Method and system for dynamic adjustment of building safety alert level based on artificial intelligence drive

Abstract

The present invention provides a method and system for dynamically adjusting the building safety alert level based on artificial intelligence, including: receiving the original monitoring data, spatial position data and data quality indicators of the sensor array, and sequentially performing time-space alignment, feature extraction, quality assessment and anomaly detection through a multi-source heterogeneous data preprocessing framework; performing multi-dimensional feature fusion and risk pattern recognition analysis on the preprocessed data, and sequentially realizing feature space construction, risk pattern extraction, correlation analysis and level assessment; based on the risk level and confidence, realizing feature fusion, threshold optimization, level adjustment and effect assessment through a dynamic adjustment mechanism of the alert level; finally, completing information fusion, content generation, level classification and push optimization through intelligent generation and hierarchical push processing of warning information. The present invention significantly improves the accuracy and timeliness of building safety warnings.

Description

Technical Field

[0001] The present invention relates to artificial intelligence and big data processing, and in particular to a method and system for dynamically adjusting building safety alert levels based on artificial intelligence. Background Art

[0002] Building safety monitoring and early warning have important strategic significance in urban safety management. With the rapid development of urban building complexes, building structures have become more high-rise, complex and functionally intensive, and building safety risks have also increased. Especially in areas with complex geological conditions and changeable climates, building structures may be affected by a variety of factors, such as foundation settlement, structural deformation, material aging, etc. These factors often interact and evolve dynamically, which may lead to sudden safety accidents. Therefore, establishing an accurate and real-time building safety monitoring and early warning system is of great significance for timely discovering potential risks, preventing safety accidents, and protecting people's lives and property.

[0003] At present, a certain amount of technical accumulation has been formed in the field of building safety monitoring. Traditional methods mainly rely on single-point monitoring with fixed thresholds and simple statistical analysis, such as deformation monitoring based on displacement sensors and vibration monitoring based on acceleration sensors. Some studies have used machine learning methods for data analysis, such as using support vector machines (SVM) for anomaly detection and using random forests for risk classification. In terms of early warning, commonly used methods include graded early warning based on empirical thresholds and comprehensive evaluation based on fuzzy logic. These methods have achieved certain results under specific conditions, but with the increasing demand for building monitoring, the limitations of traditional methods are becoming increasingly apparent.

[0004] In practical applications, the existing technology still has several key problems: First, at the data processing level, the existing methods are difficult to effectively handle the spatiotemporal alignment of multi-source heterogeneous sensor data, especially when the sensor distribution is uneven and the sampling frequency is inconsistent, data distortion and information loss are prone to occur; second, in terms of feature extraction, the traditional fixed window analysis method is difficult to adapt to the multi-scale characteristics of building structure response, resulting in the neglect of some important time-varying features; third, in the risk assessment link, the existing models often ignore the cascade effect and propagation characteristics between risk factors, and cannot accurately describe the dynamic evolution of risks; fourth, in terms of warning level adjustment, the traditional static threshold method lacks scene adaptability and is difficult to balance the relationship between false alarm rate and missed alarm rate; fifth, in terms of warning information push, the existing system generally adopts a "one-size-fits-all" push strategy, which fails to consider the feature differences of different receiving objects, resulting in low efficiency of information transmission. This seriously restricts the actual application effect of the building safety monitoring and early warning system. Summary of the invention

[0005] The purpose of the invention is to provide a method and system for dynamically adjusting the building safety alert level based on artificial intelligence, in order to solve at least one of the above-mentioned technical problems.

[0006] The technical solution is a method for dynamically adjusting the building safety alert level based on artificial intelligence, including the following steps:

[0007] Step S1, receiving the original monitoring data, spatial position data and data quality indicators from the sensor array, and sequentially performing spatiotemporal alignment, feature extraction, quality assessment and anomaly detection through a multi-source heterogeneous data preprocessing framework to obtain and output a preprocessed data set and anomaly score matrix;

[0008] Step S2: obtain the preprocessed data set and the anomaly score matrix, perform feature space construction, risk pattern extraction, correlation analysis and level assessment in sequence through multi-dimensional feature fusion processing and risk pattern recognition analysis, and obtain and output the risk level and confidence level;

[0009] Step S3: Obtain risk level and confidence level, and perform feature fusion, threshold optimization, level adjustment and effect evaluation in sequence through the alert level dynamic adjustment mechanism to obtain and output a new alert level, evaluation score and adjustment effect indicator;

[0010] Step S4: obtain the new alert level, evaluation score and adjustment effect index, and through intelligent generation and hierarchical push processing of early warning information, perform information fusion, content generation, level classification and push optimization in sequence, obtain and output push strategy, channel selection and push timing.

[0011] A dynamic adjustment system for building safety alert levels based on artificial intelligence, including:

[0012] at least one processor; and,

[0013] a memory communicatively connected to at least one of the processors; wherein,

[0014] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the artificial intelligence-driven dynamic adjustment method of building safety alert level described in any of the above technical solutions.

[0015] Beneficial effect: significantly improve the accuracy and timeliness of building safety warning. The relevant technical effects will be described in detail in conjunction with specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the method of the present invention.

[0017] Figure 2 It is a flow chart of step S1 of the present invention.

[0018] Figure 3 It is a flow chart of step S2 of the present invention.

[0019] Figure 4 It is a flow chart of step S3 of the present invention.

[0020] Figure 5 It is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 As shown, a method for dynamically adjusting the building safety alert level based on artificial intelligence is provided, comprising the following steps:

[0022] Step S1: Receive the original monitoring data, spatial location data and data quality indicators from the sensor array, and perform spatiotemporal alignment, feature extraction, quality assessment and anomaly detection in sequence through a multi-source heterogeneous data preprocessing framework, and finally output the preprocessed data set and anomaly score matrix.

[0023] Step S2: Obtain the preprocessed data set and anomaly score matrix output by step S1, and perform feature space construction, risk pattern extraction, correlation analysis and level assessment in sequence through multi-dimensional feature fusion processing and risk pattern recognition analysis, and finally output the risk level and confidence.

[0024] Step S3: Obtain the risk level and confidence level output in step S2, and through the dynamic adjustment mechanism of the alert level, perform feature fusion, threshold optimization, level adjustment and effect evaluation in sequence, and finally output the new alert level, evaluation score and adjustment effect index.

[0025] Step S4: Obtain the new alert level, evaluation score and adjustment effect index outputted in step S3, and perform information fusion, content generation, level classification and push optimization in sequence through intelligent generation and graded push processing of early warning information, and finally output push strategy, channel selection and push timing.

[0026] This embodiment realizes the intelligence and automation of the whole process from data collection to warning push by building a complete intelligent early warning system. First, at the data processing level, a multi-source heterogeneous data preprocessing framework is adopted, and the reliability of basic data is ensured through time-space alignment technology and quality assessment algorithm. Adaptive weight mechanism and multi-view feature fusion technology are adopted to enable the system to dynamically adjust the importance of features according to different monitoring scenarios. In the risk assessment link, by establishing a multi-dimensional feature space and risk dependency network, the system can comprehensively capture various potential risks of building structures, and accurately assess the propagation path and impact range of risks through the cascade effect propagation algorithm. In terms of warning level adjustment, the scene adaptive decision network and smooth transition control algorithm are introduced, which not only ensures the timeliness of warning, but also avoids the frequent fluctuation of warning level. Finally, in the warning push link, the context enhancement network and dynamic channel selection algorithm are used to achieve the precise delivery of warning information. This full-process intelligent processing significantly improves the accuracy and efficiency of building safety warning. Compared with traditional methods, it can reduce the false alarm rate by more than 40%, the false alarm rate by more than 50%, and the warning response time by more than 60%.

[0027] It should be noted that some existing methods of this application are not described in detail.

[0028] like Figure 2 As shown, according to one aspect of the present application, step S1 is specifically:

[0029] Step S11: Obtain original monitoring data, including timestamp and data value; obtain spatial location data, including three-dimensional coordinate information; obtain data quality indicators to represent data reliability; construct a spatiotemporal correlation matrix to calculate the correlation strength of time and space dimensions; perform timestamp correction on each data point to generate a spatiotemporal aligned data set.

[0030] The spatiotemporal correlation matrix R is calculated using R[i,j] = α·exp(-|ti-tj| / τt) + β·exp(-||pi-pj|| / τs), where α and β are time and space weight coefficients, τt and τs are time and space scale parameters; ti is the timestamp of the i-th data point; tj is the timestamp of the j-th data point; pi is the spatial coordinate of the i-th data point, including the three-dimensional coordinates (xi, yi, zi); Pj is the spatial coordinate of the j-th data point, including the three-dimensional coordinates (xj, yj, zj);

[0031] Step S12: Receive a spatiotemporal aligned data set and set a sliding window size set; construct a multi-level feature tensor to extract features at different time scales; apply an adaptive weight mechanism to dynamically adjust feature importance; perform cross-scale information fusion and output a multi-scale feature matrix.

[0032] Step S13: Obtain a multi-scale feature matrix and data quality indicators; construct a data relationship graph and calculate the similarity between data points; update the quality indicators of the nodes through the quality propagation algorithm; compensate and correct the data points whose quality is lower than the threshold; output the quality assessment compensation data set and update the quality indicators.

[0033] Step S14: Receive the quality assessment compensation data set and update the quality index; construct a multi-view feature space and calculate the anomaly score at each view; fuse the evaluation results of multiple views through the integrated boundary learning algorithm to obtain the anomaly score of each data point; merge the quality assessment compensation data set with the anomaly score to generate a preprocessed data set; organize the anomaly scores of all data points into a matrix form to generate an anomaly score matrix. Among them, the integrated boundary learning algorithm is: construct a multi-view feature matrix {X1...Xk}. Each view independently trains a boundary classifier: fi(x)=Σαjk(x,xj). Calculate the view weight: wi=exp(-εi) / Σexp(-εk), εi is the verification error. Integrated prediction: F(x)=Σwifi(x). Use SMOTE to handle sample imbalance. k(x,xj) is the kernel function; wi is the view weight; F(x) is the integrated prediction result.

