Cloud platform-based building structure risk assessment method and system, and storage medium

Abstract

This invention discloses a cloud-based method, system, and storage medium for building structure risk assessment, relating to the field of building safety technology. The method consists of several steps, including: Step 1: Real-time acquisition of multi-source monitoring data, including structural strain, temperature and humidity, traffic flow, and visual data of building cracks; real-time calculation of physical guidance residuals and damage mutation index using a lightweight algorithm; initial data and perceived data selected from the multi-source monitoring data based on the physical guidance residuals and damage mutation index, and uploaded to the cloud; Step 2: Filtering linear environmental interference in the perceived data through a simplified mechanical model to obtain nonlinear environmental fluctuations; fitting the nonlinear environmental fluctuations using a lightweight Transformer model to obtain pure residuals; Step 3: Constructing a healthy baseline evolution pool using the initial data as a baseline; when the healthy baseline evolution pool detects a continuous change in the environmental pattern of the building...

Description

Technical Field

[0001] This invention relates to the field of building safety technology, specifically to a cloud-based method, system, and storage medium for assessing building structural risks. Background Technology

[0002] During the long-term service of large infrastructure such as bridges and high-rise buildings, structural damage will gradually accumulate under the combined effects of material aging, load fatigue and extreme environment. If accurate early damage identification cannot be achieved, it may lead to serious accidents such as collapse. Therefore, structural health monitoring has become a core technology direction in the field of engineering safety.

[0003] Existing manual inspection and periodic offline testing methods suffer from drawbacks such as low efficiency, limited coverage, and inability to provide real-time early warnings. Meanwhile, the structural responses, such as strain and vibration frequency, are affected by environmental factors such as temperature, humidity, and traffic loads, as well as structural damage. Current technologies generally use fixed health baselines as the judgment criteria, but in actual engineering, the health status of structures evolves slowly over time, and the surrounding environmental patterns also change over a long period of time, such as the annual increase in traffic flow. Static baselines cannot adapt to this dynamic evolution, either causing frequent false alarms due to environmental drift or causing missed damage due to baseline aging. In the later stages, frequent manual calibration is required, resulting in extremely high operation and maintenance costs. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: Cloud-based building structure risk assessment methods include: Step 1: Collect multi-source monitoring data in real time, including structural strain, temperature and humidity, traffic flow and visual data of building cracks. Calculate physical guidance residuals and damage mutation index in real time using a lightweight algorithm. Based on the physical guidance residuals and damage mutation index, filter out the initial data and sensing data from the multi-source monitoring data and upload them to the cloud. Step 2: Filter linear environmental interference in the sensing data by simplifying the mechanical model to obtain nonlinear environmental fluctuations. Fit the nonlinear environmental fluctuations using a lightweight Transformer model to obtain the pure residuals. Step 3: Construct a health baseline evolution pool using the initial data as a baseline. When the health baseline evolution pool detects a continuous change in the environmental pattern of the building, it automatically iterates the baseline to incorporate the new health status. Based on the pure residual and the health baseline evolution pool, it determines whether the anomaly is a real damage and outputs a structured damage diagnosis result. Among them, the continuous change in the environmental pattern is determined by comparing multi-source monitoring data with a preset threshold. Step 4: Transform the damage diagnosis results into structural risk level, damage location and evolution trend, and generate a risk assessment report by combining historical data.

[0005] Furthermore, the process of calculating the physical guidance residual and damage mutation index in real time using a lightweight algorithm is as follows: By preloading simplified physical model parameters and lightweight operation logic in the edge computing gateway where the lightweight algorithm runs, and based on the simplified physical model parameters, combined with real-time collected temperature, humidity and traffic flow data, the theoretical response value of the structure under the current environment is calculated. The difference between the measured structural strain and the theoretical response value is used to obtain the physical guidance residual. A sliding time window of 1 hour is obtained in the edge computing gateway through lightweight computing logic. The sliding standard deviation is obtained based on the physical guidance residual. The damage mutation index is obtained by dividing the absolute difference between the physical guidance residual at the current moment and the physical guidance residual at the previous moment by the sliding standard deviation at the previous moment.

[0006] Furthermore, the process of filtering out the initial data and perceived data is as follows: Based on the results of physical guidance residuals and damage mutation index, multi-source monitoring data are classified and screened. The processed multi-source monitoring data is used as the initial data, and the preprocessed results of physical guidance residuals and damage mutation index are used as the sensing data.

[0007] Furthermore, the process of obtaining nonlinear environmental fluctuations is as follows: From the environmental monitoring data associated with the sensing data, environmental factors that are linearly related to the structural strain are extracted. The environmental factors are input into a simplified mechanical model to calculate the linear theoretical response value of the structural strain under the current environment. The measured structural strain response value in the sensing data is subtracted from the linear theoretical response value calculated by the model to complete the filtering of linear environmental interference. Based on the filtered linear environmental disturbances, a healthy sliding window with a length of 24 hours is obtained at the edge to construct the baseline distribution of nonlinear environmental fluctuations. The deviation between the measured structural strain response value and the mean of the healthy baseline is calculated based on the baseline distribution of nonlinear environmental fluctuations, which is the nonlinear environmental fluctuation.

[0008] Furthermore, the process of obtaining the pure residual is as follows: By pre-training and deploying a lightweight Transformer model in the cloud, the nonlinear environmental fluctuation features uploaded from the edge are aligned, normalized, temporally encoded, and dimensionality reduced to obtain 32-dimensional temporal features. The preprocessed features are then input into a lightweight Transformer encoder, which captures the spatiotemporal correlation of environmental fluctuations through a multi-head attention mechanism and outputs a nonlinear environmental response prediction value. The purified residual is obtained by subtracting the model-predicted nonlinear environmental response value from the perceived data after filtering by the simplified mechanical model.

