Building health monitoring and assessment method and system based on physical neural network

Abstract

The present invention proposes a building health monitoring and assessment method and system based on physical neural network, which includes collecting monitoring data such as building vibration, strain, displacement, etc. through a sensor network, preprocessing data based on wavelet transform and neural network; inputting the processed data into a physical constraint encoder, mapping it to the physical space using the Hamilton quantity encoding method, and constructing constraint conditions in combination with physical model parameters; using a physical constraint neural network to perform feature matching and extract key features; using the principal component analysis method to decompose early warning indicators, and constructing parameter mapping through a deep neural network; generating hierarchical decision rules based on a decision tree algorithm; performing constraint inspection and cost calculation on candidate solutions, and completing solution sorting through a hierarchical analysis method. The present invention embeds physical knowledge into a deep learning framework, improving the accuracy and reliability of monitoring and evaluation.

Description

Technical Field

[0001] The present invention belongs to the field of structural health monitoring and management, and in particular to a building health monitoring and evaluation method and system based on physical neural network. Background Art

[0002] Building health monitoring is important for ensuring the safety of building structures, extending their service life and reducing maintenance costs.

[0003] At present, a lot of research work has been carried out in the field of building health monitoring; in terms of data acquisition, optical fiber sensing, wireless sensor networks and other technologies are used to achieve multi-parameter monitoring; in terms of data processing, wavelet transform, empirical mode decomposition and other methods are used for signal noise reduction and feature extraction; in terms of damage identification, a variety of structural damage diagnosis models are established based on modal parameters, vibration characteristics, etc.; in terms of performance evaluation, structural performance prediction models are constructed through finite element analysis, statistical learning and other means; traditional machine learning methods, such as support vector machines and random forests, have been applied to building status assessment and early warning; however, these methods often separate physical models and data-driven models, and fail to fully utilize the physical knowledge of building structures.

[0004] The existing technical solutions have the following specific problems: First, physical constraints are ignored in the feature extraction process, resulting in the extracted features possibly violating basic physical laws such as conservation of energy and conservation of momentum, affecting the reliability of the evaluation results; second, the data fusion algorithm does not consider the physical topological relationship of the spatial distribution of sensors, resulting in the loss of correlation between local monitoring data; third, the damage diagnosis model lacks constraints on the structural dynamics equations, resulting in deviations between the diagnosis results and the actual physical mechanism; fourth, the performance evaluation method does not embed physical parameters such as mass matrix and stiffness matrix into the neural network structure, reducing the interpretability of the model; fifth, the physical knowledge contained in historical decision-making experience is not fully utilized in the generation of early warning strategies, limiting the rationality of the decision-making plan; sixth, the coupling relationship between physical parameters is not considered in the multi-objective optimization process, resulting in the optimization results may not meet the actual needs of the project; these technical problems seriously restrict the actual application effect of the building health monitoring system. Summary of the invention

[0005] The purpose of the invention is to provide a building health monitoring and assessment method and system based on physical neural network, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution, a building health monitoring and assessment method based on physical neural network, includes the following steps:

[0007] S1. Obtain the original sensor data set, perform data preprocessing, and generate a standardized data set; perform feature extraction on the standardized data set to generate a feature vector set; fuse the feature vector set and pre-stored historical feature data to generate a fused feature tensor;

[0008] S2, performing physical constraint processing on the fused feature tensor and the pre-stored physical model parameter set to generate a physical constraint feature matrix; performing time series decomposition on the physical constraint feature matrix, and using a dynamic time warping algorithm to perform feature alignment to obtain a time-varying feature matrix; performing feature mapping processing on the time-varying feature matrix to generate a structural state feature vector;

[0009] S3, performing state prediction based on the structural state feature vector and the pre-stored historical state database to obtain a predicted state sequence; performing early warning analysis based on the predicted state sequence and the pre-stored threshold parameter set to obtain an early warning indicator matrix; performing optimization analysis on the early warning indicator matrix to generate an optimized parameter vector;

[0010] S4. Read and extract key indicators and decision rules based on the early warning indicator matrix and optimization parameter vector to generate a decision rule set; generate solutions based on the decision rule set and the pre-stored historical decision library to obtain a candidate solution matrix; evaluate and optimize based on the candidate solution matrix to generate a decision solution sequence.

[0011] Building health monitoring and assessment system based on physical neural network, 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 building health monitoring and assessment method based on physical neural network.

[0015] Beneficial effects: The present invention improves the reliability of evaluation results, retains the correlation between local monitoring data, and makes the diagnosis results conform to the actual physical mechanism; improves the interpretability of the neural network model and the rationality of the decision-making plan, and ensures that the optimization results meet the actual needs of the project; effectively solves the problems existing in traditional methods, such as feature extraction not conforming to physical laws, poor data fusion effect, large deviation of diagnosis results, and poor model interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart 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, this application proposes a building health monitoring and assessment method based on physical neural network, including the following steps:

[0022] S1. Obtain the original sensor data set, perform data preprocessing, and generate a standardized data set; perform feature extraction on the standardized data set to generate a feature vector set; fuse the feature vector set and pre-stored historical feature data to generate a fused feature tensor;

[0023] S2, performing physical constraint processing on the fused feature tensor and the pre-stored physical model parameter set to generate a physical constraint feature matrix; performing time series decomposition on the physical constraint feature matrix, and using a dynamic time warping algorithm to perform feature alignment to obtain a time-varying feature matrix; performing feature mapping processing on the time-varying feature matrix to generate a structural state feature vector;

[0024] S3, performing state prediction based on the structural state feature vector and the pre-stored historical state database to obtain a predicted state sequence; performing early warning analysis based on the predicted state sequence and the pre-stored threshold parameter set to obtain an early warning indicator matrix; performing optimization analysis on the early warning indicator matrix to generate an optimized parameter vector;

[0025] S4. Read and extract key indicators and decision rules based on the early warning indicator matrix and optimization parameter vector to generate a decision rule set; generate solutions based on the decision rule set and the pre-stored historical decision library to obtain a candidate solution matrix; evaluate and optimize based on the candidate solution matrix to generate a decision solution sequence.

