristics of the insulation layer material attribute data to generate a material aging feature vector, and extract the spatio-temporal distribution characteristics of the environmental exposure history data to generate an environmental action feature tensor.

[0022] For the extraction of the material degradation characteristics of the insulation layer material attribute data, a material aging kinetics model is established. This material aging kinetics model includes a chemical bond fracture probability equation under the coupling action of temperature and humidity. Taking the EPS insulation layer as an example, the polymer degradation rate is non-linearly fitted with the time variable. Assuming that the time variable is t and the polymer degradation rate is P(t), a functional relationship P(t)=A*exp(-B / t) (where A and B are constants obtained by fitting) is obtained through non-linear fitting, and the aging acceleration factor is extracted as the first feature dimension. For example, when A = 0.1 and B = 10, as t increases, the aging acceleration factor will also change accordingly.

[0023] Perform a frequency domain transformation on the hydrophobic performance attenuation coefficient, and extract the low-frequency attenuation dominant frequency as the second feature dimension. Assuming that the data sequence of the hydrophobic performance attenuation coefficient changing with time is H(t), after Fourier transform, a dominant frequency f is found in the low-frequency band, and this frequency reflects the attenuation characteristics of the hydrophobic performance on a longer time scale.

[0024] Identify the inflection point position in the compressive strength change curve, and calculate the slope change rate before and after the inflection point as the third characteristic dimension. For example, the compressive strength change curve has an inflection point at t = 5 years. The slope before the inflection point is k1, and the slope after the inflection point is k2. Calculate (k2 - k1) / k1 to obtain the slope change rate, which reflects the trend characteristics of the compressive strength change. Normalize and splice these three characteristic dimensions to generate a material aging feature vector with temporal dependence.

[0025] For the extraction of spatio-temporal distribution characteristics of environmental exposure historical data, taking the temperature gradient change sequence as an example, in a three-dimensional space (x, y, t, where x and y are the plane coordinates of the exterior wall, and t is time), organize the temperature data at different seasons and different wall positions. In summer, the temperature on the sunny side is high, and the temperature on the shaded side is low, which is different in winter. Extract the spatio-temporal distribution characteristics of these temperature data to construct an environmental action feature tensor. For example, divide the temperature data on the sunny side in summer into certain grids in space (such as one grid per square meter). The temperature values at different time points in each grid form a part of the tensor, and then add the data on the shaded side and other seasons to form the environmental action feature tensor, which can comprehensively reflect the spatio-temporal distribution characteristics of environmental factors on the exterior wall surface.

[0026] Step S130, spatially align and fuse the structural connection strength data and the surface deformation monitoring data to generate a structural deformation correlation map, and convert the construction process record data into a process defect probability distribution.

[0027] When spatially aligning and fusing the structural connection strength data and the surface deformation monitoring data, construct a three-dimensional coordinate system mapping relationship. Taking a certain layer of the building exterior wall as an example, align the spatial distribution of the residual value of the anchor bolt pull-out force measured by the stress wave detection device with the micro-displacement measured by the laser scanner within the same coordinate grid. Assume that the exterior wall plane is divided into a 10×10 grid, and each grid corresponds to a measurement point of the residual value of the anchor bolt pull-out force and the micro-displacement. Calculate the Pearson correlation coefficient between the interfacial bond stress distribution and the crack propagation rate in each grid cell to generate a stress-deformation coupling coefficient matrix. For example, in a certain grid cell, the interfacial bond stress distribution is large, and the crack propagation rate is also fast. A high Pearson correlation coefficient is obtained through calculation, indicating a strong correlation between the two.

[0028] Perform morphological dilation on the distribution of hollow area regions to generate a mask for the radiation range affected by the hollow area, and superimpose it on the stress-deformation coupling coefficient matrix. Assume that the hollow area is a circular region with a radius of 0.5 meters. After morphological dilation, the radius expands to 1 meter, and this expanded region is the mask for the radiation range affected by the hollow area. Superimpose it on the stress-deformation coupling coefficient matrix so that the influence of the hollow area can be fully considered when considering the structural deformation correlation. Use graph convolution operations to propagate the effect of the mask for the radiation range affected by the hollow area in a three-dimensional coordinate system to generate a structural deformation correlation map with spatial topological correlation. The adjacency matrix of the graph convolution operation is determined by the material continuity parameter between grid cells. For example, if the material continuity between two adjacent grid cells is good, the corresponding element value in the adjacency matrix is larger; otherwise, it is smaller.

[0029] For converting construction process record data into the probability distribution of process defects, first perform a spatial dispersion analysis on the uniformity index of adhesive coating to generate an adhesive distribution dispersion index. Assume that on a certain wall, the measurement data of the adhesive coating thickness is a set of discrete values, and the adhesive distribution dispersion index is obtained by calculating the coefficient of variation of the coating thickness per unit area. For example, in a 1-square-meter wall area, the maximum coating thickness is 3 mm and the minimum is 1 mm, and the calculated coefficient of variation is 0.5, which reflects the uniformity of the adhesive coating.

[0030] Perform kernel density estimation on the distribution of anchor fastener density to generate an anchor fastener density distribution map. For example, set the standard anchor fastener density to 5 per square meter. In actual measurement, the number of anchor bolts per square meter in a certain wall area fluctuates between 3 and 7. The anchor fastener density distribution map generated by kernel density estimation can intuitively show the spatial deviation degree of the anchor fastener density from the standard density.

[0031] Perform a sliding window statistics on the deviation value of the mesh fabric lapping length to generate a mesh fabric lapping deviation trend vector. Assume that the sliding window length is 1 meter and the statistics are performed along the vertical and horizontal directions of the wall. In the vertical direction, it is found that the deviation value of the mesh fabric lapping length shows a gradually increasing trend in some areas, and the deviation value is relatively stable in the horizontal direction. The vector composed of these deviation values is the mesh fabric lapping deviation trend vector.

[0032] Align the adhesive distribution dispersion index, the anchor fastener density distribution map, and the mesh fabric lapping deviation trend vector in the spatial coordinate system to generate a three-dimensional process quality spatial distribution map. Then extract the residual value of the anchor bolt pulling force in the structural connection strength data, superimpose this residual value of the anchor bolt pulling force on the three-dimensional process quality spatial distribution map, and calculate the product of the process quality parameter and the structural strength in each spatial unit as the process strength coupling coefficient matrix.

[0033] Perform a morphological erosion operation on the distribution of the hollow area to generate a mask for the potential expansion area of the hollow. Assume that the preset buffer distance for the edge of the hollow area to extend outward is 0.3 meters. After the morphological erosion operation, a mask for the potential expansion area of the hollow is obtained. Multiply the mask for the potential expansion area of the hollow with the process strength coupling coefficient matrix pixel by pixel to generate a defect-sensitive area matrix. In the defect-sensitive area matrix, the high-value area represents the superposition area of process defects and structural weakening. According to the construction timestamp in the construction process record data, perform time decay weighting on the defect-sensitive area matrix. For example, the weight coefficient applied to the construction process defects 5 years ago from the current time is 0.8, and the weight coefficient applied to the construction process defects 10 years ago is 0.9. Because over time, the impact of early construction process defects on the current risk of insulation layer shedding is relatively greater, a time-varying defect-sensitive matrix is generated. Finally, perform max-min normalization on the time-varying defect-sensitive matrix to map the values of each unit in the matrix to the 0-1 interval, generating a process defect probability distribution.

[0034] Step S140, input the material aging feature vector, the environmental action feature tensor, the structure deformation correlation map, and the process defect probability distribution into a pre-trained multi-modal risk prediction network, and output the risk level of the external wall insulation layer shedding and the heat map of the risk area.

[0035] Specifically, the pre-trained multi-modal risk prediction network includes a feature cross-attention mechanism and a spatio-temporal gated recurrent unit. Input the previously obtained material aging feature vector, the environmental action feature tensor, the structure deformation correlation map, and the process defect probability distribution into this network.

[0036] In the feature cross-attention mechanism, use the material aging feature vector as the query vector, the environmental action feature tensor as the key vector, and the structure deformation correlation map as the value vector to calculate the cross-modal feature attention weight. For example, in the calculation process, the aging acceleration factor in the material aging feature vector is correlated with the temperature gradient change data in the environmental action feature tensor, and the obtained attention weight reflects the mutual influence relationship between the two. Perform time series modeling on the cross-modal feature attention weight through the spatio-temporal gated recurrent unit to capture the evolution law of risk factors under the action of environmental loads in different seasons. For example, in winter, due to the lower temperature, the stress change in the structural connection part is different from that in summer, and the impact of this seasonal change on the risk factor is modeled through the spatio-temporal gated recurrent unit.

[0037] Set a region-sensitive loss function at the output layer of the multi-modal risk prediction network, and multiply the pixel-level prediction error of the risk region heat map by the spatial weight of the process defect probability distribution. For example, for the prediction error within the radiation range affected by the hollowing, since the hollowing area has a greater impact on the risk of insulation layer shedding, the region-sensitive loss function applies a triple penalty coefficient to the prediction error within this area. Adopt an adversarial training strategy to optimize the multi-modal risk prediction network, and the discriminator network distinguishes the distribution difference between the heat map of real shedding cases and the heat map generated by the network. Through this process, the network outputs the risk level of the external wall insulation layer shedding, such as low risk, medium risk, high risk levels, as well as the risk region heat map. The risk region heat map can intuitively show the shedding risk degree of different regions of the external wall. For example, in the heat map, the red area represents a high-risk area, which may be the result of the combined action of factors such as low structural connection strength, serious process defects, and large environmental exposure in this area; the blue area represents a low-risk area, which may be because all aspects of factors are relatively good.

