Construction method of building quality evaluation model based on big data

Abstract

The application discloses a kind of based on big data's construction method of building quality evaluation model;Including obtaining first environmental state parameter flow and constructing space microclimate dissipation coefficient matrix;Second physical surface reflection wave array is obtained to construct surface texture evolution characteristic matrix;Both are input into multimodal cross attention neural network, and based on physical elasticity mechanics loss function regularization constraint output material internal hidden stress prediction sequence;The discrete numerical change difference value of stress sequence is extracted to calculate first-order discrete difference slope, and exponential extrapolation mathematical model is constructed to solve the corresponding time stamp to generate fatigue time evaluation index;In response to the index satisfying early warning threshold, traverse associated weight matrix to determine the maximum weight node, substitute into graph topological structure to carry out reverse jacobian determinant derivation and feature truncation, and the target abnormal environment position entity space coordinate causing fatigue state is solved out;The present application realizes 72 hours of advance timing quantization evaluation and cause tracing.

Description

Technical Field

[0001] This invention relates to the field of building quality and health monitoring and big data processing technology, specifically a method for constructing a building quality assessment model based on big data. Background Technology

[0002] With the deep integration of IoT, machine vision, and deep learning technologies, intelligent building lifecycle health monitoring has become a core component of smart city construction. In the operation and maintenance of modern large-scale landmarks, super high-rise buildings, and complex underground projects, how to utilize multimodal sensing data to achieve advanced prediction and temporal quantitative assessment of hidden defects within the structure has become a key research direction for ensuring building structural safety and preventing sudden disasters. Existing monitoring systems are gradually evolving from traditional single-point manual inspections to automated time-series data analysis, striving to capture potential degradation trends before obvious structural damage occurs.

[0003] However, existing building quality monitoring and assessment solutions still face numerous technical challenges and limitations in practical engineering applications. The dimensional mismatch of multimodal data and the alignment challenges of heterogeneous manifolds are significant. Traditional solutions, when fusing one-dimensional time-series parameters such as environmental temperature and humidity with multi-dimensional visual texture images of surfaces, often only stitch together data at the data layer or a simple feature layer, lacking strong constraints from physical and mechanical mechanisms. This makes it difficult to effectively mine the cross-correlation features between different modes, and the model's robustness is easily affected by noise pollution caused by harsh environments. The lag and static threshold dependence of time-series quantitative assessments are prominent limitations. Existing technologies generally use a method of preset static thresholds, that is, triggering an alarm only after physical indicators actually exceed safety boundaries. This cannot proactively extrapolate the rapid evolution trend of hidden stresses within the structure, making it difficult to provide a golden warning window of tens of hours. The black-box nature of end-to-end deep learning models and the lack of traceability mechanisms limit their practical value. Although the models can output high-dimensional predicted values, they cannot reverse-engineer the initial environmental causes and spatial coordinates that led to the sudden change in stress peaks. This makes it impossible for maintenance personnel to implement precise physical-targeted interventions and determine responsibility for quality accidents after receiving risk signals.

[0004] Therefore, this invention provides a method for constructing a building quality assessment model based on big data. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for constructing a building quality assessment model based on big data, thereby solving the technical problems described in the background section.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a building quality assessment model based on big data, comprising the following steps: S1. Obtain the first environmental state parameter stream within a preset period, perform discrete quantization operation on the first environmental state parameter stream, and extract the environmental time series feature sequence that reflects the spatial temperature and humidity evolution law. S2. Obtain the second physical surface reflection wave array within a preset period, perform a de-identification masking operation based on local differential privacy on the second physical surface reflection wave array, remove pixel clusters containing biological features, and extract the surface texture evolution feature matrix reflecting the nonlinear deformation of the material surface from the masked data. S3. Input the environmental temporal feature sequence and the surface texture evolution feature matrix into a pre-trained multimodal cross-attention neural network. The multimodal cross-attention neural network uses a preset elasticity loss function to perform physical dimension alignment and manifold mapping on the environmental temporal feature sequence and the surface texture evolution feature matrix, calculates the internal hidden stress prediction sequence used to characterize the internal stress state of the material, and simultaneously extracts the modal correlation attention weight matrix generated by the multimodal cross-attention neural network when performing manifold mapping. S4. Extract the numerical variation difference of the internal hidden stress prediction sequence at adjacent time steps, compare the numerical variation difference with the historical safety evolution benchmark rate, and calculate the fatigue time assessment index for the future generation of microcracks inside the structure. S5. In response to the fatigue time assessment index meeting the preset structural fatigue early warning threshold, the maximum weight node is selected based on the modal association attention weight matrix. The target abnormal environment location that caused the fatigue state is traced back according to the maximum weight node, and the building concealed quality assessment result containing the fatigue time assessment index and the target abnormal environment location is output.