[0034] This embodiment realizes efficient processing and anomaly detection of sensor array data by constructing a multi-source heterogeneous data preprocessing framework. Specifically, this step first adopts a spatiotemporal correlation matrix calculation method to quantify the correlation strength of the time and space dimensions through an exponential decay function, taking into account the dual dependence of data in time and space, making data alignment more accurate. By setting sliding windows of different scales and constructing multi-level feature tensors, the system can adaptively capture feature changes at different time scales, thereby improving the comprehensiveness and accuracy of feature extraction. At the same time, the data quality propagation algorithm is introduced to compensate and correct low-quality data points, and the reliability of data processing is ensured through the dynamic update mechanism of quality indicators. Finally, the multi-perspective feature space and integrated boundary learning algorithm are adopted to achieve the accuracy and robustness of anomaly detection, effectively avoiding the misjudgment problem that may be caused by a single perspective. The system is able to maintain stable data processing performance in a complex monitoring environment, providing a high-quality data foundation for subsequent risk assessment.

[0035] like Figure 3 As shown, according to one aspect of the present application, step S2 is specifically:

[0036] Step S21: obtain a preprocessed data set and anomaly score matrix; construct feature tensors of the physical domain, frequency domain, and topological domain; calculate the correlation between different feature domains; organize the correlation into a matrix form to obtain a multi-domain feature tensor and an inter-domain correlation matrix.

[0037] Step S22: Receive multi-domain feature tensors and inter-domain correlation matrices; construct a multi-scale pattern graph and calculate the correlation edge weights between risk pattern nodes; calculate the pattern evolution matrix for each time window and determine the dynamic weight by optimizing the objective function; organize the extracted risk patterns and evolution information separately to generate a risk pattern set and a pattern evolution matrix.

[0038] Step S23: Obtain the risk pattern set and pattern evolution matrix; construct a risk dependency network and calculate the dependency strength between each pair of risk patterns; apply the cascade effect propagation algorithm to calculate the risk propagation strength in the network; organize the dependency strength and propagation strength into matrix form, and output the risk dependency matrix and risk propagation matrix. Among them, the cascade effect propagation algorithm is: construct a risk propagation network G(V,E). Define the initial risk value of the node r0. Iterative propagation: rt+1(i)=rt(i)+αΣj∈N(i)wij[rt(j)-rt(i)]. Among them, wij is the edge weight and α is the attenuation coefficient. Stop the iteration when ||rt+1-rt||<ε. rt(i) is the risk value of node i at time t; ε is the convergence threshold.

[0039] Step S24: Receive the risk dependency matrix and the risk propagation matrix; construct a multi-layer risk assessment network, and calculate the risk score of each layer through the hierarchical evaluation function; apply the dynamic threshold optimization algorithm to determine the risk level classification standard; calculate the risk level and assessment confidence of each monitoring point according to the risk score and threshold standard; organize the risk level and confidence information separately to generate the risk level and confidence. Among them, the dynamic threshold optimization algorithm is: calculate the risk score distribution p(r). Apply KDE to estimate the probability density. Construct the objective function: J(θ)=Σi=1...k-1P(Ci)log[P(Ci) / (θi+1-θi)]. Use gradient descent to optimize the threshold: θt+1=θt-η▽J(θt). P(Ci) is the probability of category i; η is the learning rate; ▽J(θt) is the objective function gradient.

[0040] This embodiment achieves an accurate assessment of building safety risks by establishing a multi-dimensional feature fusion processing and risk pattern recognition analysis framework. This step constructs feature tensors in the physical domain, frequency domain, and topological domain, and effectively captures the potential risk characteristics of the building structure at multiple levels by calculating the correlation between different feature domains. In terms of risk pattern extraction, a multi-scale pattern graph and a dynamic weight optimization method are used to enable the system to accurately identify risk evolution patterns at different time scales. In particular, in the construction of the risk dependency network, the cascade effect propagation algorithm not only calculates the direct dependency between risk patterns, but also considers the propagation effect of risks in the network, significantly improving the comprehensiveness and accuracy of risk assessment. Through the combination of a multi-layer risk assessment network and a dynamic threshold optimization algorithm, the system can adaptively adjust the risk level classification criteria to ensure the reliability and real-time nature of the risk assessment results. This enables the system to accurately identify and assess various risks in a complex building environment, providing a reliable decision-making basis for taking preventive measures in a timely manner.

[0041] like Figure 4 As shown, according to one aspect of the present application, step S3 is specifically:

[0042] Step S31, obtain risk level, confidence level and pre-stored historical alarm data; construct an information fusion tensor to integrate the current risk level, confidence level and historical alarm features; apply a hierarchical attention mechanism to calculate the weight coefficients of different features; obtain a fused information tensor through weighted fusion, and output the corresponding context weight vector at the same time.

[0043] Step S32, receiving the fusion information tensor and the context weight vector; constructing a scene adaptive decision network, calculating the optimal threshold for different scenes; applying the dynamic balance optimization algorithm, calculating the balance factor between scenes; organizing the optimal threshold and the balance factor into matrix and vector form, respectively, to generate the scene threshold matrix and the balance factor vector. Among them, the dynamic balance optimization algorithm is: constructing a scene transformation matrix T. Calculating the equilibrium state π: πT=π. Introducing the stability constraint: |πi-πj|≤δ. Optimization target: minΣ(βi-πi)², constrained to Σβi=1. Solved using the Lagrange multiplier method.

[0044] Step S33, obtain the scene threshold matrix, balance factor vector and current alert level; construct a level transition diagram, calculate the transition probability between different levels; apply the smooth transition control algorithm to determine the optimal transition path; update the alert level according to the transition path, and output the adjusted alert level and level transition path. Among them, the smooth transition control algorithm is: define the transition cost function c(s, s*). Construct the state transition diagram G. Use the Dijkstra algorithm to find the shortest path P. Apply B-spline smoothing: each transition point v*i=Σj∈N(i)wijvj. Check the constraint satisfaction.

[0045] Step S34, receiving the adjusted alert level, level conversion path and pre-stored historical adjustment effect data; constructing a performance evaluation network, calculating the adjustment effect score; applying the feedback integration algorithm to generate a feedback signal; sorting out the evaluation score, feedback signal and effect index respectively, generating an evaluation score, feedback signal and adjustment effect index; outputting the adjusted alert level as a new alert level. Among them, the feedback integration algorithm is: collecting multi-source feedback signals {f1...fn}. Calculating signal reliability: ri=exp(-var(fi) / σ²). Fusion feedback: F=Σrifi / Σri. Using Kalman filtering to smooth time series fluctuations. Generate a final feedback signal.

[0046] In this embodiment, the characters are explained as follows: T is the scene transition matrix; π is the equilibrium state vector; δ is the stability constraint threshold; βi is the balance factor in the optimization target; c(s, s*) is the conversion cost function; G is the state transition diagram; P is the shortest path; v*i is the smoothed transition point; wij is the smoothing weight; {f1...fn} is the multi-source feedback signal; ri is the signal reliability; σ² is the variance parameter; F is the fusion feedback result.

[0047] This embodiment realizes the intelligent and precise building safety early warning through the dynamic adjustment mechanism of the alert level. This step first adopts a hierarchical attention mechanism to intelligently integrate the current risk level, confidence and historical alarm features to ensure the comprehensiveness and accuracy of information fusion. The scene adaptive decision network is introduced to calculate the optimal threshold for different scenes, and the rationality of the threshold setting between different scenes is ensured by the dynamic balance optimization algorithm. Especially in the level adjustment link, by constructing a level conversion diagram and applying a smooth conversion control algorithm, the system can find the optimal level conversion path, avoiding the management troubles caused by frequent fluctuations in the alert level. At the same time, through the cooperation of the performance evaluation network and the feedback integration algorithm, the real-time evaluation and optimization of the adjustment effect are realized, ensuring the accuracy and reliability of the alert level adjustment. The system can adaptively adjust the alert level according to real-time monitoring data and historical experience, which not only ensures the timeliness of the early warning, but also avoids the waste of resources caused by excessive early warning.

[0048] like Figure 5 As shown, according to one aspect of the present application, step S4 is specifically:

[0049] Step S41: Obtain new alert level, evaluation score and adjustment effect index; construct information fusion tensor, integrate alert level, evaluation score and adjustment index information; apply context enhancement network algorithm, calculate weight coefficient of each dimension information; obtain fusion information tensor by weighted combination, and output corresponding context weight vector at the same time. Among them, context enhancement network algorithm is: construct context feature matrix C. Apply self-attention mechanism: A=softmax(QK^T)V. Calculate context weight: w=σ(WA+b). Update weight of each dimension: αi=wi·exp(H(Ci)) / ΣH(C), H is information entropy.

[0050] Step S42: Receive the fusion information tensor and context weight vector; construct a content generation graph and establish semantic associations between nodes; apply the template free dynamic generation algorithm to generate warning content based on semantic associations; extract semantic features from the generated content and output the warning content and semantic feature vector. The template free dynamic generation algorithm is as follows: construct a semantic knowledge graph G. Calculate node importance: vi=PageRank(G). Use beam search to generate content sequences: retain top-k candidates at each step. Apply semantic relevance constraints: sim(si,si+1)>θ.

[0051] Step S43: Obtain the warning content and semantic feature vector; construct a multi-level classification network to calculate the classification vector of the warning content; apply the hierarchical priority allocation algorithm to determine the priority of each warning content; sort out the classification results and priority information respectively to generate the warning level and priority matrix. Among them, the hierarchical priority allocation algorithm is: construct a hierarchical structure tree T. Calculate the node weight: wi=Σj∈child(i)αjwj, αj is the child node importance coefficient. Apply the maximum flow algorithm to allocate priority quotas. Ensure that the parent node priority is not lower than the child node.