[0038] Step S150, generate a dynamic monitoring strategy according to the risk level and the risk region heat map, and the dynamic monitoring strategy includes an optimized sensor deployment plan and a detection cycle adjustment parameter.

[0039] Specifically, perform density clustering according to the risk values of each pixel in the risk region heat map to identify the risk core region and the edge diffusion region. For example, in the risk region heat map, the region with relatively high and concentrated risk values is identified as the risk core region, which may be a certain corner of a certain layer of the external wall, and the risk of this region is relatively high due to reasons such as weak structural connection and serious aging of the insulation layer; while the region with relatively low risk values but a spreading trend is identified as the edge diffusion region, such as the surrounding wall near the risk core region.

[0040] In the optimized sensor deployment plan, deploy high-precision fiber Bragg grating sensors in the risk core region. Such sensors can accurately measure minute deformation and stress changes. For example, deploy one fiber Bragg grating sensor within each 1-square-meter grid in the risk core region to monitor the status of this region in real time. Deploy low-power wireless vibration sensors in the edge diffusion region because the risk in the edge diffusion region is relatively low, and low-power sensors can not only meet the monitoring requirements but also reduce costs.

[0041] Adjust the detection cycle based on the time series prediction results of the risk level. Assume that the risk level is in a low-risk state for a period of time, then during the stable period, extend the detection cycle to twice the original cycle. For example, if the original detection cycle is once a week, now it is extended to once every two weeks. When it is predicted that the risk is about to enter the acceleration period, shorten the detection cycle to one-third of the original cycle, such as changing it to once a day, in order to timely detect the possible risk of insulation layer shedding.

[0042] Establish a game model for the power consumption-accuracy of monitoring devices and adopt a multi-objective optimization algorithm. Define the decision variables as the combination patterns of the sampling frequencies and data transmission intervals of various types of sensors. For example, for fiber Bragg grating sensors, the sampling frequency can be set to different modes such as once per minute, once every 5 minutes, etc., and the data transmission interval can be once per hour or once per day, etc. Establish the first objective function as the reciprocal of the total power consumption of the monitoring system, and the second objective function as the F1 value of the risk prediction accuracy. Use the non-dominated sorting genetic algorithm to find the Pareto optimal solution set in the decision variable space. Introduce constraint conditions to ensure that the sampling frequency in the risk core area is not less than once per minute, ensuring effective monitoring of high-risk areas. Select the optimal deployment plan that takes into account both power consumption and accuracy from the Pareto front through the fuzzy comprehensive evaluation method. For example, on the premise of meeting the risk monitoring requirements, select the combination pattern of the sensor sampling frequency and data transmission interval with lower power consumption and higher accuracy.

[0043] Based on the above steps, in the data acquisition layer of the embodiment of the present application, a multi-dimensional state data set of the outer wall of the target building is comprehensively collected, covering insulation layer material property data, environmental exposure history data, structural connection strength data, surface deformation monitoring data, and construction process record data, which can comprehensively describe the actual situation of the external wall insulation system from multiple key perspectives. Secondly, unique data feature processing is performed on different types of data. Material degradation features are extracted from the insulation layer material property data to generate material aging feature vectors, accurately capturing the aging trend of materials over time and environmental factors; spatio-temporal distribution features are extracted from the environmental exposure history data to generate environmental action feature tensors, comprehensively considering environmental factors in the time and space dimensions, comprehensively showing the dynamic impact of environmental factors on the external wall insulation system, effectively mining the potential laws behind the data, and providing more representative and valuable feature information for subsequent risk analysis.

[0044] Furthermore, in the data fusion link, the structural connection strength data and the surface deformation monitoring data are spatially aligned and fused to generate a structural deformation correlation map, and the construction process record data is converted into a process defect probability distribution. By deeply exploring the internal relationships between different data, a map and probability distribution that can accurately reflect the complex relationship between the structure and performance of the external wall insulation system are constructed, providing a more intuitive and comprehensive perspective for risk prediction and greatly improving the accuracy of risk assessment.

[0045] Then, input the material aging feature vectors, environmental action feature tensors, structural deformation correlation maps, and process defect probability distributions obtained after processing into a pre-trained multi-modal risk prediction network to output the risk level of the external wall thermal insulation layer shedding and the thermal map of the risk area. This multi-modal risk prediction network integrates data features of multiple different modalities, fully leveraging the advantages of deep learning models in processing complex data and non-linear relationships, and can more accurately analyze the comprehensive impact of various factors on the risk of external wall thermal insulation layer shedding, thereby achieving high-precision prediction of the risk of external wall thermal insulation layer shedding, accurately identifying the risk level and the distribution of the risk area, and providing clear guidance for subsequent maintenance and management.

[0046] Finally, generate a dynamic monitoring strategy based on the risk level and the thermal map of the risk area, including an optimized sensor deployment plan and detection cycle adjustment parameters. By dynamically adjusting the sensor deployment and detection cycle, while ensuring effective monitoring of the external wall thermal insulation condition, it is possible to reasonably allocate monitoring resources, avoid unnecessary waste of monitoring costs, achieve the optimal balance between monitoring efficiency and resource utilization, and provide intelligent and efficient safeguard measures for the long-term stable operation of the building external wall thermal insulation system.

[0047] In a possible implementation manner, step S110 includes: Step S111, obtaining the temperature gradient change sequence and humidity penetration depth data in the environmental exposure historical data through a distributed temperature and humidity sensor array, where the temperature gradient change sequence includes the temperature difference fluctuation patterns on the sunny and shady sides of the external wall in different seasons.

[0048] In this embodiment, in different seasons, the temperature gradient change sequence of the building external wall shows obvious characteristics. In spring, the overall temperature is relatively mild, but due to the existence of the diurnal temperature difference, the temperature difference fluctuation pattern on the sunny and shady sides of the external wall still exists. The temperature on the sunny side can reach about 20°C in the afternoon, while the shady side remains at around 15°C. This temperature difference fluctuation forms a certain temperature gradient on the surface of the building external wall. With the arrival of summer, the temperature rises, the sunlight irradiation intensity increases, the temperature on the sunny side rises significantly, and can climb to 50°C in the hot afternoon, and the temperature on the shady side will also reach about 30°C. This huge temperature difference fluctuation pattern lasts for a long time, and due to the high temperature weather, the humidity is also relatively high. The humidity penetration depth data shows that during periods with more rainfall, the humidity of the external wall can penetrate to a certain depth inside the thermal insulation layer, about 5 cm. In autumn, the temperature gradually drops, and the temperature difference fluctuation pattern changes again. The temperature difference between the sunny and shady sides is relatively smaller than that in summer. In winter, although the overall temperature is low, the temperature difference between the sunny and shady sides is still obvious under sunlight irradiation. This temperature difference fluctuation pattern will have a continuous impact on the structure of the building external wall and the thermal insulation layer.

[0049] Step S112: Use a laser scanner to collect the three-dimensional micro-displacement and crack propagation trend in the surface deformation monitoring data, and obtain the distribution of the hollow area between the thermal insulation layer and the base wall based on an infrared thermal imager.

[0050] During the use of the building, through regular scanning of the building exterior wall by a laser scanner, it is found that the three-dimensional micro-displacement gradually changes in some areas. For example, in the high-rise part of the building, due to the more obvious wind force, a micro-displacement of about 0.5 mm in the horizontal direction and a displacement of 0.3 mm in the vertical direction occur in a certain wall area within a period of time. At the same time, the crack propagation trend is also continuously monitored. At a corner of the building exterior wall, a small crack with an initial width of about 0.1 mm is initially found. Over time, through multiple laser scanning measurements, it is found that this crack is expanding at a rate of about 0.05 mm per year. Based on the infrared thermal imager to obtain the distribution of the hollow area between the thermal insulation layer and the base wall, in the middle area of the exterior wall of a certain floor of the building, the infrared thermal imager detects a hollow area with an area of about 1 square meter. This area shows obvious different temperature characteristics from the surrounding normal areas in the thermal imaging diagram, indicating that there is a separation phenomenon between the thermal insulation layer and the base wall in this area. Such a hollow area will affect the overall performance and structural stability of the thermal insulation layer.

[0051] Step S113: Extract the adhesive coating uniformity index, anchor bolt density distribution, and mesh cloth lap length deviation value in the construction process record data from the project archive.

[0052] In the detailed records of the project archive, the adhesive coating uniformity index shows differences in different wall areas. For example, in the upper area of a certain wall, the adhesive coating thickness is uneven. The measurement data shows that within an area of 1 square meter, the maximum coating thickness reaches 3 mm and the minimum is 1 mm. Such uneven coating may affect the bonding effect between the thermal insulation layer and the base wall. In terms of the anchor bolt density distribution, according to the preset standard density, there should be 5 anchor bolts per square meter. However, in actual construction, there are deviations in the anchor bolt density in some wall areas. In a certain wall area, after detailed statistics, the number of anchor bolts per square meter fluctuates between 3 and 7. Such uneven density distribution will affect the fixing effect of the thermal insulation layer. The mesh cloth lap length deviation value is also recorded in detail. In the vertical direction of the wall, the maximum deviation value of the mesh cloth lap length in a certain area reaches 5 cm. In the horizontal direction, the deviation is relatively small, but there is still a certain degree of deviation. Such deviation will affect the overall structural strength of the thermal insulation layer.