[0026] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0027] S11. Collecting original sensor data sets through acceleration sensors, strain sensors, displacement sensors and temperature sensors arranged in the building, using wavelet denoising method to eliminate high-frequency noise in the original sensor data sets to obtain denoised data sets; interpolating missing values ​​of the denoised data sets, using cubic spline interpolation method to fill data gaps to obtain a complete data set; normalizing the complete data set to obtain a standardized data set;

[0028] S12, using a sliding time window method, dividing the standardized data set into predetermined time series segments to obtain a time series segment set; calculating the time domain features and frequency domain features of each segment in the time series segment set to obtain a first feature subset; using an autoencoder to perform dimensionality reduction processing on the time series segment set, extracting potential feature representations, and obtaining a second feature subset; combining the first feature subset and the second feature subset to form a feature vector set; wherein the time domain features include mean, variance, and peak; and the frequency domain features include power spectral density and main frequency components;

[0029] S13. Using the attention mechanism, calculate the correlation weights between each feature in the feature vector set and the historical features in the pre-stored historical feature data to obtain a first weight matrix; based on the first weight matrix, perform weighted fusion on the feature vector set and the historical feature data to generate initial fusion features; use the tensor decomposition method to reduce the rank of the initial fusion features, extract feature patterns, and finally obtain a fusion feature tensor. Feature patterns are also called main feature patterns or significant feature patterns, including features with high energy concentration, features with large information entropy values, and components with high variance contribution rates.

[0030] Significance index SI(f) = w1E(f) + w2H(f) + w3V(f); where: energy ratio E(f) = ||f|| 2 / ∑||fi|| 2 ; Information entropy H(f) = -∑p(f)log(p(f)); Variance contribution V(f) = var(f) / ∑var(fi); w1, w2, w3 are weight coefficients, f is a signal or feature; The specific feature pattern types mainly include time domain features, frequency domain features, time-frequency features and morphological features, among which the time domain features include: peak value, mean, standard deviation, skewness, kurtosis; frequency domain features include: main frequency, frequency band energy, number of spectral peaks; time-frequency features include: wavelet coefficients, Hilbert spectrum; morphological features include: waveform factor, pulse factor, margin factor.

[0031] In one embodiment of the present application, an original data set D = {d1, d2, ..., dn} is obtained from n sensor nodes, and each data point di contains: acceleration (ai), strain (εi), inclination (θi), temperature (ti) and humidity (hi); a timestamp τi and spatial position coordinates (xi, yi, zi) are added to each data point. A feature matrix F = [f1, f2, ..., fm] is constructed, where fi represents the i-th feature vector; the topological entropy weight coefficient is calculated: ω(i) = -∑(pij * ln(pij)) / ln(m), where pij is the normalized value of the i-th feature in the j-th time window, and a dynamic feature vector is generated: V(t) = ∑(ω(i) * fi(t)). Construct the environmental impact matrix E = [e1, e2, ..., ek], calculate the compensation coefficient: α(t) = μ(t)*exp(-β(t)*||E(t)-E(t-1)||), where μ(t) is the dynamic learning rate, β(t) is the environmental sensitivity parameter, and generate compensated data: D'(t) = D(t) *α(t). Construct the evaluation index set Q = {q1, q2, ..., qr}, calculate the hierarchical information entropy: H(l) = -∑(wi *qi * log(qi)), where l represents the evaluation level, wi is the weight of each indicator, and output the data reliability score: R = ∏H(l) / max(H), where ∏ is the product operation.

[0032] In another embodiment of the present application, the wavelet denoising method is specifically: S(t) =∑(j=1→J)[∑(k)φj,k·δ(|dj,k|,λj)·ψj,k(t)]+∑(k)cJ,k·φJ,k(t); wherein S(t) is the reconstructed denoised signal; δ(x,λ)=sign(x)(|x|-λ)·I(|x|>λ) is an improved soft threshold function; λj= σj·sqrt(2log(N))·(1+α·|▽ 2 dj,k|) is the adaptive threshold; dj,k is the wavelet coefficient at position k of scale j; ψj,k(t) is the wavelet basis function at position k of scale j; φj,k is the scale function at position k of scale j; cJ,k is the approximate coefficient at position k under the coarsest scale J; J is the maximum number of layers of wavelet decomposition; N is the length of the original signal; σj is the standard deviation of the noise at the jth layer; α is the gradient weight coefficient; ▽ 2 represents the second-order difference operator; I(·) is the indicator function; t is the time variable; j is the scale index; k is the position index; sign(x) is the sign function; this implementation improves the denoising effect by introducing the second-order gradient information to adaptively adjust the threshold.

[0033] The specific cubic spline interpolation method is: S(x)=∑(i=1→n)[ai(x-xi)3 +bi(x-xi) 2 +ci(x-xi)+di]·I(x∈[xi,x(i+1)]); where ai=(S(i+1)-Si) / (6hi)-hiMi / 6; bi=Mi / 2; ci=[(y(i+1)-yi) / hi]-[(2Mi+M(i+1))·hi] / 6; di=yi; Mi is the second-order derivative value at node xi, which is solved by the equation system AM=B: A is a tridiagonal matrix, Aii=2, A(i, i+1)=A(i+1, i)=μi, μi=hi / (hi+h(i+1)); B is the right-hand side vector, Bi=(6 / h(i+1))[(y(i+2)-y(i+1)) / h(i+1) - (y(i+1)-yi) / hi]; S(x) is the interpolation function; x is the independent variable; xi is the horizontal coordinate of the known data point; yi is the vertical coordinate of the known data point; hi=x(i+1)-xi is the node spacing; n is the number of data points; I(·) is the indicative function; This embodiment improves the smoothness of the interpolation by optimizing the second-order derivative constraints at the nodes.

[0034] The sliding time window method is specifically as follows: W(t)=W0·exp[-β·V(t)]; V(t)=|∑(i=tL→t)[▽X(i)·▽X(i-1)]| / L; wherein W(t) is the dynamic window length at time t; W0 is the reference window length; V(t) is the signal change rate index; X(t) is the original signal; ▽X(t) is the first-order difference of the signal; β is the sensitivity coefficient; L is the length of the local calculation interval; t is the time variable; i is the time index; exp(·) is an exponential function; this embodiment adaptively adjusts the window size according to the local change characteristics of the signal, thereby improving the adaptability of feature extraction.