[0007] (III) Beneficial Effects This invention provides a method for constructing a building quality assessment model based on big data, which has the following beneficial effects: This invention constructs a spatial microclimate dissipation coefficient matrix and a visual manifold feature matrix through S1 to S3, and uses a differentiable dynamic time warping algorithm to achieve phase alignment of spatiotemporal features. By introducing a de-identification masking operation based on local differential privacy and injecting a Laplacian noise matrix, combined with a smoothing term factor design, irrelevant pixel clusters containing biological features are effectively removed, thus achieving privacy protection. By using a Pareto multi-objective adaptive gradient weighting algorithm and a cross-modal information confusion negative exponential decay mechanism, the elasticity loss function is divided and amplified to correct it, which greatly improves the model's noise resistance and confidence penalty sensitivity when facing poor appearance images such as dust occlusion and light spot pollution, and enhances its generalization robustness in complex environments. This invention extracts the difference in stress values ​​between adjacent time steps using S4 to calculate the first-order discrete difference slope, and substitutes it as an exponential growth rate factor into the exponential extrapolation mathematical model with the historical safety evolution benchmark as the initial constant term. By solving the timestamp parameter through inverse natural logarithm operation, it completely eliminates the serious lag in traditional monitoring that only alarms when the static threshold is exceeded. It transforms the method based on historical statistics into a future evolution prediction based on dynamic slope, realizing a forward-looking quantitative assessment of tens of hours before the internal stress accumulation reaches the material fracture limit, thus securing a golden intervention window for the governance of building safety engineering. This invention automatically activates the white-box reverse inference program when the fatigue time assessment index reaches the warning threshold via S5; it uses the synchronously retained modal association attention weight matrix to lock the node with the largest weight, and performs inverse Jacobian determinant differentiation on the graph topology stacked structure to calculate the extreme coordinates of the input layer whose contribution to sensitivity is greater than the preset contribution threshold, and maps them back to the three-dimensional spatial reference frame; it endows the neural network with powerful traceability and interpretability capabilities, which can not only provide future risk time, but also accurately track the initial spatial environmental causes that trigger mechanical deformation, and realize the complete traceability and precise targeted management of the logical chain of quality accidents. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the steps of the method for constructing a building quality assessment model based on big data according to the present invention. Figure 2 This is a schematic diagram illustrating the construction steps of the internal hidden stress prediction sequence in the construction method of the building quality assessment model based on big data of the present invention. Figure 3 This is a schematic diagram illustrating the steps of performing negative exponential decay mapping calculations on cross-modal information confusion using the confusion mapping decay coefficient of the big data-based building quality assessment model of this invention. Detailed Implementation

[0009] 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.

[0010] Please see Figure 1-3 This invention provides a method for constructing a building quality assessment model based on big data, comprising the following steps: S1. Obtain the first environmental state parameter stream within a preset period, perform discrete quantization operation on the first environmental state parameter stream, and extract the environmental time series feature sequence that reflects the spatial temperature and humidity evolution law. S2. Obtain the second physical surface reflection wave array within a preset period, perform a de-identification masking operation based on local differential privacy on the second physical surface reflection wave array, remove pixel clusters containing biological features, and extract the surface texture evolution feature matrix reflecting the nonlinear deformation of the material surface from the masked data. S3. Input the environmental temporal feature sequence and the surface texture evolution feature matrix into a pre-trained multimodal cross-attention neural network. The multimodal cross-attention neural network uses a preset elasticity loss function to perform physical dimension alignment and manifold mapping on the environmental temporal feature sequence and the surface texture evolution feature matrix, calculates the internal hidden stress prediction sequence used to characterize the internal stress state of the material, and simultaneously extracts the modal correlation attention weight matrix generated by the multimodal cross-attention neural network when performing manifold mapping. S4. Extract the numerical variation difference of the internal hidden stress prediction sequence at adjacent time steps, compare the numerical variation difference with the historical safety evolution benchmark rate, and calculate the fatigue time assessment index for the future generation of microcracks inside the structure. S5. In response to the fatigue time assessment index meeting the preset structural fatigue early warning threshold, the maximum weight node is selected based on the modal association attention weight matrix. The target abnormal environment location that caused the fatigue state is traced back according to the maximum weight node, and the building concealed quality assessment result containing the fatigue time assessment index and the target abnormal environment location is output.