[0053] Step S114: Use a material composition analyzer to obtain the polymer degradation rate, hydrophobic performance attenuation coefficient, and compressive strength change curve in the thermal insulation layer material property data.

[0054] For the polystyrene foam board (EPS) insulation layer used on the exterior wall of the building, the test results of the material composition analyzer show the change of the material over time. The polymer degradation rate gradually increases with the increase of the building service life. In the first 5 years after completion, the polymer degradation rate rises from an extremely low initial level to about 5%, indicating that the polymer molecular chains inside the insulation layer begin to gradually break under the action of environmental factors. The hydrophobic property attenuation coefficient is also constantly changing. Initially, the EPS insulation layer has good hydrophobic properties and can effectively prevent moisture intrusion. However, as time goes by, the hydrophobic property attenuation coefficient gradually increases. For example, after 10 years of use, the hydrophobic property decreases by about 30% compared to the initial state, making the insulation layer more vulnerable to moisture. The compressive strength change curve shows that in the first few years after completion, the compressive strength decreases slightly. The initial compressive strength is a certain value, and after 5 years of use, the compressive strength decreases by about 10%. This change in compressive strength reflects the aging degree of the insulation layer material.

[0055] Step S115: Use a stress wave detection device to measure the residual value of the anchor bolt pulling force and the interfacial bonding stress distribution in the structure connection strength data.

[0056] During the building's use, the stress wave detection device detects the structural connection part of the building's exterior wall. The residual value of the anchor bolt pulling force changes over time. For the anchor bolts on the top floor of the building, due to the long-term influence of factors such as wind load and temperature change, the residual value of the pulling force decreases by about 20% compared to the initial stage after completion, indicating that the fixing effect of the anchor bolts on the insulation layer is gradually weakening. The interfacial bonding stress distribution also shows non-uniformity. In some local areas, such as the edge part of the building's exterior wall, the bonding stress is significantly lower than other areas. The measurement data shows that the interfacial bonding stress in the normal area is a certain value, while the bonding stress in the edge part is only about 60% of that in the normal area. This non-uniform bonding stress distribution will affect the connection stability between the insulation layer and the base wall.

[0057] In a possible implementation manner, step S120 includes: Step S121: Establish a material aging kinetics model, perform non-linear fitting on the polymer degradation rate and the time variable, and extract the aging acceleration factor as the first characteristic dimension.

[0058] In this embodiment, when analyzing the polystyrene foam board (EPS) thermal insulation layer of the building exterior wall, a non-linear fitting is performed between the polymer degradation rate and the time variable. As the service life of the building increases, the time variable continuously grows, and the polymer degradation rate also gradually rises. Assuming that the year is used as the time unit and recording starts from the completion, the polymer degradation rate rises relatively slowly in the first few years and then the rising speed gradually accelerates. By collecting the polymer degradation rate data at different time points, such as the degradation rate is 1% in the first year, 3% in the third year, 5% in the fifth year, etc., a functional relationship that conforms to these data points is obtained using the non-linear fitting method. This functional relationship can reflect the law of the polymer degradation rate changing with time, and the aging acceleration factor is extracted from this fitting result as the first characteristic dimension. This aging acceleration factor can be understood as a quantitative index for the accelerated polymer degradation rate under specific temperature-humidity conditions, reflecting the accelerating effect of environmental factors on the internal polymer structure of the thermal insulation layer material over time.

[0059] Step S122: Perform a frequency domain transformation on the hydrophobic property attenuation coefficient, and extract the low-frequency attenuation dominant frequency as the second characteristic dimension.

[0060] During the use of the building, over time, the hydrophobic property of the EPS thermal insulation layer continuously decays, and its hydrophobic property attenuation coefficient gradually increases. For example, the initial hydrophobic property can effectively prevent most moisture from invading the thermal insulation layer, but as time goes by, moisture begins to gradually seep in. The values of the hydrophobic property attenuation coefficient measured at different time points are formed into a time series, and then a frequency domain transformation is performed on this series. Through frequency domain analysis, it can be found that there is a low-frequency attenuation dominant frequency, which can reflect the attenuation characteristics of the hydrophobic property on a long time scale, and it is extracted as the second characteristic dimension. This low-frequency attenuation dominant frequency is a key index that can characterize the slow decline trend of the hydrophobic property over a long time, reflecting a macroscopic property change characteristic of the thermal insulation layer material when exposed to the environment for a long time.

[0061] Step S123: Identify the inflection point position in the compressive strength change curve, and calculate the slope change rate before and after the inflection point as the third characteristic dimension.

[0062] Regarding the compressive strength of the exterior wall insulation layer of this building, regular measurements and data recording have been carried out since its completion to form a curve of compressive strength change. In this curve of compressive strength change, it can be observed that as time goes by, the compressive strength may fluctuate slightly in the early stage and start to show a more obvious change trend after a certain point in time. For example, around the 5th year, an inflection point appears in the curve of compressive strength change. Before the inflection point, the rate of decrease in compressive strength is relatively slow and the slope is relatively small, while after the inflection point, the rate of decrease accelerates. Calculate the rate of change of the slope before and after the inflection point. This rate of change of the slope can quantify the difference in the change trend of compressive strength at different stages and reflect the characteristics of the aging process of the insulation layer material in terms of compressive performance.

[0063] Step S124, normalize and splice the first feature dimension, the second feature dimension, and the third feature dimension to generate a material aging feature vector with a time series dependence relationship.

[0064] When performing the normalization process, it is necessary to ensure that the values of each feature dimension are within a reasonable numerical range for subsequent analysis and processing. For example, map the value of the aging acceleration factor to the interval of 0 - 1, and perform corresponding normalization processing on the value of the low-frequency attenuation dominant frequency, and the same is true for the rate of change of the slope before and after the inflection point. Through such normalization and splicing, three different feature dimensions are combined into a vector. This vector can comprehensively reflect the aging state of the insulation layer material, and since each feature dimension is related to time, this vector reflects the time series dependence relationship of material aging. This kind of material aging feature vector can provide an important basis for further evaluating the overall performance of the insulation layer and predicting its peeling risk.

[0065] Among them, the material aging kinetics model includes a chemical bond breakage probability equation under the coupling action of temperature and humidity.

[0066] In a possible implementation manner, step S130 includes: Step S131, construct a three-dimensional coordinate system mapping relationship to align the spatial distribution of the residual value of the anchor bolt pull-out force with the micro-displacement measured by the laser scanner within the same coordinate grid.

[0067] Taking a certain floor exterior wall of this high-rise commercial building as an example, divide the exterior wall surface into several small grid units. For example, divide a wall surface into 10×10 grid units. Each grid unit has a corresponding coordinate position. In this three-dimensional coordinate system, on the one hand, measure the residual value of the anchor bolt pull-out force at each grid unit position through a stress wave detection device, and on the other hand, collect the micro-displacement at each grid unit position through a laser scanner. Thus, a spatial correspondence relationship is established, enabling these two different types of data to be correlated and analyzed within the same coordinate framework.

[0068] Step S132: Calculate the Pearson correlation coefficient between the interfacial bond stress distribution and the crack propagation rate in each grid cell, and generate a stress-deformation coupling coefficient matrix.

[0069] Within each grid cell, the interfacial bond stress distribution reflects the bonding strength between the insulation layer and the base wall, while the crack propagation rate reflects a dynamic change in surface deformation. For example, in a certain grid cell, if the interfacial bond stress distribution is relatively large, it indicates that the bonding at this position is relatively firm. However, if the crack propagation rate is also fast at the same time, it suggests that there may be other factors affecting the stability of the insulation layer. By calculating the Pearson correlation coefficient, the degree of association between the two can be quantified. Combining the correlation coefficients of all grid cells forms the stress-deformation coupling coefficient matrix. Each element in this stress-deformation coupling coefficient matrix reflects the coupling relationship between the interfacial bond stress and the crack propagation rate in the corresponding grid cell, and this coupling relationship is an important basis for evaluating structural stability.

[0070] Step S133: Perform morphological dilation on the distribution of the hollow area to generate a mask for the radiation range affected by the hollow, and superimpose it on the stress-deformation coupling coefficient matrix.

[0071] For the external wall hollow areas detected by the infrared thermal imager before, such as the hollow area with an area of about 1 square meter in the middle of a certain floor of the external wall, perform morphological dilation on it. Assume that the original hollow area is a circular area. By setting certain dilation parameters, the radius of this circular area is gradually enlarged to form a larger area, and this area is the mask for the radiation range affected by the hollow. The area covered by this mask represents the range that may be affected by the hollow. Superimposing it on the stress-deformation coupling coefficient matrix can fully consider the hollow area and the possible impacts on its surrounding areas when considering structural stability.