[0035] This embodiment achieves high-quality feature extraction through multi-level data processing and fusion. First, the wavelet denoising method can effectively eliminate high-frequency noise in sensor data while retaining the effective components of the signal; the cubic spline interpolation method maintains the continuity and smoothness of the data when filling data gaps, avoiding the mutation caused by the traditional linear interpolation method; the normalization process maps different types of sensor data to a unified interval, solving the problem of inconsistent data scales; in the feature extraction link, the sliding time window method achieves effective segmentation of long sequence data, which is convenient for capturing local features; by calculating the dual features of time domain and frequency domain, combined with the dimensionality reduction processing of the autoencoder, the physical features of the data are retained and the potential feature representation is extracted; the introduction of the attention mechanism enables the system to adaptively adjust the importance weights of different features, improving the accuracy of feature fusion; tensor decomposition further compresses the data dimension and extracts feature patterns; this multi-level feature extraction and fusion scheme effectively solves the problems of insufficient feature extraction and poor data fusion effect in traditional methods, and provides high-quality feature input for subsequent state evaluation.

[0036] According to one aspect of the present application, step S11 is further:

[0037] S111, acquiring acceleration sensor data, performing time synchronization through the GPS timing system, and obtaining synchronized acceleration data; acquiring strain sensor data, achieving a unified sampling rate through piecewise linear interpolation, and obtaining resampled strain data; acquiring displacement sensor data and temperature sensor data, and packaging the data through a time series compression algorithm to obtain a compressed data packet; based on the synchronized acceleration data, the resampled strain data, and the compressed data packet, a time window segmentation algorithm is used to generate an original sensor data set;

[0038] S112, based on the original sensor data set, decompose it through the db4 wavelet basis function to obtain wavelet coefficients; based on the wavelet coefficients, use the soft threshold function to screen and obtain noise reduction coefficients; based on the noise reduction coefficients, identify abnormal points through the isolation forest algorithm to obtain abnormal markers; based on the noise reduction coefficients and abnormal markers, reconstruct the signal through inverse wavelet transform to generate a noise reduction data set;

[0039] S113. Based on the denoised data set, a sliding window algorithm is used to locate data gaps and obtain missing positions; based on the denoised data set, the correlation between sensors is calculated by the Pearson correlation coefficient to obtain a correlation matrix; based on the denoised data set, the missing positions and the correlation matrix, the low-rank structure of the data is extracted by Tucker decomposition to obtain a core tensor; based on the core tensor, a cubic spline interpolation algorithm is used to fill the missing values ​​to generate a complete data set;

[0040] S114. Based on the complete data set, the data distribution parameters are calculated through adaptive moment estimation; based on the data distribution parameters, the data are mapped to the [-1, 1] interval through maximum and minimum normalization to obtain normalized data; based on the normalized data, the dynamic time warping algorithm is used to align the timing features of different sensors to obtain alignment features; based on the alignment features, the tensor encapsulation algorithm is used to integrate multi-source data to generate a standardized data set.

[0041] In one embodiment of the present application, the process of obtaining the wavelet coefficients is specifically as follows: W(a, b) = (1 / sqrt(a))·∫X(t)·ψ*((tb) / a)dt; ψa, b(t) = Φ(t)·exp[jω0(tb) / a]·exp[-|tb| 2 / (2a 2 )]; wherein W(a, b) is the wavelet coefficient; X(t) is the input signal; ψa, b(t) is the improved Morlet wavelet basis function; Φ(t) is the window function; a is the scale parameter; b is the translation parameter; ω0 is the center frequency; t is the time variable; j is the imaginary unit; exp(·) is the exponential function; * represents the conjugate operation; this embodiment improves the time-frequency resolution by introducing an adaptive window function.

[0042] The specific time series compression algorithm is: Z(t)=φ(W2·ReLU(W1X(t)+b1)+b2); L=||X(t)-X*(t)|| 2 +α·||▽Z(t)||1 + β·TV(Z(t)); TV(Z) =∑|▽Z(t)|; where Z(t) is the compressed representation; X(t) is the input sequence; X*(t) is the reconstructed sequence; W1, W2 are weight matrices; b1, b2 are bias vectors; φ(·) is the activation function; ReLU is the rectified linear unit; L is the loss function; TV(·) is the total variation regularization term; α, β are weight coefficients; ▽ is the gradient operator; ||·||1 is the L1 norm; ||·|| 2 is the L2 norm; t is the time variable; this embodiment combines sparse constraints and total variation regularization to achieve efficient compression.

[0043] The isolation forest method is as follows: s(x, n) = 2 -E(h(x)) / c(n);h(x) = -∑(t∈P)[wt·log2(st / et)];c(n) = 2H(n-1)-(2(n-1) / n);wt = exp(-dt / σ); where s(x,n) is the anomaly score; h(x) is the weighted path length; c(n) is the normalization factor; n is the number of samples; P is the path from the root node to the leaf node; wt is the node weight; st is the number of samples of node t; et is the expected number of samples of node t; dt is the feature distance; σ is the kernel width parameter; H(i) is the i-th harmonic number; exp(·) is the exponential function; log2(·) is the logarithm with base 2. This embodiment introduces a feature weighting mechanism to improve the accuracy of anomaly detection.

[0044] This embodiment realizes high-quality multi-source data acquisition and preprocessing. In the data acquisition link, the GPS timing system is used for time synchronization, which solves the time alignment problem of different sensor data and ensures the time consistency of the data; the piecewise linear interpolation realizes a unified sampling rate and avoids the information loss caused by the traditional sampling method; the time series compression algorithm is used for data packaging, which reduces the storage and transmission overhead while ensuring data integrity; in the noise reduction processing link, the selection of the db4 wavelet basis function not only ensures the calculation efficiency, but also effectively separates the noise and useful signal; the use of the soft threshold function avoids the signal mutation caused by the hard threshold and improves the noise reduction effect; the isolation forest algorithm can effectively identify abnormal points and improve the data quality; the cubic spline interpolation method maintains the continuity and smoothness of the data when filling the data gaps, avoiding the mutation caused by the traditional linear interpolation method; the overall preprocessing process not only ensures the data quality, but also improves the processing efficiency, providing a high-quality data foundation for subsequent analysis.