[0011] The steps for constructing the environmental temporal feature sequence are as follows: map the first environmental state parameter stream into graph topology data containing feature vectors of multiple spatial nodes; input the graph topology data into a graph convolutional long short-term memory network; calculate the numerical gradient changes of the first environmental state parameter stream among various spatial nodes to generate a spatial microclimate dissipation coefficient matrix as the environmental temporal feature sequence.

[0012] Feature extraction is performed using a graph convolutional temporal network, with a Python script containing a configuration file loaded. System hyperparameters (in this embodiment, system hyperparameters include the number of hidden layer nodes and the initial weights of the topological adjacency matrix) are read from an external spreadsheet file. Discrete data points are used as nodes in the graph topology, and the gradient partial derivatives between nodes are calculated based on spatial relative coordinates, thereby outputting a spatial microclimate dissipation coefficient matrix representing differences in energy flow. The graph convolutional network is used to process discrete environmental parameter streams, effectively overcoming the accuracy degradation problem caused by sparse data sources, and enabling the high-precision three-dimensional spatial energy dissipation field to be inverted with extremely low-density collection points. The system hyperparameter determination steps are as follows: Construct an offline reference dataset containing time-series data of the building entity environment, and all data have undergone cubic spline interpolation for missing values ​​and range standardization; initialize the graph convolutional long short-term memory network structure; use the number of hidden layer nodes as the first hyperparameter variable and the initial weights of the topological adjacency matrix as the second hyperparameter variable; define a supervised training loss function, which is used to calculate the root mean square error between the spatial dissipation matrix predicted by the network and the calibration matrix of the entity sensor; extract multiple hyperparameter combinations with a fixed step size within a preset node number range (64 to 512 in this embodiment) and weight range (-1 to +1 in this embodiment); for each hyperparameter combination, perform pre-processing on the offline reference dataset. Backpropagation and backpropagation are performed until the root mean square error (RMSE) decreases to the convergence threshold or reaches the maximum number of iterations (the convergence threshold is preferably set to 0.001, with a reasonable range within the closed interval [0.0001, 0.005]; the maximum number of iterations is extracted into local memory and instantiated, with a preferably set value of 500 iterations, with a reasonable range within the closed interval [200, 1000] iterations); the network model state that minimizes the RMSE of the validation set is extracted; the number of hidden layer nodes corresponding to the network model state is solidified as the final first basic configuration parameter; the node connection weight matrix that converges in the network model state is extracted and solidified as the final second basic configuration parameter, and both are written into an external spreadsheet for S1 to retrieve in real time.

[0013] The specific processing flow for calculating the spatial microclimate dissipation coefficient matrix in the Python program script loaded with the configuration file in S1 is as follows: The configuration file is declared as runtime configuration parameters and stored in an external structured carrier; system hyperparameters are read from an external spreadsheet file; the number of hidden layer nodes in the network is extracted to local runtime memory and instantiated as a constant term 256 with a priority value; the initial weights of the topological adjacency matrix are read; the first environmental state parameter stream (including spatial temperature and spatial humidity parameters) and the three-dimensional spatial relative coordinates corresponding to each discrete data point are obtained; the spatial Euclidean distance between any two adjacent discrete data points is calculated, and the numerical difference of the first environmental state parameter stream between two adjacent discrete data points is calculated; the numerical difference is divided by the spatial Euclidean distance to obtain the discrete spatial gradient value, and the discrete spatial gradient value is multiplied element-wise with the preset initial weights of the topological adjacency matrix to generate the spatial microclimate dissipation coefficient matrix as an environmental temporal feature sequence for output.

[0014] The steps for constructing the surface texture evolution feature matrix are as follows: use a pre-trained human factor recognition model to locate the set of human skeleton and facial contour coordinates in the second physical surface reflection wave array; inject a Laplacian noise matrix into the data area covered by the set of human skeleton and facial contour coordinates to perform irreversible blurring; input the blurred second physical surface reflection wave array into a depth variational autoencoder for feature manifold decoding to generate a visual manifold feature matrix as the surface texture evolution feature matrix.