[0072] Step S134: Use graph convolution operations to propagate the effect of the mask for the radiation range affected by the hollow in the three-dimensional coordinate system to generate a structural deformation correlation map with spatial topological association.

[0073] Among them, the adjacency matrix of the graph convolution operation is determined by the material continuity parameters between grid cells.

[0074] In a three-dimensional coordinate system, the material continuity between each grid cell is different. For example, if the material continuity between adjacent grid cells is good, then their connection is closer in the graph convolution operation, and the corresponding element value in the adjacency matrix is larger. Through the graph convolution operation, the effect of the delamination influence radiation range mask will propagate in the three-dimensional coordinate system according to the topological relationship between grid cells. This propagation process can reflect the influence of the delamination area on the surrounding structure in the structure deformation correlation graph through a spatial topological association method. This structure deformation correlation graph can comprehensively display the complex spatial relationship between the structure connection strength, surface deformation, and delamination area, providing an important basis for further analyzing the structural stability of the insulation layer.

[0075] For example, the conversion of the construction process record data into a process defect probability distribution includes: Step S135, perform a spatial dispersion analysis on the adhesive coating uniformity index to generate an adhesive distribution dispersion index, where the adhesive distribution dispersion index characterizes the coefficient of variation of the coating thickness per unit area.

[0076] In this embodiment, during the construction of the building exterior wall, the coating uniformity of the adhesive has an important impact on the bonding effect between the insulation layer and the base wall. The adhesive coating uniformity index obtained from the engineering archive shows that there are differences in the adhesive coating thickness in different wall areas. Taking a certain wall as an example, the wall is divided into several small unit area regions, for example, 1 square meter as a unit. Measure the adhesive coating thickness within these units and find that the thickness values between different units are not the same. Calculate the coefficient of variation of the coating thickness per unit area to characterize the adhesive distribution dispersion index. Suppose in a certain 1-square-meter area, the maximum measured coating thickness is 3 mm and the minimum is 1 mm. According to the calculation formula of the coefficient of variation, calculate the adhesive distribution dispersion index for this area. This index can quantify the non-uniformity of the adhesive coating thickness per unit area. The larger the value, the more uneven the adhesive coating. This non-uniformity may cause the insulation layer to be not firmly bonded in some areas, thus affecting the overall stability of the insulation layer.

[0077] Step S136, perform a kernel density estimation on the anchor density distribution to generate an anchor density distribution map, and the anchor density distribution map reflects the spatial deviation degree of the number of anchor bolts per square meter from the preset standard density.

[0078] During construction, the number of anchor fasteners should be set according to the preset standard density per square meter. For the exterior wall of this building, the preset standard density is 5 anchor bolts per square meter. However, in actual construction, there are deviations in the density of anchor fasteners in different wall areas. By counting the number of anchor bolts in each wall area and then performing kernel density estimation. For example, in a certain wall area, the number of anchor bolts per square meter fluctuates between 3 and 7. Through the kernel density estimation method, the generated anchor fastener density distribution map can intuitively reflect the spatial deviation degree between the number of anchor bolts per square meter and the preset standard density. On this distribution map, it can be clearly seen which areas have fewer anchor bolts than the standard density and which areas have more. Areas with insufficient anchor bolts may not provide enough fixing force, while areas with excessive anchor bolts may be unreasonable layouts during the construction process, and these will all affect the structural stability of the insulation layer.

[0079] Step S137, perform a sliding window statistic on the deviation value of the lap length of the mesh cloth to generate a mesh cloth lap deviation trend vector, and the mesh cloth lap deviation trend vector includes the continuity defect gradients in the vertical and horizontal directions of the wall.

[0080] In the vertical and horizontal directions of the wall, the lap length of the mesh cloth plays an important role in the overall structural strength of the insulation layer. The deviation values of the lap length of the mesh cloth obtained from the engineering archives show different degrees of deviation in different areas. Using the sliding window statistic method, with a certain length (such as 1 meter) as the window, move along the vertical and horizontal directions of the wall, and count the deviation values of the lap length of the mesh cloth within each window. In the vertical direction, the deviation value of the lap length of the mesh cloth in a certain area may show a gradually increasing trend, while in the horizontal direction, the deviation value may be relatively stable or show different changing trends. Combining these deviation values in the vertical and horizontal directions forms a mesh cloth lap deviation trend vector. This vector can reflect the continuity defect gradients of the lap length deviation of the mesh cloth in different directions of the wall, and the existence of this gradient indicates that the structural strength of the insulation layer may vary in different areas because the lap length deviation of the mesh cloth will affect the integrity and crack resistance of the insulation layer.

[0081] Step S138, align the adhesive distribution dispersion index, the anchor fastener density distribution map, and the mesh cloth lap deviation trend vector in the spatial coordinate system to generate a three-dimensional process quality spatial distribution map.

[0082] In the three-dimensional space of the building exterior wall, each wall surface has a corresponding coordinate position. Align the previously obtained dispersion index of adhesive distribution, the density distribution map of anchor bolts, and the trend vector of the lap deviation of the mesh fabric according to the same spatial coordinate system. For example, the dispersion index of adhesive distribution is corresponding to the adhesive coating area of each wall surface unit, the density distribution map of anchor bolts is corresponding to the anchor bolt setting area of each wall surface unit, and the trend vector of the lap deviation of the mesh fabric is corresponding to the lap area of the mesh fabric of each wall surface unit. Through this alignment of the spatial coordinate system, these three indicators reflecting different aspects of the construction process are integrated into a three-dimensional space to generate a three-dimensional process quality space distribution map. This distribution map can comprehensively display the quality distribution of the construction process in the three-dimensional space of the building exterior wall, where each spatial unit contains process quality information such as adhesive coating, anchor bolt setting, and mesh fabric lapping.

[0083] Step S139, extract the residual value of the anchor bolt pulling force in the structural connection strength data, and perform spatial superposition of this residual value of the anchor bolt pulling force with the three-dimensional process quality space distribution map, and calculate the product of the process quality parameter and the structural strength in each spatial unit as the process strength coupling coefficient matrix.

[0084] The residual value of the anchor bolt pulling force measured by the stress wave detection device reflects the structural connection strength of the anchor bolt after being used for a period of time. For example, in the top layer part of the building exterior wall, due to the long-term influence of factors such as wind load, the residual value of the anchor bolt pulling force has decreased compared with that at the initial stage of completion. Performing spatial superposition of this residual value of the anchor bolt pulling force with the three-dimensional process quality space distribution map means multiplying the process quality parameter (including the process quality information reflected by the dispersion index of adhesive distribution, the density distribution map of anchor bolts, and the trend vector of the lap deviation of the mesh fabric) with the corresponding residual value of the anchor bolt pulling force in each spatial unit. Each element in the calculated process strength coupling coefficient matrix reflects the coupling relationship between the process quality and the structural strength in this spatial unit. If the process quality in a certain spatial unit is poor and the residual value of the anchor bolt pulling force is low, then the corresponding element value in this matrix will be small, indicating that the process defects and structural weakening phenomena in this area are relatively serious.

[0085] Step S1310, perform a morphological erosion operation on the distribution of the hollow area to generate a mask of the potential expansion area of the hollow, and the mask of the potential expansion area of the hollow covers a preset buffer distance extending outward from the edge of the hollow.

[0086] Previously, hollow areas were detected on the exterior wall of a building using an infrared thermal imager. For example, there was a hollow area with an area of approximately 1 square meter in the middle of the exterior wall on a certain floor. The preset buffer distance of the mask for the potential expansion area of the hollow was determined based on the humidity penetration depth data in the environmental exposure history data. Since the humidity penetration depth data showed that during periods of heavy rainfall, the humidity could penetrate to a certain depth inside the insulation layer, and this depth was directly proportional to the preset buffer distance of the potential expansion area of the hollow. At the same time, the convolution kernel size of the morphological erosion operation was negatively correlated with the polymer degradation rate in the insulation layer material property data. As the polymer degradation rate increased, the convolution kernel size decreased. Through the morphological erosion operation, a preset buffer distance was extended outward from the edge of the hollow area to form a mask for the potential expansion area of the hollow. The area covered by this mask represents the possible potential expansion range of the hollow. Considering the effects of material aging and environmental factors on the expansion of the hollow, this mask can provide more comprehensive information for subsequent analysis.

[0087] Step S1311: Multiply the mask of the potential expansion area of the hollow and the process strength coupling coefficient matrix pixel by pixel to generate a defect-sensitive area matrix. The high-value areas in the defect-sensitive area matrix represent the superposition area of process defects and structural weakening.

[0088] In the defect-sensitive area matrix, because the mask of the potential expansion area of the hollow and the process strength coupling coefficient matrix are multiplied pixel by pixel, the high-value areas represent the superposition area of process defects and structural weakening. For example, in a certain spatial unit, if the value of the mask of the potential expansion area of the hollow is high (indicating that this area is close to the potential expansion area of the hollow), and at the same time the value in the process strength coupling coefficient matrix is also high (indicating that there are serious process defects and structural weakening phenomena in this area), then the result of the multiplication will be a high value in the defect-sensitive area matrix. These high-value areas indicate that there is a high risk of insulation layer shedding in this area because the superposition of process defects and structural weakening will greatly reduce the stability of the insulation layer in these areas.