[0045] According to one aspect of the present application, step S12 is further:

[0046] S121. Based on the standardized data set, the signal is segmented by a Hamming window function to obtain signal segments; based on the signal segments, the window length and the overlap rate are adjusted by an adaptive window algorithm to obtain optimized segments; based on the optimized segments, a three-dimensional data structure is constructed by a tensor rearrangement algorithm to generate a time series segment set;

[0047] S122, based on the time series segment set, calculate the spectrum components by fast Fourier transform to obtain a spectrum matrix; based on the spectrum matrix, use the optimal basis selection algorithm to extract the frequency band energy characteristics; based on the frequency band energy characteristics, use the peak detection algorithm to extract the main frequency component to obtain the frequency domain characteristics;

[0048] S123, based on the time series segment set, calculating the signal analytical form to obtain an analytical signal; based on the analytical signal, using a variational mode decomposition algorithm to extract an intrinsic mode function to obtain a modal feature; based on the modal feature, extracting a time-frequency feature by Hilbert spectrum analysis to obtain a joint feature;

[0049] S124. Based on the time series segment set, extract local features through a one-dimensional convolutional neural network to obtain convolution features; based on the convolution features, use a self-attention algorithm to calculate feature weights to obtain weighted features; based on the weighted features, extract deep feature representations through latent variable encoding to obtain deep features;

[0050] S125. Based on frequency domain features, joint features and deep features, the Lasso regression algorithm is used to screen key features to obtain screened features; based on the screened features, the feature importance is calculated through Fisher discriminant analysis to obtain important features; based on the important features, an adaptive weighting algorithm is used to integrate multi-source features to generate a feature vector set.

[0051] In one embodiment of the present application, the adaptive window method is specifically: W(t)=Wmin+(Wmax-Wmin)·σ(F(t)); F(t)=∑(i=1→k)[αi·fi(t)]; σ(x)=1 / (1+exp(-γx)); fi(t)=||▽ i X(t)||2 / ||X(t)||2; where W(t) is the time window length; F(t) is the feature comprehensive index; σ(·) is the sigmoid function; fi(t) is the i-th order derivative feature; X(t) is the signal sequence; Wmin, Wmax are the window length limits; αi is the feature weight; γ is the sigmoid function parameter; k is the derivative order; ▽ i is the i-th order derivative operator; ||·||2 is the L2 norm; exp(·) is the exponential function; t is the time variable; the window size is adaptively adjusted according to the multi-order derivative characteristics.

[0052] The variational mode decomposition algorithm is as follows: X(t)=∑(i=1→n)IMFi(t)+r(t);IMFi(t)=∑(j=1→m)[cij(t)·exp(2πjωit)];cij(t) = hij(t)·exp[-γi(t-τij) 2]; where X(t) is the original signal; IMFi(t) is the i-th intrinsic mode function; r(t) is the residual term; cij(t) is the j-th component of the i-th IMF; ωi is the center frequency; hij(t) is the amplitude modulation function; γi is the time scale parameter; τij is the time delay parameter; t is the time variable; n is the number of IMFs; m is the number of components of each IMF; exp(·) is an exponential function; adaptive time-frequency analysis of the signal is achieved through multi-scale decomposition.

[0053] The optimal basis selection algorithm is as follows: B* = argmin[∑||X-BΦ|| 2 +λ·R(B)]; R(B) = tr(B T LB); L = DW; Wij = exp(-||xi-xj|| 2 / σ 2 )·M(xi,xj); where B* is the optimal basis matrix; X is the data matrix; Φ is the coefficient matrix; R(B) is the physical constraint term; L is the Laplace matrix; W is the similarity matrix; D is the degree matrix; M(·,·) is the physical model compatibility function; λ is the regularization parameter; σ is the kernel width; tr(·) is the matrix trace; exp(·) is the exponential function; ||·|| is the norm; the basis function selection is optimized through physical model constraints.

[0054] The specific Lasso regression method is: β* = argmin[1 / (2n)||y-Xβ|| 2 2 + λ1||β||1 + λ2||Dβ||2]; D = [▽;▽ 2 ];▽β = [β(i+1)-βi];▽ 2 β = [β(i+2)-2β(i+1)+βi]; where β* is the optimal parameter vector; y is the response variable; X is the feature matrix; β is the regression coefficient; D is the difference operator matrix; λ1, λ2 are regularization parameters; n is the number of samples; ▽ is the first-order difference; ▽ 2 is the second-order difference; ||·||1 is the L1 norm; ||·||2 is the L2 norm; i is the sample index; robust regression is achieved by combining sparsity and smoothness constraints.

[0055] This embodiment achieves a comprehensive characterization of the structural state through multi-dimensional feature extraction; the sliding time window method can adaptively adjust the window size, which not only ensures the time domain locality of the feature, but also avoids the information loss caused by the fixed window; the calculation of time domain features includes statistics such as mean, variance, peak, etc., which can reflect the basic statistical characteristics of the data; the extraction of frequency domain features includes power spectrum density and main frequency components, which effectively captures the dynamic characteristics of the structure; the introduction of the autoencoder realizes nonlinear dimensionality reduction of the data, and its encoding-decoding structure not only compresses the data dimension but also retains key information; the feature vector formed by the combination of feature subsets has a strong expressive ability and can comprehensively reflect the state characteristics of the structure. This embodiment not only improves the expressive ability of the feature, but also reduces the impact of redundant information, providing a reliable feature basis for subsequent state evaluation.

[0056] In another embodiment of the present application, the variational mode decomposition method can also be: min∑(k=1→K)[||Ψt[(δ(t)+j / πt)*uk(t)]·e -jωkt || 2 2]; subject to ∑uk(t)=f(t); ▽ 2 uk(t)≤εk; where uk(t) is the kth component; f(t) is the input signal; ωk is the center frequency; δ(t) is the Dirac function; Ψt is the time derivative operator; * indicates convolution operation; j is the imaginary unit; εk is the smoothness constraint; ▽ 2 is the Laplace operator; e is the base of natural logarithm; π is the circumference of pi; t is the time variable; K is the number of components; ||·||2 is the L2 norm; and the smoothness constraint is introduced to improve the decomposition stability.

[0057] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0058] S21. Using the Hamilton quantity encoding method, the features in the fused feature tensor are mapped to the physical space to obtain the physical space features; the pre-stored physical model parameter set is input into the parameter mapping module to obtain the physical parameter mapping; based on the physical space features and the physical parameter mapping, the constraint conditions are constructed by the Lagrange multiplier method to generate the physical constraint feature matrix; wherein the physical model parameter set includes the mass matrix, the stiffness matrix and the damping matrix;

[0059] S22. Use a recursive neural network to perform time series decomposition on the physical constraint feature matrix to obtain a time series feature sequence; apply a dynamic time warping algorithm to the time series feature sequence to align features of different time scales to obtain an aligned feature sequence; input the aligned feature sequence into the attention layer, calculate the importance weights of the time series features, and generate a weighted feature sequence; perform time series aggregation on the weighted feature sequence to obtain a time-varying feature matrix;

[0060] S23. Extract high-order features in the time-varying feature matrix through a deep residual network; perform dimensionality reduction and normalization on the high-order features to obtain normalized features; map the normalized features to the structural state space to generate a structural state feature vector.