[0015] The key parameters of the Laplacian noise matrix are obtained as follows: The location and scale parameters of the Laplacian noise are declared to be stored in an external structured carrier. A read request is sent to the JSON configuration file in the device file system through the system data interface. The location parameter is extracted to the local running memory and instantiated as the value 0. The scale parameter is extracted to the local running memory and instantiated as 0.45 in the priority range [0,1]. A Laplacian probability density distribution is constructed based on the location and scale parameters. A privacy-covered data region containing the coordinates of the human skeleton and facial contour is obtained, and the horizontal and vertical pixel counts of the privacy-covered data region are counted. A Laplacian noise matrix with spatial dimensions completely consistent with the horizontal and vertical pixel counts is generated using a pseudo-random number generator. The noise value at each position in the Laplacian noise matrix is ​​summed element-wise with the original pixel grayscale value at the corresponding coordinate in the privacy-covered data region. The desensitized second physical surface reflection wave array data is output and input into the depth variational autoencoder.

[0016] The data regions that may contain workers' faces or body shapes are clearly identified and covered with Laplacian noise to completely eliminate the bio-identifiability of pixel clusters. The cleaned pure material appearance data is then input into a variational autoencoder. The variational autoencoder compresses and projects the high-dimensional appearance deformation matrix into a low-dimensional latent space through convolution operations and distributed reparameterization techniques, thereby extracting a visual manifold feature matrix that reflects the true physical deformation law of the material.

[0017] The construction steps of the internal hidden stress prediction sequence are as follows: Extract the environmental temporal feature sequence as the query feature vector, and extract the surface texture evolution feature matrix as the key-value feature vector; calculate the minimum cost mapping path between the cumulative evolution curve corresponding to the query feature vector and the deformation rate slope corresponding to the key-value feature vector based on the differentiable dynamic time warping algorithm, so as to align the query feature vector and the key-value feature vector in time phase; based on the aligned feature vector, use the Pareto multi-objective adaptive gradient weighting algorithm to calculate the gradient norm ratio between the preset elasticity loss function and the network feature generation loss function; dynamically adjust the gradient descent direction of the multimodal cross-attention neural network during backpropagation according to the gradient norm ratio to generate the internal hidden stress prediction sequence.

[0018] The calculation of the gradient norm ratio includes the following steps: Extracting all discrete values ​​contained in the key-value feature vector, and calculating the probability proportion of each discrete value appearing in the key-value feature vector; for each extracted probability proportion, calculating the product of the probability proportion and its own natural logarithm, summing all calculated products, and taking the negative of the summation result, using the final value as the cross-modal information confusion; obtaining a preset confusion mapping decay coefficient, using the cross-modal information confusion as the base input, and performing a negative exponentiation operation on the base input using the confusion mapping decay coefficient, outputting a dynamic confidence penalty weight within the closed interval 0 to 1; extracting a preset elasticity loss function, performing a division operation between the dynamic confidence penalty weight and the preset elasticity loss function, outputting a modified elasticity loss function; inputting the modified elasticity loss function into the Pareto multi-objective adaptive gradient weighting algorithm, and calculating the gradient norm ratio between the modified elasticity loss function and the network feature generation loss function.

[0019] The steps for performing a negative exponential decay mapping operation on the cross-modal information confusion coefficient are as follows: Extract all discrete values ​​contained in the key-value feature vector and calculate the probability proportion of each discrete value appearing in the key-value feature vector; before calculating the logarithm, set a smoothing factor of 1 / 100,000 to prevent zero element overflow; for each probability proportion, sum the probability proportion and the smoothing factor, and perform a natural logarithmic operation on the sum to output the logarithmic median; calculate the product of the probability proportion and the logarithmic median, accumulate all the calculated products, and perform a negative operation on the accumulated result to obtain the cross-modal information confusion; obtain the confusion mapping attenuation coefficient retrieved from an external spreadsheet as... The multiplication scaling parameter is used to multiply the cross-modal information confusion degree with the multiplication scaling parameter to obtain the median value of the scaled confusion degree. The median value is then multiplied by -1 to convert it to the negative quadrant, resulting in the negative power exponent. The power function is solved using the natural constant e as the base and the negative power exponent as the exponent to obtain the initial confidence degree. Boundary truncation is performed: it is determined whether the initial confidence degree is greater than 1. In response to the determination result that it is greater than 1, the final dynamic confidence degree penalty weight is forcibly overwritten and limited to a fixed value of 1. Otherwise, the initial confidence degree value is retained as the dynamic confidence degree penalty weight output to perform division amplification correction on the elasticity loss function.