[0089] Step S1312: Perform time decay weighting on the defect-sensitive area matrix according to the construction timestamp in the construction process record data, where a higher weight coefficient is applied to the construction process defects that are further away from the current time to generate a time-varying defect-sensitive matrix.

[0090] During the building construction process, the impact of construction process operations carried out at different times on the current insulation layer state is different. The defects of construction processes that occurred longer ago have a relatively greater impact on the risk of current insulation layer shedding due to the long-term influence of environmental factors and material aging. Dynamically adjust the attenuation rate of time decay weighting according to the chemical bond fracture probability equation in the material aging kinetics model. For example, for construction process defects that occurred 10 years ago, a relatively high weight coefficient, such as 0.9, is calculated according to the chemical bond fracture probability equation, while for construction process defects that occurred 5 years ago, the calculated weight coefficient is 0.8. Apply these weight coefficients to each element in the defect-sensitive area matrix for weighted calculation to obtain a time-varying defect-sensitive matrix. This matrix can reflect the dynamic change of the impact of construction process defects on the risk of insulation layer shedding over time.

[0091] Step S1313, perform max-min normalization on the time-varying defect-sensitive matrix, map the numerical values of each unit in the time-varying defect-sensitive matrix to the 0-1 interval, and generate the process defect probability distribution.

[0092] Among them, the attenuation rate of the time decay weighting is dynamically adjusted according to the chemical bond fracture probability equation in the material aging kinetics model.

[0093] Among them, the convolution kernel size of the morphological corrosion operation is negatively correlated with the polymer degradation rate in the insulation layer material attribute data.

[0094] Among them, the value of the preset buffer distance is directly proportional to the humidity penetration depth data in the environmental exposure history data.

[0095] In the time-varying defect-sensitive matrix, the values of each element may vary greatly. Through max-min normalization, the minimum value in the matrix is mapped to 0, the maximum value is mapped to 1, and the values of other elements are linearly scaled to the 0-1 interval. For example, if the minimum value in the time-varying defect-sensitive matrix is 0.1 and the maximum value is 0.9, then the value of an element with a middle value of 0.5 after normalization is (0.5 - 0.1) / (0.9 - 0.1) = 0.5. The generated process defect probability distribution can intuitively represent the probability of process defects in each area of the building exterior wall. The higher the probability value, the higher the risk of insulation layer shedding caused by process defects in that area. This process defect probability distribution provides an important basis for evaluating the risk of insulation layer shedding.

[0096] In a possible implementation manner, the pre-trained multi-modal risk prediction network includes a feature cross-attention mechanism and a spatio-temporal gated recurrent unit. The optimization steps of the multi-modal risk prediction network include: Step S210, in the feature cross-attention mechanism, use the material aging feature vector as the query vector, the environmental action feature tensor as the key vector, and the structural deformation correlation map as the value vector to calculate the cross-modal feature attention weights.

[0097] In this embodiment, taking the insulation layer of the building exterior wall as an example, the material aging feature vector contains information such as the aging acceleration factor, the dominant frequency of low-frequency attenuation, and the change rate of the slope before and after the inflection point extracted from the insulation layer material property data, which reflects the degradation characteristics of the insulation layer material over time. The environmental action feature tensor contains information such as the temperature gradient change sequence between the sunny side and the shady side of the exterior wall in different seasons and the humidity penetration depth data, which reflects the influence of environmental factors on the exterior wall. The structural deformation correlation map covers structural-related information such as the spatial distribution of the residual value of the anchor bolt pull-out force, the micro-displacement measured by the laser scanner, the relationship between the interfacial bonding stress distribution and the crack propagation rate, and the influence of the hollow area.

[0098] When calculating the cross-modal feature attention weights, for example, the aging acceleration factor in the material aging feature vector will be associated with the temperature gradient change data in the environmental action feature tensor. If the temperature gradient changes greatly and the aging acceleration factor is also high during a certain period, then the correlation between the two will be reflected in the weights when calculating the cross-modal feature attention weights. This correlation calculation comprehensively considers the mutual influence between material aging and environmental factors, as well as their relationship with structural deformation, so as to obtain the attention weights that can reflect the mutual relationship between different modal features.

[0099] Step S220, perform time series modeling on the cross-modal feature attention weights through a spatio-temporal gated recurrent unit to capture the evolution law of risk factors under environmental loads in different seasons.

[0100] This high-rise commercial building faces different environmental loads in different seasons. In summer, the high temperature, large temperature difference fluctuations, and the humidity changes brought by possible heavy rain weather have completely different impacts on the building exterior wall compared with the low temperature and snow in winter. The spatio-temporal gated recurrent unit will perform time series modeling according to the environmental data in different seasons, combined with the cross-modal feature attention weights.

[0101] For example, in summer, due to the high temperature, the aging rate of the thermal insulation layer material may accelerate, and at the same time, greater stress changes may occur in the structural connection parts due to thermal expansion and contraction. In winter, the low temperature may cause the material to shrink, the residual value of the anchor bolt pull-out force may be further reduced, and the structural stability is challenged. By performing time series modeling on the environmental load data in these different seasons and the corresponding risk factors (such as material aging, structural deformation, etc.), the evolution process of the risk factors under the influence of different seasonal environmental factors over time can be accurately captured. For example, data from consecutive years show that every summer, there is an obvious increasing trend in the micro-displacement in the structural deformation correlation map of a specific wall area. This seasonal pattern can be effectively modeled and analyzed by the spatio-temporal gated recurrent unit.

[0102] Step S230, set a region-sensitive loss function in the output layer of the multi-modal risk prediction network, and multiply the pixel-level prediction error of the risk area heat map by the spatial weight of the process defect probability distribution.

[0103] The risk area heat map reflects the degree of risk of insulation layer detachment in different areas of the building exterior wall, and the process defect probability distribution reflects the defect probability caused by construction process problems in different areas. For each pixel in the risk area heat map, its prediction error is multiplied by the spatial weight of the corresponding area in the process defect probability distribution.

[0104] For example, in a certain area of the building exterior wall, the process defect probability distribution shows that there is a high process defect probability in this area due to reasons such as uneven adhesive coating and deviation of anchor bolt density. In the risk area heat map, if there is an error in the prediction of the detachment risk in this area, the region-sensitive loss function will adjust this prediction error according to the spatial weight in the process defect probability distribution. Especially for the prediction error within the radiation range affected by hollowing, since hollowing has a greater impact on the risk of insulation layer detachment, the region-sensitive loss function will apply a triple penalty coefficient. If there is a deviation between the predicted risk value and the actual risk value of a pixel within the radiation range affected by hollowing in the risk area heat map, this deviation will be amplified three times when calculating the loss function, thereby prompting the network to pay more attention to the prediction accuracy of the area affected by hollowing during the training process.

[0105] Step S240, optimize the multi-modal risk prediction network using an adversarial training strategy, where the discriminator network distinguishes the distribution difference between the heat map of real detachment cases and the heat map generated by the network.

[0106] Among them, the region-sensitive loss function applies a triple penalty coefficient to the prediction error within the radiation range affected by hollowing.

[0107] In the actual situation, heat map data of real detachment cases of this building and other similar buildings were collected as a reference. The heat map of the risk area generated by the network was compared with the heat maps of these real detachment cases. The discriminator network optimized the multi-modal risk prediction network by analyzing the distribution differences between the two.

[0108] For example, if the risk prediction value of the heat map generated by the network is too high in a certain area, while the heat map of the real detachment case shows that the risk in this area is low, the discriminator network will capture this difference and feedback it to the multi-modal risk prediction network for adjustment. On the contrary, if the risk prediction value of the network for a certain area is too low, the discriminator network will also identify this difference and prompt the network to be optimized, making the heat map of the risk area generated by the network closer to the real situation.

[0109] In a possible implementation manner, the feature cross-attention mechanism implements multi-layer feature interaction, including: Step S310, in the first attention layer, cross-correlate the channel dimensions of the material aging feature vector and the environmental action feature tensor to generate a material-environment coupling feature map.

[0110] Each feature dimension in the material aging feature vector is associated and calculated with different channel data in the environmental action feature tensor. For example, the hydrophobic performance attenuation coefficient in the material aging feature vector is cross-correlated with the relevant channels of the humidity penetration depth data in the environmental action feature tensor. If the hydrophobic performance attenuation coefficient is high, it means that the waterproof performance of the insulation layer has decreased, and if the humidity penetration depth data is large, it indicates that the environmental humidity has a greater impact on the insulation layer. In this case, the cross-correlation between the two in the first attention layer will generate a specific material-environment coupling feature map, which can reflect the coupling relationship between material aging and environmental humidity.

[0111] Step S320, in the second attention layer, align the positions of the material-environment coupling feature map with the spatial dimension of the structure deformation correlation map to generate three-dimensional spatial attention weights.

[0112] The material-environment coupling feature map reflects the relationship between material aging and environmental factors, and the structure deformation correlation map contains spatial information such as structure connection strength and surface deformation. After aligning their positions, for example, the eigenvalue corresponding to a certain wall area in the material-environment coupling feature map is associated and calculated with the structure information of the same wall area in the structure deformation correlation map to generate three-dimensional spatial attention weights. These weights can reflect the comprehensive relationship among material aging, environmental factors, and structure deformation at a specific spatial position.