[0061] In one embodiment of the present application, based on the preprocessed data set D'(t) and the reliability score R, a physical constraint matrix P = [p1, p2, ..., ph] is constructed, including the structural stiffness constraint K(t), the mass distribution constraint M(t) and the damping characteristic constraint C(t), and a physical constraint vector is generated: Φ(t) = [K(t), M(t), C(t)]. Based on the physical constraint vector Φ(t), the node association is calculated: γij = exp(-||Φi -Φj|| / σ 2 ), where σ is a dynamic adjustment parameter; construct the adjacency matrix A(t): when γij> preset threshold, aij =γij; output dynamic topology structure T(t) = {V(t), A(t)}, where V(t) is a node set and A(t) is an adjacency matrix. Based on the dynamic topology structure T(t), construct a hybrid neuron: N(t)=λ1Np(t)+λ2Nd(t), where Np(t) is the physical model output, Nd(t) is the data-driven model output, and λ1 and λ2 are adaptive weight coefficients; calculate the neuron activation function: f(N(t)) = tanh(N(t)) * g(Φ(t)), where g(Φ(t)) is the physical constraint modulation function. Based on the mixed neuron group N(t), construct the inter-layer connection strength matrix W(t): wij(t) = ρ(t) * exp(-||Ni(t) - Nj(t)||), where ρ(t) is the dynamic connection coefficient; construct the optimization objective function: L(t) = ||f(N(t)) - Y(t)|| 2 + η||W(t)||1, where Y(t) is the expected output, η is the regularization parameter, and the output is the optimized network structure G(t) = {N(t), W(t)}. Based on the network structure G(t), the performance index set is calculated: physical consistency index: δp(t)=||Φ(t)-Φ'(t)||; prediction accuracy index: δa(t)= RMSE(Y(t), Y*(t)), where RMSE is the root mean square error, and Y*(t) is the predicted output of the model; calculation efficiency index: δe(t) = t_compute / t_threshold, where t_compute is the calculation time, and t_threshold is the calculation time threshold; output comprehensive performance score S(t) = w1δp(t) + w2δa(t) + w3*δe(t), where w1, w2 and w3 are weight coefficients.

[0062] In another embodiment of the present application, an energy constraint mapping function Φ(x, p, E)=∑(ωi·Hi(x, p))·exp(-μ·|E-E0| 2 ), where Hi is the ith Hamiltonian subsystem, Hi(x, p) = p T M -1 p / 2 + x T Kix / 2 +Vi(x); E is the total energy of the system; E0 is the reference energy; ωi is the subsystem weight; μ is the energy deviation penalty factor; x is the generalized coordinate vector; p is the generalized momentum vector; M is the mass matrix; Ki is the i-th stiffness matrix; Vi is the i-th nonlinear potential energy term; i is the subsystem index, i∈[1,n]; n is the total number of subsystems; exp is the exponential function.

[0063] The parameter mapping module algorithm is as follows: construct the parameter mapping function Ψ(θ, Γ) = A·σ(W·θ + b)·exp(-λ·‖Γ(θ)‖ 2 ), where θ is the physical parameter vector; Γ is the parameter constraint matrix; A is the mapping matrix; W is the weight matrix; b is the bias vector; σ is the activation function, σ(z) = tanh(z)·(1-exp(-z 2 )); λ is the constraint strength coefficient; ‖·‖ represents the matrix norm; exp is the exponential function; tanh is the hyperbolic tangent function.

[0064] Construct the constraint fusion function C(f, g, α) = ∑(πi·fi(x))·∏(1-exp(-αj·gj(x))); where f is the feature vector function; g is the constraint vector function; πi is the feature weight; αj is the constraint strength coefficient; x is the state variable; fi is the i-th feature component; gj is the j-th constraint component; i is the feature index, i∈[1,m]; j is the constraint index, j∈[1,k]; m is the feature dimension; k is the constraint dimension; exp is the exponential function.

[0065] Construct feature learning function L(s, t, Ω) = ∑(βi·φi(s, t))·exp(-γ‖▽Ω(s)‖ 2 );where s is the state feature; t is the time variable; Ω is the dynamic operator; βi is the learning weight; φi is the dynamic basis function, φi(s, t) = si·cos(ωit); γ is the gradient penalty factor; ▽ is the gradient operator; si is the i-th state component; ωi is the i-th natural frequency; i is the modal index, i∈[1,r]; r is the modal number; exp is the exponential function; cos is the cosine function.

[0066] Construct the knowledge distillation function D(y, y', Φ) = α·KL(y‖y') + β·‖Φ(y)-Φ(y')‖ 2 + η·R(Φ); where y is the teacher model output; y' is the student model output; Φ is the physical operator; KL is the KL divergence; α is the soft label weight; β is the physical consistency weight; η is the regularization coefficient; R is the regularization term, R(Φ) = tr(Φ T Φ); tr is the matrix trace; ‖·‖ represents the norm.

[0067] The structural constraint optimization algorithm is as follows: construct the constraint optimization function O(w, λ, ε) = f(w) + ∑(λi·gi(w))·exp(-ε·‖h(w)‖ 2 ), where w is the optimization variable; λ is the Lagrange multiplier vector; ε is the penalty factor; f is the objective function; g is the inequality constraint vector; h is the equality constraint vector; gi is the i-th inequality constraint; λi is the i-th multiplier; i is the constraint index, i∈[1,p]; p is the number of constraints; ‖·‖ represents the norm; and exp is the exponential function.

[0068] Construct the residual mapping function R(x, θ, Φ) = x + F(x, θ)·exp(-η·‖Φ(x)‖ 2 );F(x,θ)= W2·σ(W1x + b1) + b2;Φ(x) is the physical constraint operator;x is the input feature;θ is the network parameter set, θ={W1,W2,b1,b2};W1,W2 are weight matrices;b1,b2 are bias vectors;η is the physical constraint coefficient;σ is the activation function, σ(z) =max(0,z) + α·min(0,z); α is the negative slope parameter;‖·‖ represents the norm;exp is the exponential function.

[0069] The specific process of the dynamic time warping algorithm is: time warping distance DTW(X, Y, ω) = min{∑(d(xi, yj)·ω(i, j))·exp(-λ·|E(xi)-E(yj)|)}; where X, Y are time series; d(xi, yj) is the local distance metric, d(xi, yj) = ‖xi-yj‖ 2 ; ω(i, j) is the path weight; E(x) is the energy function; xi is the i-th point in X; yj is the j-th point in Y; λ is the energy difference penalty factor; i, j are time indexes; ‖·‖ represents the Euclidean distance; exp is the exponential function.