[0009] Furthermore, the process of constructing a healthy baseline evolution pool using the initial data as a baseline is as follows: Seed data for health baselines were selected and spatiotemporally partitioned and aligned according to monitoring area and time period to form a structured baseline feature library. A 90-day sliding time window was configured for each monitoring area, and the baseline pool was divided into a core health layer and a candidate health layer. The initial health benchmark threshold was calculated based on the statistical characteristics of the core health layer. Incremental initial data that meets the health conditions are added to the core health layer in sequence, while old data is removed to maintain the time window length; when continuous changes in the environment pattern are detected, baseline iteration is triggered, and health data that matches the new environment in the candidate health layer is upgraded to the core health layer and the baseline threshold is updated; old data in the candidate health layer is cleaned up regularly, and historical baseline backups of key environment nodes are retained.

[0010] Furthermore, the process of outputting structured damage diagnosis results is as follows: The clean residuals are classified according to monitoring zones and matched with corresponding health baselines. Time series are completed to eliminate instantaneous anomalies, and time series features are extracted. Using the core health layer data of the health baseline evolution pool as a benchmark, the clean residuals are compared with the normal fluctuation range. If the residuals are within the range, they are judged as normal fluctuations; if they exceed the range, they are marked as potential anomalies and enter deep verification. Through mutation feature verification, it is determined whether the residuals show discontinuous jumps. Through time series persistence verification and multimodal data cross-verification, single sensor failures are excluded, and the damage deviation magnitude and mutation rate are quantified. Based on quantitative indicators and specifications, the damage is divided into four risk levels, generating a core judgment result that includes damage judgment conclusions, location, risk level, and quantitative indicators.

[0011] Furthermore, the process of generating a risk assessment report is as follows: The structured damage diagnosis results are engineered and analyzed, and trend extrapolated. Combined with historical data from the health baseline evolution pool, multi-dimensional integration is completed to ultimately form visualized structural risk information and standardized risk assessment reports.

[0012] The cloud-based building structure risk assessment system includes the following modules: The multimodal edge perception module collects multi-source monitoring data in real time, including structural strain, temperature and humidity, traffic flow and visual data of building cracks. It calculates physical guidance residuals and damage mutation index in real time through lightweight algorithms. Based on physical guidance residuals and damage mutation index, it filters out initial data and perception data from multi-source monitoring data and uploads them to the cloud. The dual-drive data decoupling module filters linear environmental interference in the sensing data by simplifying the mechanical model, obtains nonlinear environmental fluctuations, and uses a lightweight Transformer model to fit the nonlinear environmental fluctuations to obtain pure residuals. The dynamic baseline self-evolution module uses initial data as a baseline to construct a healthy baseline evolution pool. When the healthy baseline evolution pool detects a continuous change in the environmental pattern of the building, it automatically iterates the baseline to incorporate the new health status. Based on the pure residual and the healthy baseline evolution pool, it determines whether the anomaly is a real damage and outputs a structured damage diagnosis result. Among them, the continuous change in the environmental pattern is determined by comparing multi-source monitoring data with preset thresholds. The risk visualization assessment module transforms damage diagnosis results into structural risk levels, damage locations, and evolution trends, while also generating risk assessment reports by combining historical data.

[0013] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the cloud-based building structure risk assessment method described above.

[0014] The cloud-based building structure risk assessment method, system, and storage medium provided by this invention have the following beneficial effects: (1) This invention simplifies the mechanical model to accurately remove linear environmental disturbances such as temperature and uniformly distributed load, and then relies on the multi-head attention mechanism of the lightweight Transformer model to capture the spatiotemporal correlation of nonlinear environmental fluctuations such as temperature gradient and gust load and complete accurate fitting to obtain pure residuals that only reflect structural damage. At the same time, combined with the relative quantification judgment logic of damage mutation index, it effectively avoids the false alarm problem caused by slow environmental drift and improves the accuracy of damage identification under complex working conditions.

[0015] (2) This invention completes baseline hierarchical initialization based on initial health data, realizes incremental updates of health data and automatic cleaning of aging data through sliding time windows, accurately identifies continuous changes in environmental patterns and triggers automatic baseline iteration, incorporates health data under new environmental patterns into the benchmark system, and does not require frequent manual calibration and intervention throughout the process. It adapts to the long-term evolution of structural health status and surrounding environment, and reduces the system's operation and maintenance costs and manual input.