[0020] The steps for constructing the confusion mapping attenuation coefficient are as follows: Extract a historical building lifecycle image database; perform image clarity filtering on the historical building lifecycle image database to separate a sample set of poor-quality images containing severe dust obstruction, light spot pollution, and extreme exposure; use the same logarithmic product algorithm in S3 to calculate cross-modal information confusion for each sample set of poor-quality images to obtain the corresponding discrete confusion values; perform probability density distribution function fitting on the discrete confusion values ​​to generate a statistical bell-shaped curve of normal distribution of confusion; calculate the critical extreme value of confusion corresponding to the 95th percentile of the statistical bell-shaped curve; obtain the target penalty lower limit constant (in In this embodiment, to ensure that the Pareto gradient does not overflow excessively, the target penalty lower limit constant is set to 0.01. The natural logarithm of the target penalty lower limit constant is calculated to obtain the logarithmic median value. The opposite of the critical extreme value of confusion is extracted, and the logarithmic median value is divided by the opposite. The quotient obtained by the division operation is defined as the final confusion mapping decay coefficient and written into the system configuration table. The decay coefficient is back-derived using pure statistical quantiles, so that as long as the confusion calculated in real time breaks through the 95% limit edge of historical big data, the confidence penalty weight will accurately fall to the bottom of 0.01, achieving the effect of a fully data-driven automatic circuit breaker.

[0021] To address the conflict of physical dimensions in heterogeneous data, the difference between the dimensions of temperature and the grayscale value of reflected waves is eliminated, and a unified mapping is performed to a dimensionless latent space coordinate system with a feature dimension range of [-1,1]. Since heat dissipation is a slow variable that accumulates slowly, while structural deformation is a fast variable that occurs instantaneously, the cumulative evolution curve of the current environmental features is extracted as auxiliary state data. The optimal nonlinear alignment path between the cumulative curve and the slope of rapid deformation is calculated based on the D-DTW algorithm to obtain the time warping correction parameter. The correction parameter is used to adaptively adjust the alignment step of the main process time window of the cross-attention mapping. During the fusion process, the partial differential equation of elasticity is used as a physical regularization constraint, and the balance between physical conservation and manifold deformation sensitivity is achieved through Pareto gradient optimization, generating an internal hidden stress prediction sequence with clear mechanical guidance. The physical laws of mechanics are used as a translator within the neural network to bridge the dimensional gap between environmental data and apparent images. This not only solves the alignment problem of asynchronous heterogeneous data and eliminates mathematical illusions in the model that do not conform to physical common sense, but also makes the predicted stress evolution trajectory have extremely strong physical reliability.

[0022] The construction steps of the fatigue time assessment index are as follows: Obtain the predicted values ​​of the internal hidden stress sequence for the current time step and the predicted values ​​of the internal hidden stress sequence for the previous adjacent time step; perform discrete difference calculation: subtract the predicted values ​​of the internal hidden stress sequence for the previous adjacent time step from the predicted values ​​of the internal hidden stress sequence for the current time step to obtain the discrete numerical change difference; obtain the time span constant between the current time step and the previous adjacent time step (limited to 1 hour in this embodiment); divide the discrete numerical change difference by the time span constant to calculate the first-order discrete difference slope representing the current degree of change, establish an exponential extrapolation mathematical model, and use the preset... The historical safety evolution benchmark rate (i.e., the average value of historical hidden stress that has evolved stably over the past 72 hours) is used as the initial constant term of the exponential extrapolation mathematical model, and the first-order discrete difference slope is substituted as the exponential growth rate factor; the boundary conditions are solved: the dependent variable in the exponential extrapolation mathematical model is set to equal the pre-stored material fracture limit threshold (in this embodiment, the preferred value of the material fracture limit threshold is 0.85, and the value range is [0.70, 0.95]), and the natural logarithm operation is performed in reverse to solve for the corresponding timestamp parameter; the absolute value of the time difference obtained by subtracting the timestamp parameter from the current acquisition time is used as the fatigue time evaluation index and output.