[0113] Step S330: In the third attention layer, use the process defect probability distribution as a gating signal to probabilistically modulate the three-dimensional spatial attention weights.

[0114] Step S340: Through skip connections, perform residual fusion of the attention outputs of each layer with the original input features to form a multi-scale risk feature pyramid.

[0115] Among them, the probabilistic modulation uses the reciprocal of the defect probability value as the attention enhancement coefficient.

[0116] The probability value of each region in the process defect probability distribution reflects the defect possibility of that region due to construction process problems. Taking a certain wall area as an example, if the probability value of this region in the process defect probability distribution is high, it indicates a large construction process defect. Use the reciprocal of this probability value as the attention enhancement coefficient to modulate the three-dimensional spatial attention weights. For example, if the process defect probability is 0.8 and its reciprocal is 1.25, then the value of the three-dimensional spatial attention weight in this region will be increased or decreased accordingly according to this coefficient, so as to pay more attention to the risk assessment of the region with larger process defects.

[0117] Through skip connections, perform residual fusion of the attention outputs of each layer with the original input features to form a multi-scale risk feature pyramid. The attention outputs of each layer contain feature information at different levels, and the original input features contain information such as the initial material aging feature vector, environmental action feature tensor, structural deformation correlation map, and process defect probability distribution. Through skip connections, fuse these different levels of information to form a multi-scale risk feature pyramid. For example, in a certain wall area, the material-environment coupling feature map output from the first attention layer, the three-dimensional spatial attention weight output from the second attention layer, and the probabilistically modulated weight output from the third attention layer are residually fused with the eigenvalue of the original material aging feature vector, environmental action feature tensor, structural deformation correlation map, and process defect probability distribution in this region. This fusion method can retain feature information at different levels and construct a pyramid structure containing multi-scale risk features, so as to comprehensively evaluate the risk of the building exterior wall insulation layer falling off.

[0118] In a possible implementation manner, step S150 includes: Step S151: Perform density clustering based on the risk values of each pixel in the risk area heat map to identify the risk core area and the edge diffusion area.

[0119] In this embodiment, the heat map of the risk area reflects the risk degree of the insulation layer peeling off in different areas of the building exterior wall, and the risk value of each pixel represents the risk level of the corresponding area. By analyzing these pixels through the density clustering algorithm, during the clustering process, the radius parameter of the density clustering is dynamically adjusted according to the risk value gradient change.

[0120] For example, on a certain wall of the building exterior wall, pixels with higher risk values may be concentrated in a certain local area, such as near the corner or the area where structural problems have occurred. When the risk value shows a high gradient change in this local area, the radius parameter of the density clustering will be correspondingly reduced to more accurately cluster these high-risk pixels into the risk core area. While in the area where the risk value is relatively low and the change is relatively gentle, the radius parameter may be increased, so as to cluster some pixels with relatively low risk values but with a certain diffusion trend into the edge diffusion area. For example, in the middle area of a certain layer of the exterior wall, although the risk value is not very high, there is a trend of gradually spreading to the surrounding area. By appropriately increasing the radius parameter of the density clustering, this area is identified as the edge diffusion area. While in a corner of the top layer of the exterior wall, the risk value is extremely high and the surrounding risk value drops significantly. By using the density clustering with a smaller radius parameter, this corner is determined as the risk core area.

[0121] Step S152, in the optimized sensor deployment plan, high-precision fiber Bragg grating sensors are deployed in the risk core area, and low-power wireless vibration sensors are deployed in the edge diffusion area.

[0122] Due to the high risk value in the risk core area, more accurate monitoring data is required to timely detect the possible peeling off of the insulation layer. High-precision fiber Bragg grating sensors can accurately measure tiny deformation and stress changes. For example, one fiber Bragg grating sensor is deployed inside each 1-square-meter grid in the risk core area. These sensors can monitor in real time the changes of parameters such as the residual value of the anchor bolt pulling force, the interfacial bonding stress distribution, and the tiny displacement between the insulation layer and the base wall.

[0123] While the risk in the edge diffusion area is relatively low. The low-power wireless vibration sensors can not only meet the monitoring requirements but also reduce the cost. In the edge diffusion area, although the risk of the insulation layer peeling off is relatively low, monitoring is still required to prevent the further spread of the risk. The low-power wireless vibration sensors can indirectly reflect the state of the insulation layer by monitoring the vibration of the wall. For example, if the insulation layer starts to show slight loosening or structural changes in the edge diffusion area, it may cause changes in the vibration frequency or amplitude of the wall. These sensors can capture this change and transmit the data to the monitoring system.

[0124] Step S153: Based on the time series prediction results of the risk level, shorten the detection cycle to one-third of the original cycle during the risk acceleration period and extend it to twice the original cycle during the stable period.

[0125] Through the time series prediction of the risk level of the building exterior wall thermal insulation layer shedding, the situation of the risk in different stages can be analyzed. During the stable period, the risk level is relatively low and changes slowly. At this time, extending the detection cycle to twice the original cycle can reduce the consumption of unnecessary monitoring resources. For example, if the original detection cycle is once a week, it can be adjusted to once every two weeks during the stable period.

[0126] During the risk acceleration period, the risk level rises rapidly, and the possibility of the thermal insulation layer shedding increases sharply, requiring more frequent monitoring. At this time, shorten the detection cycle to one-third of the original cycle. For example, if the original cycle is once a day, it is adjusted to three times a day during the risk acceleration period. This can obtain more data in a timely manner to accurately grasp the state change of the thermal insulation layer and take corresponding measures in advance to prevent the thermal insulation layer from shedding.

[0127] Step S154: Establish a monitoring device power consumption - accuracy game model to balance the relationship between battery life and data collection frequency in the detection cycle adjustment parameters.

[0128] Among them, the radius parameter of the density clustering is dynamically adjusted with the change of the risk value gradient.

[0129] During the operation of the monitoring device, power consumption and accuracy are two mutually restrictive factors. Define the decision variable as the combination mode of the sampling frequency and data transmission interval of each type of sensor. For high-precision fiber Bragg grating sensors and low-power wireless vibration sensors, there are various combination methods for their sampling frequency and data transmission interval.

[0130] For example, the sampling frequency of the fiber Bragg grating sensor can be set to different modes such as once a minute and once every 5 minutes, and the data transmission interval can be once an hour or once a day, etc. Establish the first objective function as the reciprocal of the total power consumption of the monitoring system, that is, the lower the total power consumption, the higher the value of this objective function. The second objective function is the F1 value of the risk prediction accuracy. The higher the F1 value, the higher the accuracy of the risk prediction.

[0131] In a possible implementation manner, the monitoring device power consumption - accuracy game model adopts a multi-objective optimization algorithm, and the specific execution operations include: Step S410: Define the decision variable as the combination mode of the sampling frequency and data transmission interval of each type of sensor.

[0132] Step S420: Establish the first objective function as the reciprocal of the total power consumption of the monitoring system and the second objective function as the F1 value of the risk prediction accuracy.

[0133] Step S430, use the non-dominated sorting genetic algorithm to find the Pareto optimal solution set in the decision variable space.

[0134] Step S440, introduce constraint conditions so that the sampling frequency in the risk core area is not less than once per minute.

[0135] Step S450, select the optimal deployment plan that takes into account both power consumption and accuracy from the Pareto front by the fuzzy comprehensive evaluation method.

[0136] In this process, among the numerous combination modes of sensor sampling frequencies and data transmission intervals, those combinations that can both meet the low-power requirement and ensure high risk prediction accuracy can be found. For example, in some combination modes, although the sampling frequency is very high and more accurate data can be obtained, the power consumption is also very large, resulting in too short battery life; while in other combination modes, although the power consumption is low, the sampling frequency is too low to accurately predict risks. The non-dominated sorting genetic algorithm will screen out those combinations that achieve a better balance between power consumption and accuracy, forming the Pareto optimal solution set.

[0137] Introduce constraint conditions so that the sampling frequency in the risk core area is not less than once per minute. Due to the importance of the risk core area, in order to ensure that the insulation layer status of this area can be monitored in a timely and accurate manner, a lower limit is set for its sampling frequency. For example, for the fiber Bragg grating sensors deployed in the risk core area, even considering power consumption, at least one sampling must be guaranteed per minute so as to obtain sufficient data to analyze key information such as the structural changes and stress distribution in this area.

[0138] Select the optimal deployment plan that takes into account both power consumption and accuracy from the Pareto front by the fuzzy comprehensive evaluation method. The Pareto front contains multiple combination plans that achieve a better balance between power consumption and accuracy, but one of the most suitable plans needs to be selected from them. The fuzzy comprehensive evaluation method will comprehensively consider multiple factors, such as the performance of different sensors, the actual situation of the building exterior wall, cost, etc. For example, in a certain plan on the Pareto front, although the power consumption is low and the accuracy can meet the requirements, since the sensors used in this plan are unstable in high-temperature environments (such as the high temperature on the sunny side of the building exterior wall in summer), it may be excluded by the fuzzy comprehensive evaluation method. The finally selected plan will, under the comprehensive consideration of various factors, be able to balance the power consumption and accuracy of the monitoring equipment to the greatest extent on the premise of meeting the risk monitoring requirements, ensuring effective and economical monitoring of the risk of the building exterior wall insulation layer falling off.