[0016] (3) This invention completes real-time acquisition of multi-source data, lightweight preprocessing and anomaly screening at the edge, and only uploads suspicious data, reducing the transmission and computing pressure on the cloud. The cloud sequentially completes environmental interference decoupling, accurate damage determination, risk level verification and evolution trend prediction, and finally outputs structured damage diagnosis results and standardized risk assessment reports. At the same time, it realizes the visualization of risk information and multi-terminal push, and achieves efficient connection from data perception to risk disposal, thereby improving the level of building structure safety management. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method of the present invention; Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 Please see Figure 1 Embodiment 1 of this application provides a building structure risk assessment method based on a cloud platform, the method comprising: Step 1: Collect multi-source monitoring data in real time, including structural strain, temperature and humidity, traffic flow and visual data of building cracks. Calculate physical guidance residuals and damage mutation index in real time using a lightweight algorithm. Based on the physical guidance residuals and damage mutation index, filter out the initial data and sensing data from the multi-source monitoring data and upload them to the cloud. The process of collecting multi-source monitoring data is as follows: Structural strain data: Fiber optic strain sensors or vibrating wire strain gauges are deployed in key stress-bearing parts such as the main beam of a bridge and the floors of a high-rise building to collect strain time history data of the structure under stress and deformation in real time at a sampling frequency of 10Hz, thereby capturing the dynamic mechanical response of the structure. Temperature and humidity data: Temperature and humidity sensors are deployed on the surface of the structure and in the near-field environment to collect ambient temperature and relative humidity at a frequency of 1Hz, providing basic parameters for the subsequent separation of environmental interference. Traffic flow data: Video checkpoints or geomagnetic detectors are deployed at both ends of the bridge or on roads around the building. Through video recognition algorithms or vehicle sensing technology, traffic load information such as traffic flow and vehicle type distribution is statistically analyzed at the minute level to quantify the impact of external loads on the structure. Visual data of building cracks: High-definition industrial cameras are deployed in crack-prone areas, such as bridge piers and shear walls, to collect crack images at a frequency of 10 hours / time. Visual algorithms are used to extract apparent damage features such as crack width and length in real time. All collected multi-source data are mapped to the same timeline through a unified timestamp for spatiotemporal synchronization; at the same time, the raw data is preliminarily cleaned: outliers caused by sensor failures are removed, missing data is filled with linear interpolation, and high-frequency noise is low-pass filtered to ensure the integrity and reliability of the data. Real-time calculation of physical guidance residuals and damage mutation index: During the initialization of the lightweight algorithm at the edge, simplified physical model parameters and lightweight calculation logic are preloaded in the on-site edge computing gateway. The physical model parameters are for the monitored objects, such as bridges and high-rise buildings, and include parameters of the pre-stored simplified mechanical model, including structural span, reference stiffness, temperature correction coefficient, etc. The lightweight algorithm processing optimizes the originally complex residual calculation and mutation detection algorithm by fixing the point and pruning, so that the single sample processing time is less than 100 milliseconds. Real-time calculation of physical guidance residuals is performed synchronously on each batch of real-time monitoring data at the edge. Theoretical response prediction: Based on a simplified physical model, combined with real-time collected temperature, humidity, and traffic flow data, the theoretical response value of the structure under the current environment is calculated. For example, for a simply supported beam bridge, the theoretical deflection at mid-span is calculated using the beam deflection formula; for high-rise buildings, the theoretical strain of each floor is calculated using the strain modal model. When calculating the residuals, the difference between the measured structural strain and the theoretical response value is taken to obtain the physical guidance residuals. These residuals have been preliminarily filtered out linear and deterministic environmental interferences, making them closer to the actual damage signals.

[0020] In real-time calculation of the damage mutation index, the damage mutation index is calculated based on the time series of the physically guided residuals to quantify the degree of change in the residuals. A sliding time window with a length of 1 hour is maintained at the edge, and the sliding mean and sliding standard deviation of the residuals are updated in real time. The damage mutation index is obtained by dividing the absolute difference between the residuals at the current time and the residuals at the previous time by the sliding standard deviation of the residuals at the previous time, which reflects the relative change in the residuals and avoids misjudgment caused by slow environmental drift. Filter out the initial data and perception data and upload them to the cloud: Based on the results of physical guidance residuals and damage mutation index, multi-source monitoring data are classified and screened, and uploaded in a differentiated manner. The original multi-source monitoring data is marked as initial data; the preprocessed results such as physical guidance residuals and damage mutation index are marked as sensing data; when the damage mutation index exceeds a preset threshold, the corresponding sensing data and associated initial data are immediately uploaded to the cloud to ensure real-time transmission of suspicious abnormal signals; when no abnormality is triggered, initial data is synchronized in batches on an hourly basis, which greatly reduces the transmission pressure on the cloud. All data is cached locally at the edge, and if the network is interrupted, the data is temporarily stored and automatically re-uploaded after the network is restored to ensure data integrity.

[0021] Step 2: Filter linear environmental interference in the sensing data by simplifying the mechanical model to obtain nonlinear environmental fluctuations. Fit the nonlinear environmental fluctuations using a lightweight Transformer model to obtain the pure residuals. The process of obtaining nonlinear environmental fluctuations is as follows: A simplified mechanical model customized for the monitored object is preloaded at the edge. The model parameters have been calibrated according to the actual working conditions of the monitored structure. For bridge structures, a simplified Euler-Bernoulli beam model is used, and for high-rise buildings, a simplified inter-story strain and stiffness model is used. Complex nonlinear terms are eliminated, and only linear environmental response correlation logic is retained. Among them, parameter calibration is achieved by pre-storing structural reference mechanical parameters, such as reference stiffness, span, inter-story stiffness coefficient, and environmental linear correction coefficients, such as temperature and stiffness linear coefficients, and load and strain linear coefficients. These parameters are obtained through prior on-site calibration, thereby making the model match the actual structure. Quantitative calculation of linear environmental disturbances: Based on a simplified mechanical model, environmental factors linearly correlated with structural response are extracted from environmental monitoring data associated with sensing data. These factors include average temperature, uniformly distributed traffic load, and static humidity, while nonlinear factors such as gust load and temperature gradient are removed. The linear environmental factors are then input into the simplified mechanical model to calculate the linear theoretical response value of the structure under the current environment. For example, linear deflection caused by traffic load is calculated using a linear formula for load and deflection, and linear strain caused by temperature is corrected using a linear formula for temperature and stiffness. Linear strain of each floor is calculated using a linear formula for inter-story load and strain, and the linear influence of environmental temperature is corrected using a linear formula for temperature and inter-story stiffness. Finally, the measured structural response value from the sensing data is subtracted from the linear theoretical response value calculated by the model to filter out linear environmental disturbances.