[0023] The average distribution of historical hidden stresses that have evolved stably over the past 72 hours is extracted as a historical safety evolution benchmark rate for comparison. The difference in the predicted stress value and the first derivative within the latest one-hour window are accurately calculated. Based on the degree of deviation between the current sharp increase in the first derivative and the historical stable benchmark rate, the time period required for the internal stress accumulation to reach the material fracture limit is calculated through an extrapolation fitting formula, thus generating a fatigue time assessment index. This completely eliminates the lag in traditional quality monitoring, which only alarms when the static threshold is exceeded (in this embodiment, the static threshold value is limited to a closed interval of 0.80 to 0.95, preferably 0.90). The method based on historical statistics is transformed into extrapolation pre-simulation based on the dynamic slope of calculus, realizing advanced time-series quantitative assessment tens of hours before the hidden defects become apparent.

[0024] The construction of the target anomalous environment location includes the following steps: substituting the node with the largest weight into the graph topology stacked structure of the multimodal cross-attention neural network to perform inverse Jacobian determinant differentiation operation; locating the source network input extreme value coordinates that cause the peak mutation of the internal hidden stress prediction sequence; and mapping the source network input extreme value coordinates back to the three-dimensional spatial reference frame to generate the target anomalous environment location.

[0025] Once the fatigue time assessment index reaches the structural fatigue warning threshold (in this embodiment, the value range of the structural fatigue warning threshold is limited to a closed interval of 24 to 168 hours, with 72 hours being preferred), white-box reverse inference is triggered. Since the network retains the modal correlation attention weight matrix simultaneously when generating stress prediction results, starting from the node with the largest weight that causes the current stress concentration, the inverse Jacobian mathematical derivative is used to find the source of the input layer data that contributes the most to the node along the back propagation path of the neural network, and the location of the target abnormal environment that causes mechanical deformation is calculated. This is then combined with the prediction time and packaged into the final assessment result output. This gives the neural network the ability to trace the source and define responsibility in a white-box manner. It not only provides future risk assessment results, but also more accurately traces back to the initial environmental cause that caused the mechanical evolution, realizing the complete traceability of the quality accident logic chain.

[0026] The specific steps for obtaining the building's hidden quality assessment results are as follows: After outputting the fatigue time assessment index, the structural fatigue warning threshold comparison operation is performed cyclically: the currently output fatigue time assessment index is compared with the preset structural fatigue warning threshold; in response to the comparison judgment result that the fatigue time assessment index is less than or equal to the structural fatigue warning threshold, the reverse tracing operation is directly triggered. The reverse tracing operation includes extracting the modal correlation attention weight matrix of the multimodal cross-attention neural network within the current time period, traversing all correlation weight values ​​in the modal correlation attention weight matrix, locking the row and column index coordinates corresponding to the maximum correlation weight to determine the maximum weight node; performing inverse Jacobian determinant differentiation on the graph topology stacked structure corresponding to the maximum weight node to calculate the input layer extreme coordinates where the contribution to the maximum weight node is greater than a preset contribution threshold (in this embodiment, the contribution threshold is limited to the closed interval 0.50 to 0.80, preferably 0.65); mapping the input layer extreme coordinates back to the three-dimensional spatial reference system to generate the entity spatial coordinates of the target abnormal environment location that causes fatigue state, and encapsulating the output to include fatigue time assessment indicators and building concealment quality assessment results of the target abnormal environment location.

[0027] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a building quality assessment model based on big data, characterized in that, Includes the following steps: S1. Obtain the first environmental state parameter stream within a preset period, perform discrete quantization operation on the first environmental state parameter stream, and extract the environmental time series feature sequence that reflects the spatial temperature and humidity evolution law. S2. Obtain the second physical surface reflection wave array within a preset period, perform a de-identification masking operation based on local differential privacy on the second physical surface reflection wave array, remove pixel clusters containing biological features, and extract the surface texture evolution feature matrix reflecting the nonlinear deformation of the material surface from the masked data. S3. Input the environmental temporal feature sequence and the surface texture evolution feature matrix into a pre-trained multimodal cross-attention neural network. The multimodal cross-attention neural network uses a preset elasticity loss function to perform physical dimension alignment and manifold mapping on the environmental temporal feature sequence and the surface texture evolution feature matrix, calculates the internal hidden stress prediction sequence used to characterize the internal stress state of the material, and simultaneously extracts the modal correlation attention weight matrix generated by the multimodal cross-attention neural network when performing manifold mapping. S4. Extract the numerical variation difference of the internal hidden stress prediction sequence at adjacent time steps, compare the numerical variation difference with the historical safety evolution benchmark rate, and calculate the fatigue time assessment index for the future generation of microcracks inside the structure. S5. In response to the fatigue time assessment index meeting the preset structural fatigue early warning threshold, the maximum weight node is selected based on the modal association attention weight matrix. The target abnormal environment location that caused the fatigue state is traced back according to the maximum weight node, and the building concealed quality assessment result containing the fatigue time assessment index and the target abnormal environment location is output.