[0070] This embodiment introduces physical constraints into the neural network learning process; maps features to physical space through Hamiltonian encoding, ensuring that feature representation satisfies the law of conservation of energy; the physical parameter mapping module embeds physical parameters such as mass matrix and stiffness matrix into the network structure, enhancing the interpretability of the model; the constraints constructed by the Lagrange multiplier method ensure that the learning process conforms to physical laws; in the dynamic feature learning link, the recursive neural network realizes the effective decomposition of time series features, and the dynamic time warping algorithm solves the alignment problem of features of different time scales; the introduction of the attention mechanism enables the system to adaptively identify important time series features; the combination of deep residual network and adaptive pooling not only extracts high-order features but also realizes dimensionality reduction and normalization of features. This embodiment effectively solves the problem of unreliable features caused by traditional neural networks ignoring physical constraints, and improves the accuracy and interpretability of feature learning.

[0071] According to one aspect of the present application, step S21 is further:

[0072] S211. Based on the fused feature tensor, the energy conservation characteristics are maintained through symplectic geometry to obtain energy characteristics; based on the energy characteristics, the motion equation is constructed through the variational principle to obtain dynamic characteristics; based on the dynamic characteristics, the physical system invariants are maintained through symplectic transformation to generate physical space characteristics;

[0073] S212, based on the pre-stored physical model parameter set, maintaining the positive definiteness of the parameters through Lie algebra to obtain a first embedded feature; based on the first embedded feature, encoding the parameter symmetry through geometric algebra to obtain a constraint feature; based on the constraint feature, constructing the parameter correlation through an energy function to generate a physical parameter mapping;

[0074] S213, based on the physical space characteristics and physical parameter mapping, the numerical stability is maintained through the symplectic integral network to obtain the consistency characteristics; based on the consistency characteristics, the constraint conditions are constructed through the variational inequality; based on the constraint conditions, the augmented Lagrangian method is used to solve the optimal multiplier and generate the constraint multiplier;

[0075] S214. Based on the constraint multiplier, the activation function is modified to satisfy the energy dissipation law to obtain the neuron feature; based on the neuron feature, the physical topological structure is maintained through the graph neural network to obtain the topological feature; based on the topological feature, the feature propagation law is modeled through the wave equation to generate the propagation feature;

[0076] S215. Based on the propagation characteristics, the physical constraint loss is constructed through the energy functional to obtain the loss characteristics; based on the loss characteristics, the constraint parameters are optimized through the KKT conditions to obtain the optimization characteristics; based on the optimization characteristics, the physical characteristics are reconstructed through the Hamiltonian neural network to generate the physical constraint characteristic matrix.

[0077] In one embodiment of the present application, the process of generating physical space features is specifically as follows: physical feature mapping function H(x, p, t) = ΨS / Ψx·dp / dt - ΨS / Ψp·dx / dt + ΨS / Ψt; where S is the action functional, S=∫L(x, x*, t)dt; L is the Lagrangian, L = T - V; T is the system kinetic energy, T = (1 / 2)·x* T M x*; V is the potential energy of the system, V = (1 / 2)·x T Kx; M is the mass matrix; K is the stiffness matrix; x is the generalized coordinate vector, x* is the generalized velocity vector; p is the generalized momentum vector, p = ΨL / Ψx*; t is the time variable; Ψ represents the partial derivative operator; dx / dt represents the total derivative of x with respect to t; dp / dt represents the total derivative of p with respect to t.

[0078] The specific constraint optimization algorithm of the symplectic integral network is as follows: construct the constraint optimization function C(z, λ, t) = H(z, t) + λ T g(z); where H(z, t) is the system Hamiltonian, H(z, t) = p T M -1 p / 2 + x T Kx / 2; z is the phase space variable, z = [x;p]; g(z) is the constraint condition vector; λ is the Lagrange multiplier vector; M is the mass matrix; K is the stiffness matrix; x is the displacement vector; p is the momentum vector; t is the time variable; the phase space evolution equation dz / dt = J▽H(z, t) + G(z)λ; J is the symplectic matrix, J = [0, I; -I, 0]; G(z) is the constraint Jacobian matrix, G(z) = Ψg / Ψz; ▽ represents the gradient operator.

[0079] The modified activation function satisfies the energy dissipation law as follows: Construct the energy dissipation activation function σ(x, E) = α·tanh(βx)·exp(-γE) + δ·x; where E is the total energy of the system, E = T + V; T is the kinetic energy, T = (1 / 2)·x' T M x'; V is potential energy, V = (1 / 2)·x T Kx; α is the amplification factor; β is the scaling factor; γ is the dissipation coefficient; δ is the linear correction term; x is the input feature; tanh is the hyperbolic tangent function; exp is the exponential function; M is the mass matrix; K is the stiffness matrix; x' is the velocity vector.

[0080] This embodiment achieves a deep integration of physical constraints and feature learning. Mapping features to physical space ensures that feature representation satisfies basic physical laws such as energy conservation; the introduction of physical model parameter sets, including mass matrix, stiffness matrix, damping matrix, etc., makes the feature learning process have clear physical meaning; the parameter mapping module establishes the correspondence between physical parameters and feature space, enhancing the interpretability of the model; the constraints constructed by the Lagrange multiplier method ensure that the feature learning process conforms to physical laws and avoids feature representations that violate physical laws. This embodiment provides prior knowledge for neural networks, reduces data requirements, improves the generalization ability of the model, effectively solves the problem of unreliable features caused by ignoring physical constraints in traditional feature learning, and improves the accuracy and interpretability of feature learning.