[0022] Extraction and characterization of nonlinear environmental fluctuations: After stripping away linear interference, the remaining response difference is a mixture of nonlinear environmental fluctuations and potential damage signals. By maintaining a 24-hour health sliding window at the edge, only response difference data historically judged as healthy are included to construct a baseline distribution of nonlinear environmental fluctuations. The deviation between the current response difference and the mean of the healthy baseline is calculated, and this deviation represents the nonlinear environmental fluctuations after removing linear interference, such as nonlinear strain caused by temperature gradients and nonlinear deflection caused by gust loads. Lightweight feature extraction is performed on the extracted nonlinear environmental fluctuations, retaining only core features such as peak value, volatility, and temporal distribution as inputs for subsequent cloud-based data-driven corrections, avoiding redundant calculations, and thus extracting pure nonlinear environmental fluctuations.

[0023] Fitting nonlinear environmental fluctuations using a lightweight Transformer model: By pruning, quantizing, and distilling the standard Transformer model, only a 2-layer encoder structure and 64-dimensional feature dimension are retained, reducing the number of parameters to 15% of the original model. Based on the nonlinear environmental fluctuation characteristics in historical health data, such as temperature gradients, gust loads, and random traffic flow fluctuations and their corresponding structural response residuals, a pre-training dataset containing over 100,000 samples is constructed, covering different seasons and different load scenarios. The model learns the mapping relationship between nonlinear environmental fluctuations and structural response residuals, and enhances its ability to capture weak nonlinear features through self-supervised contrastive learning. The pre-trained lightweight model is further quantized into a fixed-point arithmetic format and deployed to an edge gateway for local real-time inference. The nonlinear environmental fluctuation features uploaded from the edge, such as temperature gradient peaks, gust load fluctuation rates, and traffic flow change frequency, are aligned with the feature dimensions during pre-training and uniformly normalized to the [0,1] interval to eliminate dimensional differences. Location encoding is added to the time series of fluctuation features to allow the model to perceive the temporal correlation of environmental fluctuations, such as the daily cycle of temperature gradient changes and the morning and evening peak patterns of traffic flow. Principal component analysis is used to compress high-dimensional fluctuation features to 32 dimensions, reducing the computational load of the model while retaining more than 95% of the fluctuation information. Nonlinear fitting of lightweight Transformer models: The preprocessed temporal features are input into a lightweight Transformer encoder, which uses a multi-head attention mechanism to capture the spatiotemporal correlation of nonlinear environmental fluctuations, such as the coupling effect between temperature gradients and traffic flow fluctuations. The model outputs the predicted structural response value corresponding to the current nonlinear environmental fluctuation, which represents all nonlinear disturbances caused by environmental factors. The model is incrementally fine-tuned every 7 days in the cloud using new health data to adapt the model to long-term changes in environmental patterns and avoid a decrease in fitting accuracy. The nonlinear environmental response value predicted by the lightweight Transformer model is subtracted from the perceived data filtered by the simplified mechanical model, i.e., nonlinear environmental fluctuations and potential damage signals, to obtain the clean residual.

[0024] Step 3: Construct a health baseline evolution pool using the initial data as a baseline. When the health baseline evolution pool detects a continuous change in the environmental pattern of the building, it automatically iterates the baseline to incorporate the new health status. Based on the pure residual and the health baseline evolution pool, it determines whether the anomaly is a real damage and outputs a structured damage diagnosis result. Among them, the continuous change in the environmental pattern is determined by comparing multi-source monitoring data with a preset threshold. Construct a healthy baseline evolution pool: In the initial screening of health data, historical data without abnormal operation were combined to select datasets identified as healthy from the uploaded initial data as baseline seeds. The screening criteria were: no abrupt changes in structural response, environmental fluctuations within the historical normal range, and no damage recorded by manual inspection. The initial data was partitioned by monitoring area, such as mid-span of bridges or floors of high-rise buildings, and by time period to ensure that health data from different areas and time periods were independently grouped, avoiding cross-regional environmental interference. Core statistical features were extracted from the health data of each partition, including response mean, moving standard deviation, fluctuation period, and linear trend coefficient, forming a structured baseline feature library. During the initial construction of the health baseline evolution pool, a 90-day sliding time window is configured for each monitoring area as the basic storage unit of the health baseline evolution pool. The length of the time window is set according to the structural service characteristics and environmental change cycle to balance baseline stability and evolution sensitivity. The baseline pool is divided into a core health layer and a candidate health layer. The core health layer stores nearly 30 days of health data for real-time damage assessment. The candidate health layer stores 30-90 days of health data to cope with long-term changes in environmental patterns. Based on the statistical characteristics of the core health layer, the initial health benchmark threshold for each monitoring area is calculated, such as the residual mean ± 3 times the standard deviation, as the initial basis for damage assessment. Dynamic update mechanism of the baseline evolution pool: After the initial data uploaded from the edge undergoes real-time health assessment, if it meets the health conditions of no damaging mutations and residuals within the threshold range, it is added to the core health layer in chronological order. Simultaneously, the oldest 1-day data in the core health layer is automatically removed to maintain a stable time window length. When a continuous change in the environmental pattern is detected, such as traffic flow increasing by more than 20% for 7 consecutive days or seasonal temperature fluctuations exceeding historical thresholds, the system automatically triggers baseline iteration: health data matching the current environmental pattern in the candidate health layer is upgraded to the core health layer, and the health baseline threshold is recalculated to ensure the baseline is adapted to the current environment. Old data exceeding 90 days in the candidate health layer is periodically cleaned to avoid redundant storage. At the same time, health data for key environmental nodes, such as extreme weather and major traffic events, is retained as a historical baseline backup for retrospective analysis. Regularly compare the baseline characteristics of different monitoring areas. If significant differences are found, such as the baseline fluctuation of a certain floor being much greater than that of other floors, manual review is triggered to rule out the possibility of sensor failure or local hidden damage. The current baseline pool is verified monthly using historical damage event data. If false alarms or missed reports occur, the time window length or health threshold is adjusted to optimize the judgment accuracy of the baseline pool. If the baseline characteristics of crack visual data and strain data are combined, and a single-modal anomaly occurs but the multimodal data is not correlated, it is determined to be environmental interference to avoid baseline failure caused by single sensor data deviation.