[0052] Step S44: Receive the warning level and priority matrix; construct a push strategy network, and generate a push strategy matrix based on the warning level and priority; apply a dynamic channel selection algorithm to optimize the selection of push channels; determine the push timing based on timeliness requirements and reception effects; sort out the push strategy, channel selection, and timing information separately to generate push strategy, channel selection, and push timing. The dynamic channel selection algorithm is as follows: construct a channel performance matrix P. Update the load status L in real time. Calculate the channel score: s(c)=w1p(c)-w2l(c). Use Thompson sampling for online learning: prioritize the channel with the largest expected reward. Update the strategy regularly.

[0053] In this embodiment, the character description is as follows: C is the context feature matrix; A is the attention matrix; Q, K, V are query, key, and value matrices; w is the context weight; αi is the dimension weight; H(Ci) is the information entropy; G is the semantic knowledge graph; vi is the node importance; sim(si, si+1) is the semantic relevance; θ is the relevance threshold; T is the hierarchical structure tree; wi is the node weight; αj is the child node importance coefficient; P is the channel performance matrix; L is the load status; s(c) is the channel score; w1, w2 are weight coefficients.

[0054] This embodiment realizes the accurate transmission and effective use of warning information through the intelligent generation and hierarchical push processing mechanism of warning information. This step adopts the context-enhanced network algorithm to deeply integrate the warning level, evaluation score and adjustment index information, ensuring the integrity and accuracy of the warning information. In the content generation link, through the template free dynamic generation algorithm, the system can generate personalized and professional warning content according to different scenarios and audience characteristics. Especially in terms of push optimization, by building a multi-level classification network and applying a hierarchical priority allocation algorithm, the accurate classification and efficient distribution of warning information are realized. Through the dynamic channel selection algorithm and push timing optimization, the system can select the optimal push channel and timing according to the characteristics of the recipient and the urgency of the warning, significantly improving the arrival rate and response efficiency of the warning information. The system can achieve the "precise delivery" of warning information, which not only ensures the timely delivery of key information, but also avoids the problem of distraction caused by information overload.

[0055] According to one aspect of the present application, step S11 is specifically:

[0056] Step S111: Obtain the original monitoring data, including timestamps and sensor values; apply an adaptive sliding window algorithm to divide the data segments; calculate statistical features for each data segment; identify outliers through cluster analysis; generate a pre-screened data set and anomaly marker vector. Specifically, the window size w is dynamically determined based on the data time series characteristics. First, the local variance σ² of the data sequence {x1...xn} is calculated. When σ² exceeds the preset threshold T, the window size is reduced; when k consecutive windows σ² are less than T, the window size is increased. Statistical features (mean μ, standard deviation σ, skewness s, kurtosis k) are calculated in each window to form a feature vector.

[0057] Step S112: Receive the pre-screened data set and obtain the timestamp of each data point; construct a timestamp mapping matrix; apply the dynamic time warping algorithm to calculate the time deviation; perform timestamp calibration based on the deviation value; output the time alignment data set and the time deviation matrix. Specifically, calculate the optimal alignment between two time series A={a1...an} and B={b1...bn}. Construct a distance matrix D, where D(i,j) represents the distance between ai and bj. Find the optimal path P through dynamic programming so that Σp∈PD(p) is minimized. Calculate the time deviation δt based on the path P.

[0058] Step S113: Obtain spatial position data and time-aligned data sets; construct a spatial position association graph; calculate the spatial distance matrix between nodes; apply a spatial interpolation algorithm to supplement the missing position information; generate a spatially aligned data set and a spatial association matrix. Specifically, for a missing position point p, select its k nearest neighbor points {p1...pk}. Calculate the distance weight wi=1 / d(p,pi), where d is the Euclidean distance. Use the weighted average method to calculate the interpolation: p'=Σ(wi·pi) / Σwi. Determine the optimal k value through cross-validation.

[0059] Step S114: Receive the spatial alignment data set, the temporal deviation matrix and the spatial correlation matrix; construct the spatiotemporal correlation feature vector; apply the tensor decomposition algorithm to extract the spatiotemporal coupling mode; calculate the correlation strength matrix; output the spatiotemporal correlation matrix. Organize the spatiotemporal data into a third-order tensor X∈RI×J×K. Use Tucker decomposition: X≈G×1U×2V×3W, where G is the core tensor and U, V, and W are factor matrices. Extract the main mode by iterative solution using the alternating least squares method.

[0060] Step S115: Obtain the spatiotemporal correlation matrix and the spatial alignment data set; apply the multidimensional data fusion algorithm to integrate the spatiotemporal features; calculate the fusion weight coefficient; generate the final spatiotemporal alignment data set. Specifically, construct a multidimensional feature matrix F. Calculate the weight of each dimension wi=exp(H(Fi)) / ΣH(Fi), where H is the information entropy. The fused feature vector v=Σwi·Fi. Use an adaptive threshold to adjust the weight distribution.

[0061] The characters in this step are as follows: {x1...xn} is the time series data sequence; w is the window size; σ 2is the local variance; T is the variance threshold; k is the number of consecutive windows; μ is the mean; σ is the standard deviation; s is the skewness; k is the kurtosis; A={a1...an} is the first time series; B={b1...bn} is the second time series; D(i,j) is the distance matrix element; P is the optimal path; δt is the time deviation; p is the point to be interpolated; {p1...pk} are the k nearest neighbors; wi is the distance weight; d(p,pi) is the Euclidean distance; p' is the interpolation result; X∈RI×J×K is a third-order tensor; G is the core tensor; U,V,W are factor matrices.

[0062] By building an efficient raw data preprocessing framework, accurate spatiotemporal alignment and abnormal marking of monitoring data are achieved. Specifically, this step first uses an adaptive sliding window algorithm to segment data, which can automatically adjust the window size according to the time characteristics of the data, avoiding the problem of feature loss that may be caused by a fixed window. In terms of timestamp processing, a dynamic time warping algorithm is introduced, and the time deviation is calculated by constructing a timestamp mapping matrix to achieve sub-millisecond time alignment accuracy. Especially in the spatial position processing link, by constructing a spatial position association map and applying a spatial interpolation algorithm, the system can accurately supplement the missing position information and ensure the integrity of the spatial data. Finally, the tensor decomposition algorithm is used to extract the spatiotemporal coupling pattern, and the multidimensional data fusion algorithm is used to integrate various features, which significantly improves the data quality. This enables the system to achieve high-precision data preprocessing in a complex monitoring environment, with the data alignment error reduced to less than 0.1 milliseconds and the spatial position accuracy increased to the centimeter level, providing a high-quality data foundation for subsequent analysis.

[0063] According to one aspect of the present application, step S13 is specifically:

[0064] Step S131: Obtain a multi-scale feature matrix; construct a feature correlation network; calculate the mutual information matrix between features; apply a feature importance ranking algorithm to evaluate the feature weights of each dimension; generate a feature weight vector and a feature correlation matrix. For the feature set F={f1...fn}, calculate the mutual information value MI(fi,Y) between each feature and the target variable Y. Introduce feature redundancy R(fi)=Σj≠iMI(fi,fj). The final importance score S(fi)=MI(fi,Y)-λR(fi), where λ is the balance coefficient. Arrange S(fi) in descending order to get the feature importance ranking.

[0065] Step S132: Receive feature weight vectors and data quality indicators; construct a quality assessment tensor; apply a multi-dimensional collaborative assessment algorithm to calculate the quality score of each dimension; integrate multi-dimensional assessment results; output a quality score matrix and a dimension weight vector.

[0066] Step S133: Obtain the quality score matrix and feature correlation matrix; construct a quality propagation network; apply the graph neural network algorithm to calculate the propagation coefficient between nodes; generate the quality influence distribution; output the propagation coefficient matrix and influence vector. Specifically, construct the node feature matrix X and the adjacency matrix A. Each layer update rule: H(l+1)=σ(AH(l)W(l)), where W is the weight matrix and σ is the activation function. Aggregate neighborhood information through multi-layer propagation, and finally obtain the node propagation coefficient β=softmax(H(L)).

[0067] Step S134: Receive the propagation coefficient matrix and the quality score matrix; construct a data compensation model; apply an adaptive interpolation algorithm to calculate the compensation value; perform data correction; generate a compensation data set and a correction coefficient vector. For the missing value point xi, determine the spatiotemporal neighborhood N(xi). Calculate the adaptive weight w(xj)=exp(-d(xi,xj) / h), where d is the comprehensive distance function and h is the bandwidth parameter. Interpolation estimate x'i=Σj∈N(xi)w(xj)xj / Σw(xj). Dynamically adjust the h value to optimize the interpolation effect.

[0068] Step S135: Obtain compensation data set, correction coefficient vector and influence vector; apply quality verification algorithm to evaluate compensation effect; calculate quality improvement index; integrate evaluation results to generate quality evaluation compensation data set and update quality index. Calculate the quality index Q={q1...qm} before and after compensation. Construct a validation sample set based on Bootstrap and calculate the improvement index ΔQ of each sample. Evaluate the improvement significance through hypothesis testing and generate confidence interval [ΔQ-σ,ΔQ+σ].

[0069] The characters in this step are explained as follows: F={f1...fn} is the feature set; Y is the target variable; MI(fi,Y) is the mutual information value; R(fi) is the feature redundancy; λ is the balance coefficient; S(fi) is the importance score; X is the node feature matrix; A is the adjacency matrix; H(l) is the l-th layer feature; W(l) is the weight matrix; σ is the activation function; β is the propagation coefficient; xi is the missing value point; N(xi) is the spatiotemporal neighborhood; w(xj) is the adaptive weight; h is the bandwidth parameter; Q={q1...qm} is the quality indicator set; ΔQ is the improvement indicator.