[0081] According to one aspect of the present application, step S22 is further:

[0082] S221, input the physical constraint feature matrix C into the variational Bayes module, decompose the temporal dependency through the probabilistic graphical model, and obtain the probability distribution G1; input the probability distribution G1 into the gated recurrent unit, extract the sequence pattern through the long short-term memory network, and obtain the recurrent feature G2; input the recurrent feature G2 into the bidirectional recurrent layer, extract the context information through forward and backward propagation, and generate the temporal feature sequence Q1;

[0083] S222, input the time series feature sequence Q1 into the subsequence decomposition module, extract multi-scale features through the sliding window algorithm, and obtain the scale sequence H1; input the scale sequence H1 into the similarity calculation module, calculate the sequence distance through the Sakoe-Chiba band constraint, and obtain the distance matrix H2; input the distance matrix H2 into the path optimization module, solve the optimal alignment path through the dynamic programming algorithm, and generate the alignment feature sequence Q2;

[0084] S223, input the aligned feature sequence Q2 into the multi-head attention module, calculate the self-attention score by scaling the dot product, and obtain the attention score J1; input the attention score J1 into the weight learning module, optimize the feature weights through the residual connection network, and obtain the learning weight J2; input Q2 and the learning weight J2 into the feature weighting module, calculate the weighted features through adaptive normalization, and generate the weighted feature sequence Q3;

[0085] S224, input the weighted feature sequence Q3 into the spatiotemporal coding module, embed the temporal position information through the position coding algorithm, and obtain the coding feature K1; input the coding feature K1 into the feature aggregation module, model the relationship between features through the graph convolution network, and obtain the aggregated feature K2; input the aggregated feature K2 into the temporal fusion module, fuse the multi-layer features through the jump connection, and generate the fused feature K3;

[0086] S225. Input the fused feature K3 into the nonlinear mapping module, extract the high-order feature relationship through kernel function transformation, and obtain the mapping feature L1; input the mapping feature L1 into the dimensionality reduction optimization module, maintain the feature manifold structure through the t-SNE algorithm, and obtain the reduced dimensionality feature L2; input the reduced dimensionality feature L2 into the feature reconstruction module, reconstruct the time-varying feature representation through tensor decomposition, and obtain the time-varying feature matrix D.

[0087] In one embodiment of the present application, the position encoding algorithm is specifically: PE(t, i) = p(t)·∑(k=1→d)[wk·sin(ωkt+φk, i)]; p(t) = sigmoid(Wp·h(t)); ωk = 1 / (10000 2k / d );PE(t,i) is the position encoding vector; p(t) is the adaptive weight function; h(t) is the hidden state vector; Wp is the weight matrix; wk is the frequency weight; ωk is the angular frequency; φk,i is the phase shift; t is the time variable; i is the feature dimension; d is the encoding dimension; k is the frequency index; sigmoid(·) is the activation function; sin(·) is the sine function; the temporal feature expression is enhanced through the hierarchical structure.

[0088] This embodiment achieves efficient dynamic feature learning. Temporal decomposition of physical constraint features can effectively capture the temporal dependencies of features; dynamic time warping solves the alignment problem of features at different time scales and improves the temporal consistency of features; the introduction of the attention mechanism enables the system to adaptively identify important temporal features and improves the pertinence of feature learning; the aggregation process of temporal features retains key dynamic information while reducing data dimensions. This embodiment can not only accurately capture the time-varying characteristics of structural states, but also has strong adaptive capabilities, providing reliable dynamic feature support for structural state assessment.

[0089] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0090] S31, using an empirical mode decomposition method to decompose the structural state feature vector into a predetermined number of intrinsic mode functions to obtain a mode sequence; based on the mode sequence, extracting the time-varying trend of each mode through wavelet transform to obtain trend characteristics; based on a pre-stored historical state database, matching the trend characteristics through a dynamic time warping algorithm to generate a matching feature sequence;

[0091] S32. Based on the matching feature sequence, a variational autoencoder is used to extract the potential feature representation to obtain the encoding feature; based on the encoding feature, the feature dependency relationship of different time scales is calculated to obtain the attention feature; based on the attention feature, a multi-step prediction result is generated through a long short-term memory network to obtain a prediction state sequence;

[0092] S33, based on the predicted state sequence, calculating the statistical characteristics of each state indicator to obtain a statistical characteristic matrix; based on the pre-stored threshold parameter set, generating a dynamic threshold boundary through fuzzy mapping to obtain a threshold boundary set; based on the statistical characteristic matrix and the threshold boundary set, calculating the warning probability through a probability density estimation model to obtain a warning indicator matrix;

[0093] S34. Based on the early warning indicator matrix, the Bayesian optimization method is used to adjust the early warning model parameters to obtain the optimized parameter set; based on the optimized parameter set, the parameter performance is evaluated through cross-validation to obtain the verification results; based on the verification results, the adaptive learning algorithm is used to update the model parameters to generate the optimized parameter vector.

[0094] In one embodiment of the present application, based on the network structure G(t) and the performance score S(t), a state feature tensor Ψ(t) is constructed: displacement feature vector: u(t) = [u1(t), u2(t), ..., u m (t)]; stress eigenvector: σ(t) = [σ1(t), σ2(t), ..., σ m (t)]; vibration eigenvector: v(t) = [v1(t), v2(t), ..., v m (t)]; output state feature set F(t) = {u(t), σ(t), v(t)}. Based on the state feature set F(t), construct multi-layer health indicators: local indicator: IL(t) = ∑(βi*fi(t)), fi∈F(t); regional indicator: IR(t) = φ(IL1, IL2, ..., IL k ); Overall index: IG(t)= ψ(IR1,IR2,...,IR m ), where φ and ψ are regional and overall aggregation functions; output health indicator system H(t) = {IL(t), IR(t), IG(t)}. Based on the health indicator system H(t) and historical data series, construct the state transition matrix: Π(t)=[π 11 (t) ... π 1n (t)] [ ... ... ... ] [π n1 (t) ... π nn (t)], where π ij Represents the probability of state i transferring to state j; predicts the state at time future_t: H(t+future_t) = H(t) * Π(t) future_t, output the predicted state sequence P(t) = {H(t+1), H(t+2), ..., H(t+k)}. Based on the current state H(t) and the predicted state sequence P(t), construct the risk assessment matrix: R(t) = [r 11 (t) ... r 1m (t)] [ ... ... ... ] [r n1 (t) ... r nm (t)], where r ij Represents the risk value of the jth category in the ith area; calculates the warning level: W(t) = μ(R(t)) * ν(P(t)) * ξ(H(t)), where μ, ν, ξ are risk weight functions; outputs the warning information package W_info(t) = {level, location, type, recommendation}. Based on the warning information package W_info(t) and the actual status data, the warning accuracy is calculated: true positive rate: TPR = TP / (TP+FN); false positive rate: FPR = FP / (FP+TN); warning timeliness: τ=t_warning - t_event; optimize the warning parameters: θnew = θold - α▽L(θ), where L(θ) is the warning loss function; output the optimized warning parameters θnew and performance report, where TP is true positive, FN is false negative, FP is false positive, TN is true negative, t_warning is the warning time, t_event is the event time, θold is the old warning parameter, α is the learning rate, and ▽ is the gradient.