[0025] The process of detecting whether the environmental patterns of a building are undergoing continuous change is as follows: From the initial data and sensing data uploaded from the edge, core environmental factors strongly correlated with structural response are extracted as environmental model features, including key indicators such as average temperature, traffic flow, humidity change rate, and wind load level. Each feature is processed into a time series, with a statistical granularity of 1 hour, to record the numerical changes and distribution patterns of the features in real time. At the same time, an environmental model feature library is established in the health baseline evolution pool to continuously store nearly 90 days of environmental model data, providing a reference benchmark for subsequent change detection. Based on the environmental pattern feature library, the system accurately identifies continuous changes in environmental patterns through dual-dimensional judgment rules, distinguishing them from short-term, random environmental fluctuations. It calculates the average environmental pattern feature over the current seven consecutive days and compares it with the average feature over the past 30 days in the health baseline evolution pool to obtain the feature change rate. If the feature change rate of a single core environmental factor, such as traffic flow or average temperature, exceeds a preset threshold (e.g., 20%), or the combined change rate of multiple environmental factors exceeds 15%, the system proceeds to the next verification step. The system verifies whether the feature change has temporal continuity. If the trend remains unchanged for 10 consecutive days without decline, it is determined that the environmental pattern of the building has undergone continuous change, triggering a baseline iteration command. If it is a short-term fluctuation, such as a single-day traffic peak or a short-term sharp drop in temperature, it is determined to be environmental interference, and iteration is not triggered. After triggering the baseline iteration command, the system selects health data matching the new environmental pattern from the incremental initial data in the health baseline evolution pool as the data source for baseline iteration. The process of automatically iterating the baseline to incorporate new health states: Starting from the time point when the environmental pattern is determined to be continuously changing, initial data and sensing data are collected for the next 14 consecutive days. Based on the damage determination rules of no damage mutation, pure residual within the health threshold range, and no abnormality in multimodal data cross-validation, the batch of data is strictly screened for health, and abnormal and interfering data are removed. The effective data after screening is labeled as health status data under the new environmental pattern and added to the candidate health layer of the health baseline evolution pool to complete the storage of new health data. The health data matching the new environmental pattern in the candidate health layer are gradually replaced with the old data corresponding to the old environmental pattern in the core health layer. The replacement rate is increased by 10% per day to avoid baseline abrupt changes affecting the accuracy of the judgment. Based on the iterated core health layer data, the statistical characteristics of the health baseline of each monitoring area are recalculated, including key judgment parameters such as the mean of pure residuals, the moving standard deviation, and the damage mutation index threshold. Continuous changes in the environmental pattern are determined by comparing multi-source monitoring data with preset thresholds. According to the new statistical characteristics, the threshold range for damage judgment is updated, such as pure residuals ± 3 times the standard deviation, so that the threshold matches the normal structural response under the new environmental pattern. The updated health baseline parameters and threshold ranges are synchronized to the processing modules at the edge and in the cloud, replacing the old baseline parameters and enabling the new baseline to be implemented across the entire system. At the same time, information such as the type of environmental mode change, the content of the baseline iteration, and the iteration accuracy are recorded in the historical log of the health baseline evolution pool.