[0070] This step first calculates the mutual information matrix between features through the feature correlation network, and uses the feature importance ranking algorithm to accurately evaluate the weights of features in each dimension, avoiding the evaluation bias caused by fixed feature weights in traditional methods. In terms of quality assessment, a multi-dimensional collaborative evaluation algorithm is introduced, and a three-dimensional evaluation of data quality is achieved by constructing a quality assessment tensor. In particular, in the construction of the quality propagation network, the graph neural network algorithm is used to calculate the propagation coefficient between nodes, which not only considers the direct quality impact, but also captures the indirect quality propagation effect. Through the adaptive interpolation algorithm for data compensation, the system can automatically select the optimal compensation strategy according to the data characteristics, significantly improving the compensation effect. Practice has shown that the system can improve data quality by more than 30%, and the compensation accuracy is 50% higher than that of traditional methods, providing more reliable data guarantee for subsequent risk assessment.

[0071] According to one aspect of the present application, step S22 is specifically:

[0072] Step S221: Obtain multi-domain feature tensors and inter-domain correlation matrix; construct basic feature space; apply nonlinear feature decomposition algorithm to extract potential patterns; calculate pattern significance score; generate basic pattern matrix and significance vector. Specifically, for the input feature matrix X∈Rm×n, construct kernel function K(xi,xj)=exp(-||xi-xj|| 2 / 2σ 2 ). Calculate the eigenvalues ​​λi and eigenvectors vi of the kernel matrix. Select the first k principal components, and the projection matrix P=V[:,:k]. Obtain the nonlinear feature F=KP, where K is the kernel matrix of the original data.

[0073] Step S222: Receive the basic pattern matrix; construct the time series feature sequence; apply the dynamic time series segmentation algorithm to identify the time series evolution features; extract the time series change pattern; output the time series pattern matrix and the time series evolution vector. Specifically, calculate the local trend feature vector L of the sequence S={s1...sn}. Use the sliding variance window to detect the mutation point: vi=var(S[i:i+w]). When vi>μv+ασv, mark it as a candidate segmentation point. Apply dynamic programming to merge adjacent segments and minimize the variance within the segment.

[0074] Step S223: Obtain the basic pattern matrix and the inter-domain correlation matrix; construct a spatial feature network; apply the spatial correlation mining algorithm to identify the spatial aggregation pattern; calculate the spatial pattern features; generate the spatial pattern matrix and the spatial feature vector. Specifically, construct the spatial adjacency matrix W. Calculate the local spatial autocorrelation index: Ii=Σjwij(xi-μ)(xj-μ) / σ 2 . Identify significant clustering areas: Calculate the p value through Monte Carlo simulation, and determine that p < α is a significant clustering.

[0075] Step S224: Receive the temporal pattern matrix, temporal evolution vector, spatial pattern matrix and spatial feature vector; construct a pattern evolution network; apply the spatiotemporal coupling analysis algorithm to calculate the evolution characteristics; integrate temporal and spatial information; output the evolution feature matrix and evolution weight vector. Specifically, construct the spatiotemporal feature tensor T∈Rm×n×t. Calculate the correlation matrix Rt between time slices and the correlation matrix Rs between spatial positions. Fusion obtains the coupling strength matrix C=αRt+(1-α)Rs, where α is determined by maximizing mutual information.

[0076] Step S225: Obtain the evolution feature matrix, evolution weight vector and significance vector; apply the pattern fusion optimization algorithm to integrate the multi-dimensional pattern features; calculate the pattern importance index; generate the final risk pattern set and pattern evolution matrix. For multiple pattern features {M1...Mk}, calculate the complementarity score: ci,j=1-|corr(Mi,Mj)|. Construct the fusion weight: wi=exp(Σjci,j) / ΣkΣjck,j. The final fusion pattern F=Σwi·Mi.

[0077] The characters in this step are as follows: X∈Rm×n is the feature matrix; K(xi,xj) is the kernel function; σ 2 is the kernel width parameter; λi is the eigenvalue; vi is the eigenvector; S={s1...sn} is the time series; L is the local trend vector; vi is the local variance; w is the window size; W is the spatial adjacency matrix; Ii is the local spatial autocorrelation index; T∈Rm×n×t is the spatiotemporal feature tensor; Rt is the temporal correlation matrix; Rs is the spatial correlation matrix; α is the fusion weight; {M1...Mk} is the pattern feature set; ci,j is the complementarity score.

[0078] This step first processes the multi-domain feature tensor through a nonlinear feature decomposition algorithm, extracts potential patterns and calculates significance scores, breaking through the limitations of traditional linear decomposition methods and being able to capture more complex nonlinear risk patterns. In the temporal feature analysis phase, the dynamic time series segmentation algorithm is used, and the system can adaptively identify key temporal change points, which improves the pattern recognition accuracy by 40% compared with fixed window segmentation. In particular, in terms of spatial feature network construction, the spatial association mining algorithm is used to identify spatial aggregation patterns, and combined with the inter-domain correlation matrix, a three-dimensional characterization of the local and overall risk patterns of the building structure is achieved. By integrating temporal and spatial information through the spatiotemporal coupling analysis algorithm, the system can accurately track the dynamic evolution of risk patterns, and the prediction accuracy is 35% higher than that of traditional methods. Finally, the pattern fusion optimization algorithm is used to integrate multi-dimensional pattern features, and the representativeness and integrity of the final risk pattern set are ensured by calculating the pattern importance index, so that the coverage rate of risk identification reaches more than 95%.

[0079] According to one aspect of the present application, step S24 is specifically:

[0080] Step S241: Obtain risk dependency matrix and risk propagation matrix; construct multidimensional risk indicator space; apply indicator decomposition algorithm to extract key risk dimensions; calculate dimension weight coefficients; generate risk indicator matrix and dimension weight vector. Construct a hierarchical structure for indicator set I={i1...in}. Calculate inter-layer contribution: cij=dij / di(j-1). Extract critical path K={k1...km} to maximize Π(ck). The weight of each node is wk=ck / Σck.

[0081] Step S242: receiving the risk indicator matrix and the risk dependency matrix; constructing a risk association network; applying a network propagation algorithm to calculate node influence; evaluating association strength; and outputting an association network matrix and a node influence vector.

[0082] Specifically, construct the initial node influence vector v0. Iterate and propagate: vt+1=αWvt+(1-α)v0, where W is the normalized adjacency matrix. When ||vt+1-vt||<ε, stop the iteration and get the steady-state influence distribution.

[0083] Step S243: Obtain the node influence vector and dimension weight vector; construct a dynamic weight calculation model; apply an adaptive optimization algorithm to calculate the time-varying weights of each dimension; integrate multi-dimensional weights; generate a comprehensive weight matrix and a time-varying coefficient vector.

[0084] Construct the objective function f(w)=Σλigi(w). Apply gradient descent to update the weights: wt+1=wt-η▽f(wt). Dynamically adjust the learning rate: η=η0 / sqrt(t). Use Adam optimizer to accelerate convergence.

[0085] Step S244: Receive the comprehensive weight matrix, the association network matrix and the risk propagation matrix; construct a hierarchical assessment model; apply a multi-level assessment algorithm to calculate the risk score of each layer; fuse the results of different levels; output the risk score matrix and the hierarchical weight vector. Specifically, calculate the basic score sl for each layer l. Introduce the inter-layer influence matrix M and update the score: s'l=sl+ΣkMklsk. Apply the attention mechanism to calculate the layer weight: αl=softmax(Wsl+b). The final score S=Σαls'l.

[0086] Step S245: Obtain the risk scoring matrix, hierarchical weight vector and time-varying coefficient vector; apply the risk level optimization algorithm to calculate the optimal classification threshold; determine the risk level classification standard; generate preliminary risk levels and initial confidence levels. Specifically, construct a risk score sequence R={r1...rn}. Use KDE to estimate the probability density p(r). Find the optimal split point θ={θ1...θk} to minimize the intra-class variance: minΣi=1...kΣr∈Ci(r-μi)².

[0087] Step S246: Receive the risk level and initial confidence; apply the validation evaluation algorithm to calculate the stability of the evaluation results; optimize the confidence calculation; output the final risk level and confidence. Use k-fold cross validation. Calculate the stability index for each fold: si=1-var(yi) / var(y). Estimate the confidence interval through Bootstrap resampling. Calculate the consistency coefficient κ to evaluate the reliability of the results.

[0088] The characters in this step are explained as follows: I={i1...in} is the indicator set; cij is the inter-layer contribution; K={k1...km} is the critical path; wk is the node weight; v0 is the initial influence vector; W is the normalized adjacency matrix; α is the damping coefficient; ε is the convergence threshold; f(w) is the objective function; λi is the weight coefficient; η is the learning rate; sl is the score of the lth layer; M is the inter-layer influence matrix; αl is the layer weight; R={r1...rn} is the risk score sequence; θ={θ1...θk} is the segmentation point set; μi is the class mean.

[0089] This step first extracts key risk dimensions from the risk dependency matrix and risk propagation matrix through the indicator decomposition algorithm, constructs a multidimensional risk indicator space, and increases the dimension coverage of risk assessment to 98%. In the risk association network construction stage, the network propagation algorithm is applied to calculate the node influence, which can accurately assess the importance and influence range of risk nodes in the entire network, and improves the influence assessment accuracy by 40% compared with the traditional centrality measurement method. Especially in the calculation of dynamic weights, the adaptive optimization algorithm is used to achieve accurate calculation of time-varying weights of each dimension, so that the system can automatically adjust the assessment strategy according to the dynamic changes of risk status. By integrating risk scores at different levels through a multi-level assessment algorithm and combining dynamic threshold optimization, the system can not only accurately divide risk levels, but also provide confidence assessments with a credibility of more than 90%, providing a reliable quantitative basis for management decisions.