2. The method for constructing a building quality assessment model based on big data according to claim 1, characterized in that: The steps for constructing the environmental temporal feature sequence are as follows: map the first environmental state parameter stream into graph topology data containing feature vectors of multiple spatial nodes; input the graph topology data into a graph convolutional long short-term memory network; calculate the numerical gradient changes of the first environmental state parameter stream among various spatial nodes to generate a spatial microclimate dissipation coefficient matrix as the environmental temporal feature sequence.

3. The method for constructing a building quality assessment model based on big data according to claim 1, characterized in that: The steps for constructing the surface texture evolution feature matrix are as follows: use a pre-trained human factor recognition model to locate the set of human skeleton and facial contour coordinates in the second physical surface reflection wave array; inject a Laplacian noise matrix into the data area covered by the set of human skeleton and facial contour coordinates to perform irreversible blurring; input the blurred second physical surface reflection wave array into a depth variational autoencoder for feature manifold decoding to generate a visual manifold feature matrix as the surface texture evolution feature matrix.

4. The method for constructing a building quality assessment model based on big data according to claim 1, characterized in that: The construction steps of the internal hidden stress prediction sequence are as follows: Extract the environmental temporal feature sequence as the query feature vector, and extract the surface texture evolution feature matrix as the key-value feature vector; calculate the minimum cost mapping path between the cumulative evolution curve corresponding to the query feature vector and the deformation rate slope corresponding to the key-value feature vector based on the differentiable dynamic time warping algorithm, so as to align the query feature vector and the key-value feature vector in time phase; based on the aligned feature vector, use the Pareto multi-objective adaptive gradient weighting algorithm to calculate the gradient norm ratio between the preset elasticity loss function and the network feature generation loss function; dynamically adjust the gradient descent direction of the multimodal cross-attention neural network during backpropagation according to the gradient norm ratio to generate the internal hidden stress prediction sequence.

5. The method for constructing a building quality assessment model based on big data according to claim 4, characterized in that: The calculation of the gradient norm ratio includes the following steps: Extract all discrete values ​​contained in the key-value feature vector and calculate the probability proportion of each discrete value appearing in the key-value feature vector. For each extracted probability proportion, calculate the product of the probability proportion and its natural logarithm. Accumulate all calculated products and take the negative of the accumulated result. Use the final value as the cross-modal information confusion. Obtain a preset confusion mapping decay coefficient. Use the cross-modal information confusion as the base input and perform a negative exponentiation operation on the base input using the confusion mapping decay coefficient to output the dynamic confidence penalty weight. Extract a preset elasticity loss function. Divide the dynamic confidence penalty weight and the preset elasticity loss function to output the modified elasticity loss function. Input the modified elasticity loss function into the Pareto multi-objective adaptive gradient weighting algorithm and calculate the gradient norm ratio between the modified elasticity loss function and the network feature generation loss function.

6. The method for constructing a building quality assessment model based on big data according to claim 5, characterized in that: The steps for performing a negative exponential decay mapping operation on the cross-modal information confusion coefficient are as follows: Extract all discrete values ​​contained in the key-value feature vector, and calculate the probability proportion of each discrete value appearing in the key-value feature vector; before calculating the logarithm, set a smoothing term factor to prevent zero element overflow; for each probability proportion, sum the probability proportion with the smoothing term factor, and perform a natural logarithmic operation on the summation result to output the intermediate logarithm value. The product of the probability ratio and the logarithmic median is calculated. All calculated products are summed, and the sum is inversely divided to obtain the cross-modal information confusion. The confusion mapping attenuation coefficient retrieved from an external spreadsheet is used as the multiplication ratio parameter. The cross-modal information confusion and the multiplication ratio parameter are multiplied to obtain the scaled confusion median. The median is multiplied by -1 to convert it to the negative quadrant to obtain the negative power exponent. The power function is solved using the natural constant e as the base and the negative power exponent as the exponent to obtain the initial confidence. Boundary truncation is performed: it is determined whether the initial confidence is greater than 1. In response to the determination result of being greater than 1, the final dynamic confidence penalty weight is forcibly overwritten and limited to a fixed value of 1. Otherwise, the initial confidence value is retained as the dynamic confidence penalty weight output to perform division amplification correction on the elasticity loss function.