[0026] The process of determining whether an abnormality is a real injury is as follows: The clean residuals are preprocessed and classified according to monitoring zones, such as mid-span of bridge main beams, bearings, shear walls of high-rise buildings, and beam-column joints. The clean residuals are matched with the dynamic benchmarks of the corresponding zones in the healthy baseline evolution pool to avoid cross-regional data interference with the judgment results. The clean residuals are time series completed to remove isolated outliers caused by instantaneous sensor failures and retain the continuous residual change trend. The current clean residuals are spatiotemporally correlated with the residual data of the same region in the past 1 hour and 24 hours to extract the temporal characteristics such as the rate of change and fluctuation amplitude of the residuals. Using the core health layer data of the corresponding monitoring zone in the health baseline evolution pool as a benchmark, basic anomaly screening of pure residuals is performed to distinguish between normal fluctuations and potential anomalies. By retrieving the latest health baseline statistical parameters of the zone, including the health mean, moving standard deviation, and normal fluctuation range of the pure residuals (mean ± 3 times the standard deviation), the current value of the pure residuals is compared with the normal fluctuation range of the baseline. If the residuals are within the range, they are judged as environmental random noise or normal fluctuations and directly marked as no damage, terminating the current judgment process. If the residuals exceed the normal fluctuation range, they are judged as potential anomalies, and in-depth verification is performed to distinguish between false anomalies and real damage. Based on the screened potential anomalies, and combining historical data from the healthy baseline evolution pool with the inherent characteristics of structural damage, a triple verification process is used to eliminate false anomalies and identify true damage. All verifications are based on comparisons with historical benchmark data from the healthy baseline evolution pool. The damage mutation index benchmark threshold for that partition in the healthy baseline evolution pool is retrieved, and the residual mutation index corresponding to the current potential anomaly is calculated. If the index does not exceed the threshold, it is determined to be a gradual drift during the baseline iteration transition period, i.e., a false anomaly. If the index exceeds the threshold, it indicates that the residual exhibits discontinuous jump characteristics, consistent with the response law of structural damage, and proceeds to the next level of verification. Potential anomalies are then examined. If the duration of the abnormal residual is a single isolated jump followed by a rapid return to the baseline range, it is judged as transient sensor interference, i.e., a false anomaly. If the abnormal residual lasts for more than 5 minutes and shows a stable upward trend or a continuous deviation trend, it proceeds to the next level of verification. When multiple source data of the same monitoring zone are correlated, such as when the strain pure residual is abnormal, the crack visual data and vibration data of that area are retrieved simultaneously. If only a single data dimension shows an anomaly, and other dimensions match the healthy baseline, it is judged as a single sensor failure, i.e., a false anomaly. If multiple dimensions of data show coordinated anomalies and the anomaly trends are consistent, it is judged as real structural damage. After determining the damage to be genuine, based on historical data from the healthy baseline evolution pool and structural engineering specifications, the damage is quantitatively analyzed and risk levels are classified. Key indicators corresponding to the damage are calculated, including the deviation of the pure residual from the healthy mean, the mutation rate, and the duration. Simultaneously, by combining correlation data from historical damage cases in the healthy baseline evolution pool, the scope of the damage's impact is preliminarily determined. Based on the sensor deployment locations corresponding to abnormal pure residuals, and combined with the structural stress system (e.g., the mid-span and support locations corresponding to abnormal beam strain), the physical location of the damage is precisely pinpointed, and specific monitoring zones and component names are identified. Based on the damage quantification indicators and structural design specifications, the damage is classified into four risk levels: Level I Minor: The residual deviation is less than 5 times the standard deviation, with no continuous increasing trend, and does not affect the normal service of the structure; Level II (General): 5 times standard deviation ≤ residual deviation < 8 times standard deviation, slowly widening, requiring enhanced monitoring; Level III Severe: Residual deviation of 8 times standard deviation ≤ residual deviation < 10 times standard deviation, rapidly expanding and affecting the stress of local components; Level IV Critical: Residual deviation ≥ 10 times the standard deviation, multi-dimensional coordinated anomalies, which may endanger the overall structural safety; The damage assessment conclusion determines whether it is a real injury, the physical location of the injury including monitoring zones and specific components, the injury risk level, core quantitative indicators combined with residual deviation magnitude, mutation index and duration, and the assessment time; details of abnormal data from multimodal cross-validation, preliminary prediction of the injury development trend, comparison benchmark parameters of the health baseline evolution pool, and a record of the verification process for this assessment; generates structured data in standardized JSON format, which is simultaneously pushed to the cloud-based risk visualization module and terminal management platform. At the same time, a dedicated file for this injury is established in the health baseline evolution pool to record the entire life cycle data of the injury. False anomalies screened out in this assessment are also classified and archived, such as sensor failures and transient environmental interference, and the interference feature library of the health baseline evolution pool is updated synchronously.

[0027] Step 4: Convert the damage diagnosis results into structural risk level, damage location and evolution trend, and generate a risk assessment report by combining historical data. Judgment indicators are extracted from the diagnostic results, including the physical location of the damage, the initial risk level assessment, the deviation of the pure residual, the mutation index, the duration of the anomaly, and multimodal anomaly data. Location mapping is completed according to the structural component dimension, converting sensor monitoring zones into engineering component names. For example, bridge monitoring zone 1 is mapped to the left mid-span box girder of the main bridge, and building monitoring zone 5 is mapped to the 20th floor east shear wall of a high-rise office building, thus determining the specific component to which the damage belongs and the key stress-bearing parts. Quantitative indicators are then converted into engineering expressions, transforming algorithmic indicators such as the deviation of the pure residual and the mutation index into expressions that engineers can interpret. For example, a residual deviation of 6 times the standard deviation of the healthy baseline is converted into a significant deviation of the structural response from the normal state, exceeding the normal fluctuation range. Based on the load-bearing status of the damaged component (e.g., main load-bearing beams of bridges and core tubes of buildings are classified as Level 1 critical components, while auxiliary enclosure structures are classified as Level 3 critical components), a risk weighting coefficient is set. When a Level 1 component experiences an anomaly, the risk level increases by one level; when a Level 3 component experiences an anomaly, the risk level remains unchanged. This avoids excessive warnings for minor component anomalies and insufficient warnings for critical component anomalies. Historical data on the same component and type of anomalies are retrieved from the health baseline evolution pool and compared with the quantitative indicators of the current anomaly, namely the deviation magnitude and development rate. If the magnitude of the current anomaly is significantly greater than similar historical cases, or the development rate is faster, the risk level is appropriately increased. If it is the first occurrence of this type of anomaly, the risk level is determined with reference to the code requirements. Combining the weighted results with historical verification, the structural risk level is determined. The final assessment categorized the damage into four levels, consistent with the damage assessment process and aligned with the actual engineering conditions. Level I minor risk damage involves only slight localized response anomalies that do not affect the normal load-bearing capacity of the components and require no immediate intervention, only enhanced monitoring. Level II general risk damage causes significant deviations in the localized response of the components, with a slow development trend, requiring the development of a regular monitoring plan to assess the development trend. Level III severe risk damage has affected the localized load-bearing performance of the components, with the response anomalies continuing to expand, potentially leading to a decline in the performance of localized components, requiring on-site inspection and repair as soon as possible. Level IV critical risk damage involves core load-bearing components, with severe and rapidly developing abnormal responses that may endanger the overall structural safety, requiring immediate activation of the emergency response plan, closure of the relevant areas, and emergency inspection. Based on the analyzed damage location information, combined with the structural BIM model or floor plan, the accurate spatial location and visual annotation of the damage are completed. The spatial coordinate system of the component to which the damage belongs is matched with that of the structure to clarify the three-dimensional spatial location of the damage. For example, the damage location of the box girder at the mid-span of the bridge is X:1250m, Y:35m, Z:28m. The damage location of the building shear wall is the 20th floor, facing east, 62m from the ground and 1.5m from the wall end. Based on the time series data of the clean residuals, historical monitoring data of the healthy baseline evolution pool, and the development characteristics of this injury, short-term prediction of the injury evolution trend is achieved through trend fitting. Lightweight algorithms such as linear fitting and exponential fitting are used to fit the trend of quantitative indicators of the injury, deviation magnitude, and abnormal duration to determine whether the injury development rate is slow, uniform, or rapid. Based on the fitting results, the injury development trend is predicted for the next 7 days and 30 days. For example, if the residual deviation magnitude is predicted to increase from 6 times the standard deviation to 7.2 times the standard deviation in the next 7 days, the risk level may rise from level II to level III, clarifying the development trend of the injury if left untreated. The report integrates three aspects of historical data retrieved from the health baseline evolution pool: long-term health monitoring data of the damaged component, showing the component's historical response status and change patterns; historical treatment records of anomalies of the same type and location, providing a reference for this treatment; and statistical data on the overall health status of the structure, clarifying whether the damage is a local anomaly or a common problem in the overall structure. The verified risk level, damage location, evolution trend, and other visualized risk information, as well as the generated risk assessment report, are simultaneously pushed to all terminals of the system.