[0090] According to one aspect of the present application, step S32 is specifically:

[0091] Step S321: Obtain fusion information tensor and context weight vector; construct scene feature extraction model; apply multi-dimensional decomposition algorithm to extract scene key features; calculate feature importance score; generate scene feature matrix and feature importance vector. Specifically, construct feature matrix X∈Rm×n. Apply SVD decomposition: X=UΣV^T. Select the principal component that satisfies Σλi / ΣΛ>η. Calculate the contribution rate of each dimension: ci=λi / ΣΛ. Extract and reconstruct features F=UkΣkVk^T, where k is the retained dimension.

[0092] Step S322: receiving a scene feature matrix and a feature importance vector; constructing an adaptive clustering model; applying a dynamic scene classification algorithm to identify typical scene categories; calculating category representative features; and outputting a scene category matrix and a category feature vector.

[0093] Specifically, calculate the cosine similarity sij between scene feature vectors. Construct affinity matrix A: aij=exp(-||si-sj||² / 2σ²). Apply spectral clustering: calculate the normalized Laplace matrix L, solve Lv=λv to get the cluster indicator vector.

[0094] Step S323: Obtain scene category matrix and category feature vector; construct scene association network; apply correlation analysis algorithm to calculate the probability of transition between scenes; evaluate scene stability; generate scene association matrix and stability vector. Specifically, construct scene transition frequency matrix T. Calculate conditional probability P(j|i)=tij / Σtik. Apply Markov chain analysis to solve steady-state distribution π: πT=π. Calculate transition probability entropy H=-ΣP(j|i)logP(j|i) to evaluate stability.

[0095] Step S324: receiving the scene association matrix, stability vector and feature importance vector; constructing a threshold optimization model; applying a multi-objective optimization algorithm to calculate the optimal threshold for each scene; evaluating the threshold adaptability; and outputting an initial threshold matrix and an adaptability index vector. Specifically,

[0096] Define the objective function set {f1...fm}. Construct the Pareto dominance relation: xi≻xj if and only if for any k, fk(xi)≥fk(xj). Use the NSGA-II algorithm: non-dominated sorting + crowding distance sorting. Select the non-dominated solution set as the optimal threshold.

[0097] Step S325: Obtain the initial threshold matrix, adaptability index vector and scene association matrix; apply the scene collaborative adjustment algorithm to optimize the threshold difference between adjacent scenes; calculate the balance factor; generate the final scene threshold matrix and balance factor vector. Specifically, calculate the scene threshold difference matrix D. Construct a smooth constraint: minΣΣwij(θi-θj)². Introduce a monotonicity constraint: θi≥θj. Apply quadratic programming to solve the optimal threshold distribution.

[0098] The characters in this step are explained as follows: X∈Rm×n is the feature matrix; λi is the eigenvalue; ci is the contribution rate; sij is the cosine similarity; A is the affinity matrix; L is the Laplace matrix; T is the transition frequency matrix; P(j|i) is the conditional probability; π is the steady-state distribution; H is the transition entropy; {f1...fm} is the target function set; wij is the scene weight; θi is the scene threshold.

[0099] This step first processes the fused information tensor through a multi-dimensional decomposition algorithm to extract the key features of the scene, which can capture more than 98% of the scene feature information. In the scene classification link, the dynamic scene classification algorithm is adopted, and the system can adaptively identify typical scene categories, and the scene recognition accuracy rate reaches 95%. In particular, in terms of scene association network construction, the probability of transition between scenes is calculated through the correlation analysis algorithm, and the scene stability is evaluated, so that the system can accurately predict the scene change trend, and the prediction accuracy rate reaches 85%. The multi-objective optimization algorithm is introduced to calculate the optimal threshold of each scene, and the threshold difference between adjacent scenes is optimized through the scene collaborative adjustment algorithm, which not only ensures the rationality of the threshold setting within a single scene, but also maintains the smoothness of the threshold change between different scenes, significantly reducing the false alarm rate caused by threshold mutations, and improving the overall warning accuracy by more than 40%.

[0100] According to one aspect of the present application, step S33 is specifically:

[0101] Step S331: Obtain the scene threshold matrix, balance factor vector and current alert level; construct an adjustment demand assessment model; apply the difference analysis algorithm to calculate the deviation between the current level and the threshold; assess the necessity of adjustment; generate the adjustment demand matrix and the necessity score vector. Specifically, calculate the distance matrix between the current level vector c and the target threshold vector t: dij=|ci-tj|. Apply the DTW algorithm to find the optimal alignment path. Calculate the deviation score: s=Σwidi, wi is the position weight.

[0102] Step S332: Receive the adjustment requirement matrix and the necessity score vector; construct the constraint model; apply the rule extraction algorithm to identify the conversion constraints; calculate the constraint strength; output the constraint matrix and the constraint weight vector. Specifically, construct the historical conversion sequence set S. Use the Apriori algorithm to mine frequent patterns. Calculate the rule confidence: conf(X→Y)=sup(X∪Y) / sup(X). Keep the rule set that meets the minimum support and confidence.

[0103] Step S333: Obtain the constraint matrix, constraint weight vector and balance factor vector; construct a path planning network; apply a dynamic programming algorithm to calculate multiple feasible conversion paths; evaluate the path cost; generate a path candidate set and a path cost vector. Specifically, define the state space S and the transfer cost c(s,s'). Construct the optimal value function: V(s)=min[c(s,s')+γV(s')]. Obtain the optimal path by backtracking: π(s)=argmin[c(s,s')+γV(s')].

[0104] Step S334: Receive the path candidate set and the path cost vector; construct a smooth transition model; apply the progressive transition algorithm to calculate the optimal transition step; evaluate the transition smoothness; output the transition sequence matrix and the smoothness index vector. Specifically, calculate the distance between the target states d(si,sj). Design the transition step function: δ(t)=δ0exp(-βt). Update each step: st+1=st+δ(t)(starget-st). Check whether the constraints are satisfied.

[0105] Step S335: Obtain the transformation sequence matrix and smoothness index vector; apply the conflict detection algorithm to identify potential transformation conflicts; calculate the conflict resolution method; generate the conflict resolution matrix and solution scoring vector. Specifically, construct the transformation dependency graph G(V,E). Apply topological sorting to detect circular dependencies. Calculate the critical path: use the Floyd-Warshall algorithm. Identify the timing conflict points: solve the maximum flow problem.

[0106] Step S336: Receive the conflict resolution matrix, solution scoring vector and conversion sequence matrix; apply the solution verification algorithm to evaluate the reliability of the adjustment solution; select the optimal conversion path; output the final adjusted alert level and level conversion path. Specifically, construct a historical conversion effect dataset D. Train and verify the model: use XGBoost to predict the conversion success rate. Calculate the solution reliability score: R=E[p(success|x)]. Estimate the confidence interval.

[0107] The characters in this step are explained as follows: c is the current level vector; t is the target threshold vector; dij is the distance matrix element; wi is the position weight; S is the set of historical transition sequences; conf(X→Y) is the rule confidence; sup(X) is the support; c(s,s') is the transfer cost; γ is the discount factor; V(s) is the value function; δ(t) is the step size function; β is the attenuation coefficient; G(V,E) is the transition dependency graph.

[0108] This step first evaluates the deviation between the current level and the threshold through a difference analysis algorithm, and constructs an adjustment demand assessment model, which can accurately determine the necessity and urgency of the adjustment, avoiding more than 90% of unnecessary adjustments. In terms of constraint processing, the rule extraction algorithm is applied to identify the conversion constraints, and the importance of each constraint is quantified through the constraint weight vector, so that the level adjustment is always kept within a reasonable range. Especially in the path planning stage, multiple feasible conversion paths are calculated through a dynamic programming algorithm, and the optimal conversion step size is determined through a progressive conversion algorithm, which ensures both the smoothness of the adjustment and the timeliness of the response. The conflict detection algorithm identifies and resolves potential conversion conflicts, and the scheme verification algorithm is combined to evaluate the reliability of the adjustment scheme, so that the success rate of level conversion reaches more than 95%, which is 40% higher than the traditional method.

[0109] According to one aspect of the present application, step S42 is specifically:

[0110] Step S421: Obtain fusion information tensor and context weight vector; construct content template library; apply template classification algorithm to match basic templates for different scenes; calculate template applicability score; generate basic template matrix and template score vector. Specifically, construct template feature vector set M={m1...mn}. Calculate the matching degree between scene feature vector s and template: r(s,mi)=cosine(s,mi). Use hierarchical clustering method to group templates. Calculate intra-class similarity: sim(C)=Σi,j∈Cr(mi,mj) / |C| 2 . Select the best matching category.

[0111] Step S422: Receive the fusion information tensor and the template score vector; construct an information extraction model; apply a multidimensional feature extraction algorithm to identify key warning factors; evaluate the importance of factors; output a key factor matrix and a factor weight vector. Specifically, construct a multidimensional feature tensor X∈Rm×n×k. Apply Tucker decomposition: X≈G×1U×2V×3W. Calculate feature importance: wi=||Ui|| 2 / Σ||U|| 2 . Extract key dimensions: Select feature combinations whose cumulative contribution rate exceeds θ.

[0112] Step S423: Obtain the key element matrix and element weight vector; construct a semantic association network; apply the semantic analysis algorithm to calculate the logical relationship between elements; evaluate the association strength; generate a semantic association matrix and a logical strength vector. Specifically, construct the element word vector matrix E. Calculate the cosine similarity between elements: sij=ei^T·ej / (||ei||·||ej||). Apply the PageRank algorithm to calculate the element centrality. Construct a semantic association graph: retain the edges with weights greater than the threshold.