[0139] In a possible implementation manner, the method further includes: Step S510, the edge computing nodes deployed on the target building exterior wall receive newly generated multi-dimensional state data subsets in real time.

[0140] In this embodiment, the edge computing node is located near the building exterior wall and can timely obtain newly generated data related to the state of the building exterior wall. These newly generated data continuously reflect the current state of the thermal insulation layer of the building exterior wall, covering some of the material property data of the thermal insulation layer, environmental exposure history data, structural connection strength data, surface deformation monitoring data, and construction process record data, jointly constituting a multi-dimensional state data subset. For example, over time, new temperature gradient change sequences, new surface micro-displacement measurement values, recent humidity penetration depth data, and possibly newly discovered adhesive coating conditions, etc. will be received by the edge computing node.

[0141] Step S520: When it is detected that the Mahalanobis distance of the material aging feature vector corresponding to the multi-dimensional state data subset exceeds the training data distribution threshold, trigger the local fine-tuning mode.

[0142] The Mahalanobis distance is an effective indicator for measuring the distribution difference between new data and training data. For the thermal insulation layer of this building exterior wall, different material types (such as expanded polystyrene (EPS) thermal insulation layer) have different characteristics. Therefore, the Mahalanobis distance threshold sets different critical values according to different material types. Taking the EPS thermal insulation layer as an example, during the long-term monitoring process, as new data is continuously generated, calculate the Mahalanobis distance between the material aging feature vector corresponding to the new multi-dimensional state data subset and the material aging feature vector in the previous training data. If this Mahalanobis distance exceeds the threshold set for the EPS thermal insulation layer, it indicates that there is a large difference between the new data and the training data in terms of material aging characteristics. At this time, it is necessary to trigger the local fine-tuning mode. This difference may be caused by the building exterior wall under new environmental conditions (such as the influence of long-term extreme climate, new changes in the building's surrounding environment, etc.) or the emergence of new aging patterns in the thermal insulation layer material over time.

[0143] Step S530: In the local fine-tuning mode, freeze the underlying feature extraction layer of the multi-modal risk prediction network and only update the weight parameters of the risk prediction layer.

[0144] The multi-modal risk prediction network consists of multiple layers. The underlying feature extraction layer is mainly responsible for extracting various features from the original data. For example, it extracts features such as the polymer degradation rate and the coefficient of hydrophobic performance attenuation from the insulation layer material property data, and extracts features such as the temperature gradient change sequence and the humidity penetration depth data from the environmental exposure history data. These feature extraction processes are relatively stable and have been well optimized in a large amount of data training in the early stage. The risk prediction layer is directly related to predicting the risk level of insulation layer shedding and the thermal map of the risk area. In the local fine-tuning mode, since the feature extraction mode of the underlying feature extraction layer is effective in most cases, this layer is frozen, and only the weight parameters of the risk prediction layer are updated. For example, the risk prediction layer may adjust the weight relationship of each factor in predicting the risk of insulation layer shedding according to the changes in data such as the new material aging feature vector, the environmental action feature tensor, the structural deformation correlation map, and the process defect probability distribution.

[0145] Step S540, adopt an incremental learning strategy to preserve the feature distribution difference between the historical risk prediction results and the newly generated multi-dimensional state data subset, and prevent the multi-modal risk prediction network from experiencing catastrophic forgetting.

[0146] Among them, the Mahalanobis distance threshold sets different critical values according to different material types.

[0147] Over time and with the continuous influx of new data, if no appropriate measures are taken, the multi-modal risk prediction network may gradually forget the knowledge learned before, which is the phenomenon of catastrophic forgetting. The incremental learning strategy aims to solve this problem by preserving the feature distribution difference between the historical risk prediction results and the newly generated data subset to maintain the performance of the network. For the risk prediction of the building exterior wall insulation layer, the historical risk prediction results include the evaluations of the risk of insulation layer shedding in each area at different time points before, and these results reflect the past state of the building exterior wall and the risk evolution process. The newly generated multi-dimensional state data subset reflects the current state of the exterior wall. By analyzing the feature distribution difference between the two, such as the accelerated material aging speed in the new data and the change in the influence mode of environmental factors on the structural connection strength, the incremental learning strategy can enable the multi-modal risk prediction network to learn new data without losing the memory of historical data, thus maintaining the accuracy of the risk prediction of insulation layer shedding.

[0148] Among them, the incremental learning strategy implements a knowledge distillation framework, and the specific implementation steps include: Step S610, use the original multi-modal risk prediction network as the teacher network and the fine-tuned multi-modal risk prediction network as the student network.

[0149] The original multi-modal risk prediction network has been trained with a large amount of data and shown certain performance in previous risk predictions. When entering the local fine-tuning mode, the fine-tuned multi-modal risk prediction network (i.e., the student network) is a network that has been partially adjusted for new data based on the original network. There is a knowledge transfer relationship between the teacher network and the student network. Through this relationship, the student network can learn useful knowledge from the teacher network while adapting to the new data situation.

[0150] Step S620: Add a feature distribution alignment loss function between the teacher network and the student network to minimize the KL divergence of the hidden layer features.

[0151] The hidden layer features contain the internal feature representations of the network during data processing. The KL divergence (Kullback-Leibler Divergence) is a metric for measuring the difference between two probability distributions. In this scenario, by calculating the KL divergence between the hidden layer features of the teacher network and the student network and adding a feature distribution alignment loss function, the hidden layer feature distribution of the student network is made as close as possible to that of the teacher network. For example, during the process of predicting the risk of insulation layer shedding on the exterior wall of a building, when the teacher network and the student network process the same input data (such as a new subset of multi-dimensional status data), their respective hidden layers will generate different feature representations. By minimizing the KL divergence, the student network can move closer to the teacher network in terms of feature representation, thereby inheriting some knowledge from the teacher network while being able to adjust according to the new data.

[0152] Step S630: Apply temperature scaling processing to the prediction output of the student network to make its softened probability distribution consistent with that of the teacher network.

[0153] When predicting the risk of insulation layer shedding, both the student network and the teacher network will output a probability distribution regarding the risk level and the heat map of the risk area. The temperature scaling processing adjusts the prediction output probability distribution of the student network by introducing a temperature coefficient. As the number of model updates increases, the temperature coefficient shows a logarithmic decay trend. For example, in the early stage of model updates, the temperature coefficient is relatively large, which makes the probability distribution of the student network more softened, that is, the probability distribution is smoother and closer to that of the teacher network. As the number of model updates increases, the temperature coefficient gradually decreases, and the prediction output probability distribution of the student network gradually stabilizes while always maintaining consistency with the teacher network. This temperature scaling processing helps the student network better learn from the teacher network and maintain knowledge coherence with the original network while adapting to new data.

[0154] Step S640: Retain the central vector of the historical data features and add a feature space exclusion constraint term during the training with new data. Among them, the temperature coefficient of the temperature scaling process shows a logarithmic decay trend with the number of model updates.

[0155] The central vector of the historical data features represents a central tendency of the historical data in the feature space. During the training with new data, adding a feature space exclusion constraint term can prevent the feature distribution of the new data from deviating too much from that of the historical data. For example, for the multi-dimensional status data of the building exterior wall insulation layer, data such as the insulation layer material properties, environmental factors, and structural features in the historical data form a specific feature space distribution, and its central vector reflects the core features of this distribution. When new data enters the training, by adding a feature space exclusion constraint term, it is ensured that the new data will not destroy the original feature space structure during the process of integrating into the network learning, thus avoiding the occurrence of catastrophic forgetting and ensuring that the multi-modal risk prediction network always maintains accurate performance in the long-term risk prediction task.

[0156] Based on the above description, in another embodiment, the embodiment of the present invention further provides an exterior wall insulation peeling risk prediction system based on a deep learning model. Refer to Figure 2 , Figure 2 FIG. is a structural diagram of the exterior wall insulation peeling risk prediction system 100 based on a deep learning model provided by the embodiment of the present invention. The exterior wall insulation peeling risk prediction system 100 based on a deep learning model may vary greatly due to configuration or performance, and may include one or more central processing units (CPUs) 112 (for example, one or more processors) and a memory 111. Among them, the memory 111 can be short-term storage or persistent storage. The program stored in the memory 111 may include one or more modules, and each module may include a series of instruction operations for the exterior wall insulation peeling risk prediction system 100 based on a deep learning model. Further, the central processing unit 112 may be configured to communicate with the memory 111 and execute a series of instruction operations in the memory 111 on the exterior wall insulation peeling risk prediction system 100.

[0157] The exterior wall insulation peeling risk prediction system 100 based on a deep learning model may further include one or more power supplies, one or more communication units 113, one or more transfer to output interfaces, and / or one or more operating systems, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.

[0158] The steps performed by the external wall thermal insulation peeling risk prediction system based on the deep learning model in the above embodiments can be combined with Figure 2 the structure of the external wall thermal insulation peeling risk prediction system based on the deep learning model shown.

[0159] In addition, an embodiment of the present invention further provides a storage medium for storing a computer program for executing the method provided in the above embodiments.

[0160] An embodiment of the present invention further provides a computer program product including instructions, which when running on a computer, causes the computer to execute the method provided in the above embodiments.