[0028] Example 2 Please see Figure 2 Based on Example 1, Example 2 of this application also provides a cloud-based building structure risk assessment system, including the following specific modules: The multimodal edge perception module collects multi-source monitoring data in real time, including structural strain, temperature and humidity, traffic flow and visual data of building cracks. It calculates physical guidance residuals and damage mutation index in real time through lightweight algorithms. Based on physical guidance residuals and damage mutation index, it filters out initial data and perception data from multi-source monitoring data and uploads them to the cloud. The dual-drive data decoupling module filters linear environmental interference in the sensing data by simplifying the mechanical model, obtains nonlinear environmental fluctuations, and uses a lightweight Transformer model to fit the nonlinear environmental fluctuations to obtain pure residuals. The dynamic baseline self-evolution module uses initial data as a baseline to construct a healthy baseline evolution pool. When the healthy baseline evolution pool detects a continuous change in the environmental pattern of the building, it automatically iterates the baseline to incorporate the new health status. Based on the pure residual and the healthy baseline evolution pool, it determines whether the anomaly is a real damage and outputs a structured damage diagnosis result. Among them, the continuous change in the environmental pattern is determined by comparing multi-source monitoring data with preset thresholds. The risk visualization assessment module transforms damage diagnosis results into structural risk levels, damage locations, and evolution trends, while also generating risk assessment reports by combining historical data.

[0029] Example 3 Based on Example 1, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the cloud-based building structure risk assessment method described above. For example, storage media can be deployed independently or integrated, adapting to the distributed computing architecture of cloud platforms. Specific adaptation scenarios include: Integrated and deployed on the edge computing gateway, by embedding the storage medium into the edge gateway at the building structure monitoring site, lightweight program instructions and locally collected data are stored to achieve real-time preprocessing and anomaly screening at the edge. Deployed on cloud platform server nodes, by configuring storage media in the physical / virtual servers of the cloud platform, it stores the complete program instruction set and full monitoring data, health baseline evolution pool, risk assessment reports, etc., to achieve deep computing and data archiving in the cloud. Distributed deployment across edge and cloud storage nodes, the edge storage media stores real-time collected data and lightweight programs, while the cloud storage media stores complete programs and historical data. Data synchronization and command interaction are achieved through the network, balancing real-time performance and data integrity.

[0030] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0031] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0032] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A cloud-based method for assessing building structural risks, characterized in that, The method includes: Step 1: Collect multi-source monitoring data in real time, including structural strain, temperature and humidity, traffic flow and visual data of building cracks. Calculate physical guidance residuals and damage mutation index in real time using a lightweight algorithm. Based on the physical guidance residuals and damage mutation index, filter out the initial data and sensing data from the multi-source monitoring data and upload them to the cloud. Step 2: Filter linear environmental interference in the sensing data by simplifying the mechanical model to obtain nonlinear environmental fluctuations. Fit the nonlinear environmental fluctuations using a lightweight Transformer model to obtain the pure residuals. Step 3: Construct a health baseline evolution pool using the initial data as a baseline. When the health baseline evolution pool detects a continuous change in the environmental pattern of the building, it automatically iterates the baseline to incorporate the new health status. Based on the pure residual and the health baseline evolution pool, it determines whether the anomaly is a real damage and outputs a structured damage diagnosis result. Among them, the continuous change in the environmental pattern is determined by comparing multi-source monitoring data with a preset threshold. Step 4: Transform the damage diagnosis results into structural risk level, damage location and evolution trend, and generate a risk assessment report by combining historical data.

2. The cloud-based building structure risk assessment method according to claim 1, characterized in that, The process of calculating the physical guidance residual and damage mutation index in real time using a lightweight algorithm is as follows: By preloading simplified physical model parameters and lightweight operation logic in the edge computing gateway where the lightweight algorithm runs, and based on the simplified physical model parameters, combined with real-time collected temperature, humidity and traffic flow data, the theoretical response value of the structure under the current environment is calculated. The difference between the measured structural strain and the theoretical response value is used to obtain the physical guidance residual. A sliding time window of 1 hour is obtained in the edge computing gateway through lightweight computing logic. The sliding standard deviation is obtained based on the physical guidance residual. The damage mutation index is obtained by dividing the absolute difference between the physical guidance residual at the current moment and the physical guidance residual at the previous moment by the sliding standard deviation at the previous moment.