7. The method for constructing a building quality assessment model based on big data according to claim 6, characterized in that: The steps for constructing the confusion mapping attenuation coefficient are as follows: Extract a historical building lifecycle image database; perform clarity filtering on the historical building lifecycle image database to separate a sample set of poor-quality appearance images containing severe dust obstruction, light spot pollution, and extreme exposure; use the same logarithmic product algorithm in S3 to calculate cross-modal information confusion for each sample set of poor-quality appearance images to obtain the corresponding discrete confusion values; perform probability density distribution function fitting on the discrete confusion values ​​to generate a statistical bell-shaped curve of normal distribution of confusion; calculate the critical extreme value of confusion corresponding to the 95th percentile of the statistical bell-shaped curve; obtain the target penalty lower limit constant; calculate the natural logarithm of the target penalty lower limit constant to obtain the logarithmic median value; Extract the opposite of the critical extreme value of confusion, divide the logarithmic median value by the opposite, and define the quotient obtained from the division operation as the final confusion mapping attenuation coefficient.

8. The method for constructing a building quality assessment model based on big data according to claim 1, characterized in that: The construction steps of the fatigue time assessment index are as follows:

1. Obtain the predicted values ​​of the internal hidden stress sequence at the current time step and the predicted values ​​of the internal hidden stress sequence at the previous adjacent time step; 2. Perform discrete difference calculation: Subtract the predicted values ​​of the internal hidden stress sequence at the previous adjacent time step from the predicted values ​​of the internal hidden stress sequence at the current time step to obtain the discrete numerical change difference; 3. Obtain the time span constant between the current time step and the previous adjacent time step; 4. Divide the discrete numerical change difference by the time span constant to calculate the first-order discrete difference slope representing the current degree of change, establish an exponential extrapolation mathematical model, use the preset historical safety evolution benchmark rate as the initial constant term of the exponential extrapolation mathematical model, and substitute the first-order discrete difference slope as the exponential growth rate factor; 5. Solve the boundary conditions: Set the dependent variable in the exponential extrapolation mathematical model equal to the pre-stored material fracture limit threshold, perform the natural logarithm operation in reverse, and solve for the corresponding timestamp parameter in reverse; 6. Subtract the current acquisition time from the timestamp parameter, and the absolute value of the time difference is output as the fatigue time assessment index.

9. The method for constructing a building quality assessment model based on big data according to claim 1, characterized in that: The construction of the target anomalous environment location includes the following steps: substituting the node with the largest weight into the graph topology stacked structure of the multimodal cross-attention neural network to perform inverse Jacobian determinant differentiation operation; locating the source network input extreme value coordinates that cause the peak mutation of the internal hidden stress prediction sequence; and mapping the source network input extreme value coordinates back to the three-dimensional spatial reference frame to generate the target anomalous environment location.

10. The method for constructing a building quality assessment model based on big data according to claim 1, characterized in that: The specific steps for obtaining the building's hidden quality assessment results are as follows: After outputting the fatigue time assessment index, the structural fatigue warning threshold comparison operation is performed cyclically: the currently output fatigue time assessment index is compared with the preset structural fatigue warning threshold. In response to the comparison and judgment result that the fatigue time assessment index is less than or equal to the structural fatigue early warning threshold, a reverse tracing operation is directly triggered. The reverse tracing operation includes extracting the modal correlation attention weight matrix of the multimodal cross-attention neural network within the current time period, traversing all correlation weight values ​​in the modal correlation attention weight matrix, locking the row and column index coordinates corresponding to the maximum correlation weight to determine the maximum weight node; performing inverse Jacobian determinant differentiation on the graph topology stacked structure corresponding to the maximum weight node to calculate the input layer extreme coordinates where the contribution to the maximum weight node is greater than a preset contribution threshold; mapping the input layer extreme coordinates back to the three-dimensional spatial reference system to generate the entity spatial coordinates of the target abnormal environment location that causes fatigue state, and encapsulating the output to include fatigue time assessment indicators and the building concealment quality assessment results of the target abnormal environment location.

Images