[0113] Step S424: Receive the basic template matrix, semantic association matrix and logic strength vector; construct a dynamic generation model; apply the content organization algorithm to generate preliminary warning text; evaluate the text integrity; output the initial warning content and integrity indicator vector. Specifically, construct a content generation sequence based on the semantic association graph G. Use the attention mechanism to calculate the context weight: α=softmax(QK^T)V. Iterate to generate content: select the next element with the maximum probability at each step. Check logical integrity.

[0114] Step S425: Obtain the initial warning content, integrity index vector and context weight vector; apply the content evaluation algorithm to calculate the text quality index; optimize the content expression; extract the semantic features; generate the final warning content and semantic feature vector. Specifically, calculate the multidimensional quality index: {clarity, completeness, accuracy, timeliness}. Use fuzzy comprehensive evaluation: construct the factor set U and the evaluation set V. Calculate the fuzzy relationship matrix R. Obtain the evaluation vector B=AOR.

[0115] The characters in this step are explained as follows: M={m1...mn} is the template feature vector set; r(s,mi) is the matching degree; sim(C) is the intra-class similarity; X∈Rm×n×k is the feature tensor; wi is the feature importance; θ is the cumulative contribution rate threshold; E is the element word vector matrix; sij is the cosine similarity; G is the semantic association graph; α is the attention weight; Q, K, V are query, key, and value matrices; U is the factor set; V is the evaluation set; R is the fuzzy relationship matrix.

[0116] By building a content template library and using a template classification algorithm to achieve accurate matching of different scenarios, the template adaptation accuracy rate reaches over 96%. In the information extraction stage, a multidimensional feature extraction algorithm is used to identify key warning elements, and the importance of each element is quantified through the element weight vector, ensuring that the key information coverage rate of the warning content reaches 98%. In particular, in terms of semantic association network construction, the logical relationship between elements is calculated through a semantic analysis algorithm, which improves the logical coherence of the generated warning content by 45%. The preliminary warning text is generated through a dynamic generation model and a content organization algorithm, and the text expression is optimized by combining a content evaluation algorithm, which not only ensures the accuracy and completeness of the warning content, but also improves the readability and comprehensibility of the information. Practice shows that the warning content generated by this step improves the information transmission efficiency by 50% compared with the traditional template generation, and the receiver's understanding accuracy rate is improved by 40%, greatly reducing the decision delay caused by information understanding deviation.

[0117] According to one aspect of the present application, step S44 is specifically:

[0118] Step S441: Obtain warning level and priority matrix; construct receiving object feature library; apply object classification algorithm to identify different types of receiving objects; calculate object response features; generate object classification matrix and response feature vector. Specifically, construct receiving object feature matrix X. Apply K-means++ to initialize cluster center. Iterative optimization: Update class center μj=Σxi∈Cjxi / |Cj|. Calculate silhouette coefficient to select optimal k value. Generate object category label.

[0119] Step S442: Receive the object classification matrix and response feature vector; build a push condition model; apply the condition recognition algorithm to extract key constraints such as time, location, and personnel; evaluate the importance of constraints; output the push condition matrix and condition weight vector. Specifically, extract the spatiotemporal constraint feature vector f(t, s, p). Use decision trees to mine key constraint rules. Calculate rule importance: importance(r)=ΔI(Sr) / |Sr|. Keep the rule set whose importance exceeds the threshold.

[0120] Step S443: Obtain the push condition matrix and priority matrix; construct a channel evaluation model; apply a multidimensional evaluation algorithm to calculate the transmission efficiency and reliability of different channels; comprehensively evaluate channel performance; generate a channel characteristic matrix and a performance score vector. Specifically, construct a channel performance indicator set P = {efficiency, reliability, cost}. Standardize the indicator value: z = (x-μ) / σ. Calculate the comprehensive score: s = Σwipi, and the weight wi is determined by the entropy weight method. Sort and select the optimal channel combination.

[0121] Step S444: Receive channel characteristic matrix, performance score vector and conditional weight vector; construct timing generation model; apply timing optimization algorithm to calculate optimal push time window; evaluate timing rationality; output timing scheduling matrix and rationality index vector. Specifically, establish time window utility function U(t). Consider the receiver activity pattern A(t). Optimization goal: max∫U(t)A(t)dt. Solve using genetic algorithm: Encode time window as chromosome, optimize by crossover mutation.

[0122] Step S445: Obtain the time scheduling matrix and channel characteristic matrix; apply the resource allocation algorithm to optimize channel load balancing; calculate resource utilization efficiency; generate the resource allocation matrix and efficiency index vector. Specifically, construct the load balancing objective function: minΣ(Li-L*) 2 Consider the capacity constraint: Σxij≤ci. Use the Hungarian algorithm to solve the optimal allocation. Dynamically adjust the allocation strategy: update according to the real-time load situation.

[0123] Step S446: Receive the resource allocation matrix, efficiency index vector and timing matrix; apply the push effect evaluation algorithm to calculate the expected push effect; optimize the push plan; integrate all information to generate the final push strategy, channel selection and push timing. Specifically, collect historical push data D. Extract the influencing factor set F = {time, channel, object}. Train the XGBoost model to predict the delivery rate and reading rate. Calculate the area under the ROC curve to evaluate the model effect. Predict the optimal push plan.

[0124] The characters in this step are explained as follows: X is the object feature matrix; μj is the class center; f(t, s, p) is the spatiotemporal constraint feature; P is the performance indicator set; U(t) is the time window utility function; A(t) is the activity pattern; L̄ is the average load; ci is the capacity constraint; xij is the allocation variable; D is the historical push data; F is the impact factor set.

[0125] This step first builds a refined receiving object feature library through the object classification algorithm, identifies the response characteristics of different types of receiving objects, and improves the accuracy of receiving object classification to 97%. In the push condition model construction stage, the condition identification algorithm is applied to extract key constraints, and the importance of constraints is quantified through the condition weight vector, which improves the adaptability of push conditions by 55%. Especially in terms of channel evaluation, the transmission efficiency and reliability of different channels are calculated through the multi-dimensional evaluation algorithm, which can achieve a message delivery rate of more than 95%. The optimal push time window is determined by the timing generation model and the timing optimization algorithm, and the channel load balancing is optimized by combining the resource allocation algorithm, which not only ensures the instant delivery of emergency warnings, but also shortens the overall push response time by 60%. Finally, the expected push effect is evaluated and optimized by the push effect evaluation algorithm, so that the effective reach rate of warning information is increased to more than 92%, and the information reception efficiency is improved by 45% compared with the traditional push method. Practice has proved that this step significantly improves the accuracy and timeliness of warning push, and effectively solves the information redundancy and attention distraction problems caused by the "one-size-fits-all" and "cast-net" push in traditional warning push.

[0126] In this embodiment, the implementation process of the data processing layer is as follows: the system receives three types of input data: D_raw is the original monitoring data, including the timestamp t and the data value v; P_loc is the spatial position data, including the three-dimensional coordinates (x, y, z); Q_ind is the data quality indicator, with a value range of [0,1];

[0127] The spatiotemporal correlation matrix R is calculated using R[i,j] = α·exp(-|ti-tj| / τt) + β·exp(-||pi-pj|| / τs); where α is the time dimension weight coefficient, ranging from [0,1]; β is the spatial dimension weight coefficient, β = 1-α; τt is the time scale parameter used to control the time correlation decay rate; τs is the spatial scale parameter used to control the spatial correlation decay rate; ti, tj are the timestamps of data points i and j; pi, pj are the spatial position vectors of data points i and j;

[0128] Adaptive adjustment of window size W: W = W_base * (1 + γ·σ_t); where W_base is the base window size; γ is the adaptive coefficient, ranging from [0,2]; σ_t is the standard deviation of the data in the current time period;

[0129] The feature tensor F is calculated using F = {F_k | k ∈ W_set}; F_k = Transform(D_aligned,w_k);

[0130] Among them, W_set is the sliding window size set; w_k is the kth window size; D_aligned is the data set after time-space alignment; Transform() is the feature transformation function, including: time domain features are mean, variance, peak factor, skewness, and kurtosis; frequency domain features are power spectral density and main frequency components; time-frequency features are wavelet coefficients and Hilbert spectrum;

[0131] The dynamic update of feature weights adopts w_i = softmax(η·S_i + (1-η)·T_i); where w_i is the weight of the i-th feature; S_i is the feature significance score; T_i is the temporal correlation score; η is the balance factor, with a value range of [0,1]; softmax() is the normalization function;

[0132] Data relationship graph construction: node similarity calculation Sim(i,j) = exp(-||f_i f_j|| 2 / σ 2 ), where f_i, f_j are the feature vectors of nodes i and j; σ is the kernel width parameter, which is determined by cross-validation;

[0133] Quality indicator update equation Q_t+1 = λ·A·Q_t + (1-λ)·Q_0;

[0134] Among them, Q_t is the quality index vector at time t; A is the normalized adjacency matrix; λ is the propagation factor, ranging from [0,1]; Q_0 is the initial quality index;

[0135] Multi-view feature space construction: The anomaly score of each view v is calculated as Score_v(x) = Σ(d_k·K(x,x_k)) / Σ K(x,x_k); where d_k is the degree of anomaly of the kth reference point; K(·,·) is the kernel function; x_k is the reference data point; the comprehensive anomaly score is calculated as S_final(x) = Σ(ω_v·Score_v(x)); where ω_v is the weight of the vth view; Score_v is the anomaly score under each view;

[0136] Multi-domain feature tensor construction, inter-domain correlation matrix calculation C[i,j] =<F_i, F_j> / (||F_i||·||F_j||); where F_i, F_j are feature tensors of different feature domains; <·,·> is the tensor inner product operation; ||·|| is the tensor norm;