3. The cloud-based building structure risk assessment method according to claim 2, characterized in that, The process of filtering out the initial data and the perceived data is as follows: Based on the results of physical guidance residuals and damage mutation index, multi-source monitoring data are classified and screened. The processed multi-source monitoring data is used as the initial data, and the preprocessed results of physical guidance residuals and damage mutation index are used as the sensing data.

4. The cloud-based building structure risk assessment method according to claim 3, characterized in that, The process of obtaining nonlinear environmental fluctuations is as follows: From the environmental monitoring data associated with the sensing data, environmental factors that are linearly related to the structural strain are extracted. The environmental factors are input into a simplified mechanical model to calculate the linear theoretical response value of the structural strain under the current environment. The measured structural strain response value in the sensing data is subtracted from the linear theoretical response value calculated by the model to complete the filtering of linear environmental interference. Based on the filtered linear environmental disturbances, a healthy sliding window with a length of 24 hours is obtained at the edge to construct the baseline distribution of nonlinear environmental fluctuations. The deviation between the measured structural strain response value and the mean of the healthy baseline is calculated based on the baseline distribution of nonlinear environmental fluctuations, which is the nonlinear environmental fluctuation.

5. The cloud-based building structure risk assessment method according to claim 1, characterized in that, The process of obtaining the pure residual is as follows: By pre-training and deploying a lightweight Transformer model in the cloud, the nonlinear environmental fluctuation features uploaded from the edge are aligned, normalized, temporally encoded, and dimensionality reduced to obtain 32-dimensional temporal features. The preprocessed 32-dimensional temporal features are then input into a lightweight Transformer encoder, which captures the spatiotemporal correlation of environmental fluctuations through a multi-head attention mechanism and outputs a nonlinear environmental response prediction value. The nonlinear environmental response prediction value is then subtracted from the perception data filtered by the simplified mechanical model to obtain the clean residual.

6. The cloud-based building structure risk assessment method according to claim 5, characterized in that, The process of constructing a healthy baseline evolution pool using initial data as a baseline is as follows: Seed data for health baselines were selected and spatiotemporally partitioned and aligned according to monitoring area and time period to form a structured baseline feature library. A 90-day sliding time window was configured for each monitoring area, and the baseline pool was divided into a core health layer and a candidate health layer. The initial health benchmark threshold was calculated based on the statistical characteristics of the core health layer. Incremental initial data that meet the health conditions are added to the core health layer in sequence, while old data is removed to maintain the length of the time window; When a continuous change in the environment pattern is detected, a baseline iteration is triggered. The health data matching the new environment in the candidate health layer is upgraded to the core health layer and the baseline threshold is updated. The old data in the candidate health layer is cleaned up regularly, and the historical baseline backup of the environment node is retained.

7. The cloud-based building structure risk assessment method according to claim 6, characterized in that, The process of outputting structured damage diagnosis results is as follows: The clean residuals are classified according to the monitoring zones and matched with the corresponding health baselines. The time series is completed to remove instantaneous anomalies. Temporal features are extracted. Based on the core health layer data of the health baseline evolution pool, the clean residuals are compared with the normal fluctuation range. If the residuals are within the range, they are judged as normal fluctuations; if they exceed the range, they are marked as potential anomalies and enter deep verification. By using mutation feature verification, it is determined whether the residual shows discontinuous jumps; through time series persistence verification and multimodal data cross-verification, single sensor failures are eliminated, the damage deviation magnitude and mutation rate are quantified, and the damage is divided into four risk levels according to quantitative indicators and specifications, generating core judgment results that include damage judgment conclusions, location, risk level and quantitative indicators.

8. The cloud-based building structure risk assessment method according to claim 1, characterized in that, The process of generating a risk assessment report is as follows: The structured damage diagnosis results are engineered and analyzed, and trend extrapolated. Combined with historical data from the health baseline evolution pool, multi-dimensional integration is completed to ultimately form visualized structural risk information and standardized risk assessment reports.

9. A cloud-based building structure risk assessment system, characterized in that, Includes the following modules: The multimodal edge perception module collects multi-source monitoring data in real time, including structural strain, temperature and humidity, traffic flow and visual data of building cracks. It calculates physical guidance residuals and damage mutation index in real time through lightweight algorithms. Based on physical guidance residuals and damage mutation index, it filters out initial data and perception data from multi-source monitoring data and uploads them to the cloud. The dual-drive data decoupling module filters linear environmental interference in the sensing data by simplifying the mechanical model, obtains nonlinear environmental fluctuations, and uses a lightweight Transformer model to fit the nonlinear environmental fluctuations to obtain pure residuals. The dynamic baseline self-evolution module uses initial data as a baseline to construct a healthy baseline evolution pool. When the healthy baseline evolution pool detects a continuous change in the environmental pattern of the building, it automatically iterates the baseline to incorporate the new health status. Based on the pure residual and the healthy baseline evolution pool, it determines whether the anomaly is a real damage and outputs a structured damage diagnosis result. Among them, the continuous change in the environmental pattern is determined by comparing multi-source monitoring data with preset thresholds. The risk visualization assessment module transforms damage diagnosis results into structural risk levels, damage locations, and evolution trends, while also generating risk assessment reports by combining historical data.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the cloud-based building structure risk assessment method as described in any one of claims 1 to 8.

Images