[0161] The above is only a specific implementation step of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for predicting the risk of external wall thermal insulation peeling based on a deep learning model, characterized in that, Including: Collecting a multi-dimensional status data set of the exterior wall of the target building, where the multi-dimensional status data set includes thermal insulation layer material property data, environmental exposure history data, structural connection strength data, surface deformation monitoring data, and construction process record data; Extracting material degradation characteristics from the thermal insulation layer material property data to generate a material aging feature vector, and extracting spatio-temporal distribution characteristics from the environmental exposure history data to generate an environmental action feature tensor; Performing spatial alignment and fusion on the structural connection strength data and the surface deformation monitoring data to generate a structural deformation correlation map, and converting the construction process record data into a process defect probability distribution; Inputting the material aging feature vector, environmental action feature tensor, structural deformation correlation map, and process defect probability distribution into a pre-trained multi-modal risk prediction network to output the risk level of the exterior wall thermal insulation layer shedding and a heat map of the risk area; Generating a dynamic monitoring strategy according to the risk level and the heat map of the risk area, where the dynamic monitoring strategy includes an optimized sensor deployment plan and detection period adjustment parameters.

2. The method for predicting the risk of external wall thermal insulation peeling based on a deep learning model according to claim 1, wherein The collecting the multi-dimensional status data set of the exterior wall of the target building includes: Obtaining the temperature gradient change sequence and humidity penetration depth data in the environmental exposure history data through a distributed temperature and humidity sensor array, where the temperature gradient change sequence includes the temperature difference fluctuation patterns on the sunny side and the shady side of the exterior wall in different seasons; Collecting the three-dimensional micro-displacement and crack propagation trend in the surface deformation monitoring data by using a laser scanner, and obtaining the distribution of the hollow area between the thermal insulation layer and the base wall based on an infrared thermal imager; Extracting the adhesive coating uniformity index, anchor bolt density distribution, and grid cloth lap length deviation value in the construction process record data from the project archive; Obtaining the polymer degradation rate, hydrophobic performance attenuation coefficient, and compressive strength change curve in the thermal insulation layer material property data through a material composition analyzer; Measuring the residual value of the anchor bolt pulling force and the interfacial bonding stress distribution in the structural connection strength data by using a stress wave detection device.

3. The method for predicting the risk of external wall thermal insulation peeling based on a deep learning model according to claim 2, wherein, The extracting material degradation characteristics from the thermal insulation layer material property data to generate a material aging feature vector includes: Establishing a material aging kinetics model, performing non-linear fitting on the polymer degradation rate and the time variable, and extracting the aging acceleration factor as the first feature dimension; Performing frequency domain transformation on the hydrophobic performance attenuation coefficient and extracting the low-frequency attenuation dominant frequency as the second feature dimension; Identifying the inflection point position in the compressive strength change curve and calculating the slope change rate before and after the inflection point as the third feature dimension; Normalizing and splicing the first feature dimension, the second feature dimension, and the third feature dimension to generate a material aging feature vector with a time series dependence relationship; Wherein, the material aging kinetics model includes a chemical bond fracture probability equation under the coupling action of temperature and humidity.

4. The method for predicting the risk of external wall thermal insulation shedding based on a deep learning model according to claim 2, wherein, The performing spatial alignment and fusion on the structural connection strength data and the surface deformation monitoring data to generate a structural deformation correlation map includes: Construct a three-dimensional coordinate system mapping relationship to align the spatial distribution of the residual value of the anchor bolt pulling force with the micro-displacement measured by the laser scanner within the same coordinate grid; Calculate the Pearson correlation coefficient between the interfacial bonding stress distribution and the crack propagation rate in each grid cell to generate a stress-deformation coupling coefficient matrix; Perform morphological dilation on the distribution of the delaminated area to generate a mask for the radiation range affected by the delamination, and superimpose it on the stress-deformation coupling coefficient matrix; Use graph convolution operations to propagate the effect of the mask for the radiation range affected by the delamination in the three-dimensional coordinate system to generate a structural deformation correlation map with spatial topological associations; Among them, the adjacency matrix of the graph convolution operation is determined by the material continuity parameters between grid cells.

5. The method for predicting the risk of external wall thermal insulation peeling based on a deep learning model according to claim 1, wherein The pre-trained multi-modal risk prediction network includes a feature cross-attention mechanism and a spatio-temporal gated recurrent unit. The optimization steps of the multi-modal risk prediction network include: In the feature cross-attention mechanism, use the material aging feature vector as the query vector, the environmental action feature tensor as the key vector, and the structural deformation correlation map as the value vector to calculate the cross-modal feature attention weight; Perform time series modeling on the cross-modal feature attention weight through the spatio-temporal gated recurrent unit to capture the evolution law of risk factors under environmental loads in different seasons; Set a region-sensitive loss function in the output layer of the multi-modal risk prediction network to multiply the pixel-level prediction error of the risk area heat map by the spatial weight of the process defect probability distribution; Use an adversarial training strategy to optimize the multi-modal risk prediction network. Among them, the discriminator network distinguishes the difference in the distribution of the heat map of real detachment cases and the heat map generated by the network; Among them, the region-sensitive loss function applies a triple penalty coefficient to the prediction error within the radiation range affected by the delamination.

6. The method for predicting the risk of external wall thermal insulation peeling based on a deep learning model according to claim 5, wherein The feature cross-attention mechanism implements multi-layer feature interactions, including: In the first attention layer, cross-correlate the channel dimensions of the material aging feature vector and the environmental action feature tensor to generate a material-environment coupling feature map; In the second attention layer, align the position of the material-environment coupling feature map with the spatial dimension of the structural deformation correlation map to generate a three-dimensional spatial attention weight; In the third attention layer, use the process defect probability distribution as a gating signal to probabilistically modulate the three-dimensional spatial attention weight; Perform residual fusion of the attention outputs of each layer with the original input features through skip connections to form a multi-scale risk feature pyramid; Among them, the probabilistic modulation uses the reciprocal of the defect probability value as the attention enhancement coefficient.

7. The method for predicting the risk of external wall thermal insulation peeling based on a deep learning model according to claim 1, characterized in that, The generation of the dynamic monitoring strategy according to the risk level and the risk area heat map includes: Perform density clustering on the risk values of each pixel in the risk area heat map to identify the risk core area and the edge diffusion area; In the optimized sensor deployment plan, deploy high-precision fiber Bragg grating sensors in the risk core area and low-power wireless vibration sensors in the edge diffusion area; Based on the time series prediction results of the risk level, shorten the detection period to one-third of the original period during the risk acceleration period and extend it to twice the original period during the stable period; Establish a monitoring device power consumption-accuracy game model to balance the relationship between battery life and data acquisition frequency among the parameters adjusted in the detection period; Among them, the radius parameter of the density clustering is dynamically adjusted with the gradient change of the risk value.

8. The method for predicting the risk of external wall thermal insulation peeling based on a deep learning model according to claim 7, wherein, The monitoring device power consumption-accuracy game model adopts a multi-objective optimization algorithm, and the specific implementation operations include: Define the decision variable as the combination mode of the sampling frequency and data transmission interval of each type of sensor; Establish the first objective function as the reciprocal of the total power consumption of the monitoring system, and the second objective function as the F1 value of the risk prediction accuracy; Use the non-dominated sorting genetic algorithm to find the Pareto optimal solution set in the decision variable space; Introduce constraint conditions so that the sampling frequency in the risk core area is not less than once per minute; Select the optimal deployment plan that takes into account both power consumption and accuracy from the Pareto front through the fuzzy comprehensive evaluation method.

9. The method for predicting the risk of external wall thermal insulation peeling based on a deep learning model according to claim 1, wherein The method further includes: The edge computing node deployed on the outer wall of the target building receives the newly generated multi-dimensional state data subset in real time; When it is detected that the Mahalanobis distance of the material aging feature vector corresponding to the multi-dimensional state data subset exceeds the training data distribution threshold, trigger the local fine-tuning mode; In the local fine-tuning mode, freeze the underlying feature extraction layer of the multi-modal risk prediction network and only update the weight parameters of the risk prediction layer; Adopt an incremental learning strategy to save the feature distribution difference between the historical risk prediction results and the newly generated multi-dimensional state data subset to prevent the multi-modal risk prediction network from experiencing catastrophic forgetting; Among them, the Mahalanobis distance threshold sets different critical values according to different material types; Among them, the incremental learning strategy implements a knowledge distillation framework, and the specific implementation steps include: Use the original multi-modal risk prediction network as the teacher network and the fine-tuned multi-modal risk prediction network as the student network; Add a feature distribution alignment loss function between the teacher network and the student network to minimize the KL divergence of the hidden layer features; Apply temperature scaling processing to the prediction output of the student network to make its softened probability distribution consistent with the teacher network; Retain the central vector of the historical data features and add a feature space exclusion constraint term during the training of new data; among them, the temperature coefficient of the temperature scaling processing shows a logarithmic decay trend with the number of model updates.

10. An external wall thermal insulation peeling risk prediction system based on a deep learning model, characterized in that, Include: A processor; A memory, in which a computer program is stored, and when the computer program is executed, it implements the method for predicting the risk of external wall thermal insulation shedding based on a deep learning model described in any one of claims 1-9.

Images