[0137] Fusion feature calculation F_fusion = Σ(α_i·F_i + β_i·CrossDomain(F_i, F_j)); where α_i is the single domain feature weight; β_i is the cross-domain feature weight; CrossDomain() is the cross-domain feature extraction function;

[0138] Multi-scale pattern graph construction, node-to-node edge weight calculation: W_edge(i,j) = φ(M_i, M_j)·ψ(T_i, T_j)·θ(S_i, S_j); where φ(M_i, M_j) is the pattern similarity function; ψ(T_i, T_j) is the time correlation function; θ(S_i, S_j) is the spatial correlation function; M_i, M_j are the pattern features of nodes i and j; T_i, T_j are time features; S_i, S_j are spatial features;

[0139] Evolution feature extraction E(t) = G(P_t, P_t-1)·H(Δt); where P_t is the mode state vector at time t; G(·,·) is the state transfer function; H(Δt) is the time decay function; Δt is the time interval;

[0140] The risk dependency network is constructed and the dependency strength is calculated as D(i,j) = μ·Direct(i,j) + (1-μ)·Indirect(i,j); where Direct(i,j) is the direct dependency strength; Indirect(i,j) is the indirect dependency strength; μ is the direct dependency weight coefficient;

[0141] Cascade effect propagation, the propagation intensity matrix is ​​updated P_t+1 = σ(W·P_t + b)·(1 exp(-k·t)); where W is the propagation weight matrix; b is the bias vector; σ is the activation function; k is the propagation rate parameter; t is the propagation time;

[0142] Multi-layer risk assessment network is constructed, and the hierarchical assessment function R_l = f_l(W_l·X_l + B_l)·α_l + R_l-1·(1-α_l);

[0143] Among them, R_l is the risk score of the lth layer; f_l is the hierarchical evaluation function; W_l is the hierarchical weight matrix; X_l is the hierarchical feature vector; B_l is the hierarchical bias vector; α_l is the hierarchical fusion coefficient.

[0144] Threshold update equation T_new = T_old + η·(▽L_risk + λ·▽L_smooth); where T_new, T_old are the new and old threshold vectors; η is the learning rate; L_risk is the risk loss function; L_smooth is the smoothing loss function; λ is the smoothing coefficient.

[0145] Fusion feature calculation F_fusion = Attention(F_current, F_history, F_context); where F_current is the current risk feature; F_history is the historical alarm feature; F_context is the context feature; Attention() is the attention weighting function.

[0146] Attention weight calculation α_ij = softmax(Q_i·K_j^T / sqtr d)·V_j; where Q_i is the query vector; K_j is the key vector; V_j is the value vector; d is the feature dimension; sqrt d is the scaling factor.

[0147] The optimal threshold calculation for the scene is T_opt(s) = argmin_{T} [L_miss(T,s)·w_m + L_false(T,s)·w_f]; where s is the scene feature vector; L_miss is the missed detection loss function; L_false is the false alarm loss function; w_m is the missed detection weight; and w_f is the false alarm weight.

[0148] Calculation of the balance factor between scenes, B(i,j) = exp(-||F_i F_j|| 2 / τ)·(1 + ρ·Corr(i,j)); where F_i, F_j are the feature vectors of scenes i and j; τ is the temperature parameter; ρ is the correlation weight coefficient; Corr(i,j) is the correlation coefficient between scenes;

[0149] The transition probability matrix is ​​calculated as P(i→j) = softmax(E_ij·S_ij / κ); where E_ij is the transition energy from level i to j; S_ij is the safety constraint coefficient; κ is the temperature parameter controlling the transition probability distribution; and softmax is the normalization function.

[0150] Transformation path optimization Path_opt = argmin_{path}[Σ C_trans + λ_t·T_delay + λ_s·S_risk];

[0151] Among them, C_trans is the conversion cost matrix; T_delay is the time delay penalty; S_risk is the security risk penalty; λ_t, λ_s are the balance coefficients;

[0152] Evaluation score calculation Score = Σ(w_i·M_i)·(1 + β·ΔQ); where M_i is the performance indicator set; w_i is the indicator weight; ΔQ is the quality improvement; β is the gain coefficient.

[0153] Feedback signal generation F_signal = G(ΔS, ΔT, ΔR)·(1 exp(-γ·t)); where ΔS is the state change; ΔT is the time change; ΔR is the risk change; γ is the time attenuation coefficient; and G is the feedback gain function.

[0154] Context weight calculation W_context = Transformer(Q_c, K_c, V_c) Mask(R_t); where Q_c, K_c, V_c are context query, key, and value matrices; R_t is the temporal association matrix; Transformer is the attention transformation function; Mask is the temporal mask function;

[0155] Fusion feature calculation, F_merged = BiLSTM(F_base ⊕ F_context)·(1 + δ·Att_score); where,

[0156] F_base is the base feature vector; F_context is the context feature vector; ⊕ is the feature concatenation operation; BiLSTM is the bidirectional long short-term memory network; Att_score is the attention score; δ is the attention gain coefficient.

[0157] Template matching score calculation Score_template(i) = ω_s·Sim_semantic(i) + ω_c·Sim_context(i) + ω_u·Sim_urgency(i); where Sim_semantic is semantic similarity; Sim_context is context similarity; Sim_urgency is urgency matching; ω_s, ω_c, ω_u are corresponding weight coefficients;

[0158] The content generation optimization objective L_gen = -[λ_1·log P(y|x) + λ_2·R_coh + λ_3·R_acc +λ_4·R_comp]; where P(y|x) is the generation probability; R_coh is the content coherence reward; R_acc is the accuracy reward; R_comp is the completeness reward; λ_1~λ_4 are balance coefficients.

[0159] Classification probability calculation P(c|x) = Softmax(MLP(CNN(x) ⊕ RNN(x)) · W_att); where CNN() is convolution feature extraction; RNN() is sequence feature extraction; MLP() is multi-layer perceptron; W_att is the attention weight matrix; ⊕ is the feature fusion operation;

[0160] Priority calculation Priority(i) = Base_pri(i)·(1 + α·Urg(i) + β·Imp(i))·Decay(t); where Base_pri is the basic priority; Urg is the urgency factor; Imp is the importance factor; Decay is the time decay function; α, β are weight coefficients;

[0161] Channel selection probability calculation P_channel(c) = softmax(η·Eff(c) + θ·Rel(c) + φ·Load(c)); where Eff is the channel efficiency index; Rel is the channel reliability index; Load is the load balancing index; η, θ, φ are weight coefficients;

[0162] The push time is optimized as follows: T_opt = argmin_{t} [w_1·D_delay(t) + w_2·C_conflict(t) + w_3·R_response(t)]; where D_delay is the delay loss function; C_conflict is the conflict loss function; R_response is the response loss function; w_1, w_2, w_3 are weight coefficients;

[0163] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for dynamically adjusting building safety alert levels based on artificial intelligence, characterized in that: The steps include: Step S1, receiving the original monitoring data, spatial position data and data quality indicators from the sensor array, and sequentially performing spatiotemporal alignment, feature extraction, quality assessment and anomaly detection through a multi-source heterogeneous data preprocessing framework to obtain and output a preprocessed data set and anomaly score matrix; Step S2: obtain the preprocessed data set and the anomaly score matrix, perform feature space construction, risk pattern extraction, correlation analysis and level assessment in sequence through multi-dimensional feature fusion processing and risk pattern recognition analysis, and obtain and output the risk level and confidence level; Step S3: Obtain risk level and confidence level, and perform feature fusion, threshold optimization, level adjustment and effect evaluation in sequence through the alert level dynamic adjustment mechanism to obtain and output a new alert level, evaluation score and adjustment effect indicator; Step S4: obtain the new alert level, evaluation score and adjustment effect index, and perform information fusion, content generation, level classification and push optimization in sequence through intelligent generation and hierarchical push processing of warning information, and obtain and output push strategy, channel selection and push timing; Step S1 is specifically as follows: Step S11, obtaining original monitoring data, including timestamp and data value; obtaining spatial position data, including three-dimensional coordinate information; obtaining data quality indicators, indicating data reliability; Construct a spatiotemporal correlation matrix to calculate the strength of association in the temporal and spatial dimensions; Correct the timestamp of each data point to generate a spatiotemporal aligned dataset; The spatiotemporal correlation matrix R is calculated using R[i,j] = α·exp(-|ti-tj| / τt) + β·exp(-||pi-pj|| / τs), where α and β are time and space weight coefficients, τt and τs are time and space scale parameters; ti is the timestamp of the i-th data point; tj is the timestamp of the j-th data point; pi is the spatial coordinate of the i-th data point, including the three-dimensional coordinates (xi, yi,zi); Pj is the spatial coordinate of the j-th data point, including the three-dimensional coordinates (xj, yj, zj); Step S12, receiving a spatiotemporal aligned data set, setting a sliding window size set; constructing a multi-level feature tensor to extract features at different time scales; applying an adaptive weight mechanism to dynamically adjust feature importance; performing cross-scale information fusion to obtain a multi-scale feature matrix; Step S13, obtaining a multi-scale feature matrix and data quality indicators; constructing a data relationship diagram, calculating the similarity between data points; updating the quality indicators of nodes through a quality propagation algorithm; compensating and correcting data points whose quality is lower than a threshold; obtaining a quality assessment compensation data set and updating the quality indicators; Step S14, receiving a quality assessment compensation data set and updating quality indicators; constructing a multi-view feature space, calculating anomaly scores at each view; fusing the evaluation results of multiple views through an integrated boundary learning algorithm to obtain an anomaly degree score for each data point; merging the quality assessment compensation data set with the anomaly degree score to generate a preprocessed data set; organizing the anomaly degree scores of all data points into a matrix form to generate an anomaly score matrix; Step S2 is specifically as follows: Step S21, obtaining a preprocessed data set and anomaly score matrix; constructing feature tensors of the physical domain, frequency domain, and topological domain; calculating the correlation between different feature domains; organizing the correlation into a matrix form to obtain a multi-domain feature tensor and an inter-domain correlation matrix; Step S22, receiving multi-domain feature tensors and inter-domain correlation matrices; constructing a multi-scale pattern graph, calculating the correlation edge weights between risk pattern nodes; calculating the pattern evolution matrix for each time window, and determining the dynamic weight by optimizing the objective function; organizing the extracted risk patterns and evolution information respectively, and generating a risk pattern set and a pattern evolution matrix; Step S23, obtaining a risk pattern set and a pattern evolution matrix; constructing a risk dependency network, calculating the dependency strength between each pair of risk patterns; applying a cascade effect propagation algorithm, calculating the risk propagation strength in the network; organizing the dependency strength and propagation strength into matrix forms, respectively, to form a risk dependency matrix and a risk propagation matrix; Step S24, receiving the risk dependency matrix and the risk propagation matrix; constructing a multi-layer risk assessment network, and calculating the risk score of each layer through a hierarchical assessment function; using a dynamic threshold optimization algorithm to determine the risk level classification standard; calculating the risk level and assessment confidence of each monitoring point according to the risk score and threshold standard; organizing the risk level and confidence information separately to generate the risk level and confidence; Step S3 is specifically as follows: Step S31, obtain risk level, confidence level and pre-stored historical alarm data; construct an information fusion tensor to integrate the current risk level, confidence level and historical alarm features; use a hierarchical attention mechanism to calculate the weight coefficients of different features; obtain a fused information tensor through weighted fusion, and search and output the corresponding context weight vector; Step S32, receiving the fusion information tensor and the corresponding context weight vector; constructing a scene adaptive decision network, and calculating the optimal threshold for different scenes; calculating the balance factor between scenes through a dynamic balance optimization algorithm; organizing the optimal threshold and the balance factor into a matrix and a vector form respectively, and generating a scene threshold matrix and a balance factor vector; Step S33, obtaining the scene threshold matrix, the balance factor vector and the current alert level; constructing a level transition diagram, calculating the transition probability between different levels; using a smooth transition control algorithm to determine the optimal transition path; updating the alert level according to the transition path, and outputting the adjusted alert level and level transition path; Step S34, receiving the adjusted alert level, level conversion path and pre-stored historical adjustment effect data; constructing a performance evaluation network, calculating the adjustment effect score; applying a feedback integration algorithm, generating a feedback signal; sorting the evaluation score, feedback signal and effect index respectively, generating an evaluation score, feedback signal and adjustment effect index; outputting the adjusted alert level as a new alert level; Step S4 is specifically as follows: Step S41, obtaining a new alert level, evaluation score and adjustment effect index; constructing an information fusion tensor to integrate the alert level, evaluation score and adjustment index information; applying a context enhancement network algorithm to calculate the weight coefficient of each dimension of information; obtaining a fusion information tensor through weighted combination, and outputting the corresponding context weight vector at the same time; Step S42, receiving the fusion information tensor and the context weight vector; constructing a content generation graph, establishing semantic associations between nodes; applying a template free dynamic generation algorithm, generating warning content according to the semantic association; extracting semantic features from the generated content, and outputting the warning content and the semantic feature vector; Step S43, obtaining warning content and semantic feature vectors; constructing a multi-level classification network, calculating the classification vector of the warning content; determining the priority level of each warning content through a hierarchical priority allocation algorithm; sorting the classification results and priority information respectively, and generating a warning level and priority matrix; Step S44, receiving the warning level and priority matrix; constructing a push strategy network, generating a push strategy matrix according to the warning level and priority; using a dynamic channel selection algorithm to optimize the selection of push channels; determining the push timing according to timeliness requirements and receiving effects; sorting out the push strategy, channel selection and timing information respectively, generating the push strategy, channel selection and push timing.

2. The method for dynamically adjusting building safety alert levels based on artificial intelligence drive according to claim 1 is characterized in that: The step S11 is specifically as follows: Step S111, obtaining original monitoring data, including timestamps and sensor values; applying an adaptive sliding window algorithm to divide the data into segments; calculating statistical features for each data segment; identifying outliers through cluster analysis; Generate pre-screening dataset and anomaly labeling vector; Step S112, receiving the pre-screened data set and obtaining the timestamp of each data point; Construct timestamp mapping matrix; Apply the dynamic time warping algorithm to calculate the time deviation; Perform timestamp calibration based on the deviation value; Output time-aligned dataset and time deviation matrix; Step S113, obtaining spatial position data and a time-aligned data set; constructing a spatial position association graph; calculating a spatial distance matrix between nodes; applying a spatial interpolation algorithm to supplement missing position information; generating a spatially aligned data set and a spatial association matrix; Step S114, receiving the spatial alignment data set, the time deviation matrix and the spatial correlation matrix; constructing a spatiotemporal correlation feature vector; Apply tensor decomposition algorithm to extract spatiotemporal coupling patterns; Calculate the association strength matrix; Output spatiotemporal correlation matrix; Step S115, obtaining a spatiotemporal correlation matrix and a spatial alignment data set; Apply multi-dimensional data fusion algorithms to integrate spatiotemporal features; Calculate fusion weight coefficient; Generate the final spatiotemporally aligned dataset.

3. The method for dynamically adjusting building safety alert levels based on artificial intelligence drive according to claim 1 is characterized in that: The step S13 is specifically as follows: Step S131, obtaining a multi-scale feature matrix; constructing a feature correlation network; Calculate the mutual information matrix between features; apply the feature importance ranking algorithm to evaluate the feature weights of each dimension; Generate feature weight vector and feature correlation matrix; Step S132, receiving feature weight vectors and data quality indicators; constructing quality assessment tensors; applying multi-dimensional collaborative assessment algorithms to calculate quality scores for each dimension; integrating multi-dimensional assessment results; outputting quality score matrices and dimension weight vectors; Step S133, obtaining a quality score matrix and a feature correlation matrix; Construct a quality propagation network; apply the graph neural network algorithm to calculate the propagation coefficient between nodes; generate quality influence distribution; output the propagation coefficient matrix and influence vector; Step S134, receiving the propagation coefficient matrix and the quality score matrix; constructing a data compensation model; applying an adaptive interpolation algorithm to calculate compensation values; performing data correction; generating a compensation data set and a correction coefficient vector; Step S135, obtaining a compensation data set, a correction coefficient vector and an influence vector; Apply quality verification algorithms to evaluate compensation effects; calculate quality improvement indicators; integrate evaluation results to generate quality assessment compensation datasets and update quality indicators.

4. The method for dynamically adjusting building safety alert levels based on artificial intelligence drive according to claim 1 is characterized in that: The step S22 is specifically as follows: Step S221, obtaining multi-domain feature tensors and inter-domain correlation matrices; constructing a basic feature space; applying a nonlinear feature decomposition algorithm to extract potential patterns; and calculating pattern significance scores; Generate base pattern matrix and saliency vector; Step S222, receiving a basic pattern matrix; constructing a time series feature sequence; Apply dynamic time series segmentation algorithm to identify time series evolution characteristics; Extract the timing change pattern; output the timing pattern matrix and timing evolution vector; Step S223, obtaining a basic pattern matrix and an inter-domain correlation matrix; constructing a spatial feature network; applying a spatial correlation mining algorithm to identify spatial aggregation patterns; and calculating spatial pattern features; Generate spatial pattern matrices and spatial eigenvectors; Step S224, receiving a temporal pattern matrix, a temporal evolution vector, a spatial pattern matrix and a spatial eigenvector; Construct a model evolution network; apply a spatiotemporal coupling analysis algorithm to calculate evolution characteristics; integrate temporal and spatial information; output an evolution characteristic matrix and an evolution weight vector; Step S225, obtaining an evolution feature matrix, an evolution weight vector and a significance vector; applying a pattern fusion optimization algorithm to integrate multi-dimensional pattern features; Calculate the model importance index; Generate the final risk pattern set and pattern evolution matrix.

5. The method for dynamically adjusting building safety alert levels based on artificial intelligence drive according to claim 1 is characterized in that: The step S24 is specifically as follows: Step S241, obtain a risk dependency matrix and a risk propagation matrix; construct a multidimensional risk indicator space; apply an indicator decomposition algorithm to extract key risk dimensions; calculate dimension weight coefficients; generate a risk indicator matrix and a dimension weight vector; Step S242, receiving the risk indicator matrix and the risk dependency matrix; constructing a risk association network; applying a network propagation algorithm to calculate node influence; and evaluating association strength; Output the correlation network matrix and node influence vector; Step S243, obtaining the node influence vector and the dimension weight vector; constructing a dynamic weight calculation model; applying an adaptive optimization algorithm to calculate the time-varying weights of each dimension; integrating multi-dimensional weights; Generate a comprehensive weight matrix and a time-varying coefficient vector; Step S244, receiving the comprehensive weight matrix, the association network matrix and the risk propagation matrix; Construct a hierarchical assessment model; apply a multi-level assessment algorithm to calculate the risk scores of each layer; Integrate results from different levels; Output risk score matrix and level weight vector; Step S245, obtaining a risk scoring matrix, a hierarchical weight vector, and a time-varying coefficient vector; applying a risk level optimization algorithm to calculate an optimal classification threshold; and determining a risk level classification standard; Generate preliminary risk levels and initial confidence levels; Step S246, receiving the risk level and initial confidence level; Apply the verification evaluation algorithm to calculate the stability of the evaluation results; optimize the confidence calculation; Output the final risk level and confidence level.

6. A dynamic adjustment system for building safety alert levels based on artificial intelligence, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions executed by the processor, and the instructions are used to be executed by the processor to implement the artificial intelligence-driven dynamic adjustment method of building safety alert level as described in any one of claims 1 to 5.

Images