[0095] In another embodiment of the present application, the optimization algorithm of the multi-step prediction result is specifically as follows: construct a prediction function P(X, h, Ω) = ∑(αk·fk(X))·exp(-ρ·‖Ω(X")-Y‖ 2 );X is the historical data; h is the prediction step; Ω is the dynamic constraint; fk is the k-step predictor; X" is the prediction result; Y is the true value; αk is the time weight, αk = exp(-k / h); ρ is the constraint violation penalty factor; k is the number of prediction steps, k∈[1,h]; ‖·‖ represents the prediction error norm; exp is the exponential function.

[0096] The Bayesian optimization method is as follows: p(y|x, D) = N(μ(x), σ 2 (x)); μ(x) = k(x) T (K+σn 2 I) - ¹y;σ 2 (x) = k(x, x) - k(x) T (K+σn 2 I) -¹k(x);k(xi,xj) = σf 2 ·exp[-0.5(xi-xj) T Λ - ¹(xi-xj)]·R(xi,xj); where p(y|x,D) is the posterior distribution; μ(x) is the predicted mean; σ 2 (x) is the prediction variance; k(·,·) is the improved kernel function; K is the kernel matrix; y is the observation vector; x is the input variable; D is the historical data; σf 2 is the signal variance; σn 2 is the noise variance; Λ is the characteristic scale matrix; R(·,·) is the physical constraint related function; I is the unit matrix; N(·,·) represents Gaussian distribution; the introduction of physical constraints improves the optimization efficiency.

[0097] This embodiment achieves accurate state prediction and early warning analysis. The empirical mode decomposition method decomposes the state characteristics into multiple intrinsic mode functions, effectively capturing dynamic characteristics of different scales; the wavelet transform extracts the time-varying trend of each mode and retains the key information of state evolution; the dynamic time warping algorithm achieves accurate matching with historical data; the potential feature representation extracted by the variational autoencoder has strong expressive power, and the multi-head attention mechanism enhances the capture of long-term dependencies; the multi-step prediction results generated by the long short-term memory network have high accuracy; in the early warning analysis link, the dynamic threshold boundary generated by the fuzzy mapping is highly adaptable, and the probability density estimation provides a reliable early warning probability; the Bayesian optimization method optimizes the early warning model parameters, cross-validation ensures the validity of the parameters, and the adaptive learning algorithm realizes the dynamic update of the parameters. This embodiment not only has high prediction accuracy, but also has strong adaptability and robustness.

[0098] According to one aspect of the present application, step S32 is further:

[0099] S321, based on the matching feature sequence, estimate the feature distribution through the Gaussian mixture model to obtain the distribution parameters; based on the distribution parameters, generate latent variables through random sampling to obtain sampling features; based on the sampling features, optimize the coding representation through KL divergence to generate coding features;

[0100] S322, based on the encoding features, construct a query matrix through linear projection to obtain query features; based on the query features, calculate the attention weights through dot product operations to obtain a second weight matrix; based on the second weight matrix, capture multi-scale dependencies through a parallel attention mechanism to generate attention features;

[0101] S323, based on the attention feature, injecting sequence order information through position encoding to obtain a second embedded feature; based on the second embedded feature, controlling the information flow through the GRU unit to obtain an updated feature; based on the updated feature, storing long-term dependencies through an external memory network to generate an enhanced feature;

[0102] S324, based on the enhanced features, generating an initial prediction through a transformer decoder; based on the initial prediction, optimizing the prediction accuracy through residual learning to obtain a revised prediction; based on the revised prediction, generating a multi-step sequence through cyclic prediction to obtain a predicted state sequence;

[0103] S325. Based on the prediction state sequence, the prediction confidence interval is estimated by Monte Carlo sampling to obtain the confidence boundary; based on the confidence boundary, the prediction stability is evaluated by the Lyapunov index to obtain the stability index; based on the prediction state sequence and the stability index, the prediction results are optimized by Bayesian posterior inference to obtain the optimized prediction state sequence.

[0104] In one embodiment of the present application, the process of capturing multi-scale dependency is specifically as follows: performing multi-scale feature decomposition F(s, ψ) = ∑(Wi·φi(s))·exp(-|s-si| 2 / 2ψi 2 ), where s is the feature vector; Wi is the scale weight coefficient; φi is the basis function, φi(s) = cos(πs / λi); ψi is the scale parameter; λi is the characteristic wavelength; si is the reference feature point; i is the scale level index, i∈[1,n]; n is the total number of scales; exp is the exponential function; |s-si| represents the characteristic distance.

[0105] The specific optimization of prediction accuracy through residual learning is as follows: construct the error compensation function E(r, θ) = ∑(αi·ri)·exp(-βi·θi) + γ·▽ 2 r; where r is the residual vector; θ is the physical constraint vector; αi is the compensation coefficient; βi is the attenuation coefficient; γ is the smoothing factor; ▽ 2 is the Laplace operator; ri is the i-th residual component; θi is the i-th constraint component; i is the component index, i∈[1,m]; m is the feature dimension; exp is the exponential function.

[0106] This embodiment constructs an accurate state prediction model. The extracted latent feature representation has a strong probabilistic interpretability and can characterize the uncertainty of the state; by calculating the feature dependencies at different time scales, the ability to capture long-term dependencies is enhanced; the gradient vanishing problem in traditional recurrent neural networks is solved, and the accuracy of long sequence prediction is improved; the generation of multi-step prediction results takes into account the continuity and smoothness of state evolution, avoiding sudden changes in prediction results. This embodiment not only ensures the accuracy of the prediction, but also provides uncertainty estimation of the prediction results, providing a reliable prediction basis for early warning analysis.

[0107] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0108] S41. Based on the early warning indicator matrix, the principal component analysis method is used to extract key indicators to obtain a key indicator set; based on the optimized parameter vector, a mapping relationship between parameters and decision rules is constructed through a deep neural network to obtain mapping features; based on the key indicator set and mapping features, a decision tree algorithm is used to construct hierarchical decision rules to generate a decision rule set;

[0109] S42, based on the decision rule set, using fuzzy reasoning method to analyze the decision rules and obtain reasoning results; based on the pre-stored historical decision library, using cosine similarity to calculate the degree of matching with the reasoning results, to obtain a similarity vector; based on the reasoning results and the similarity vector, using reinforcement learning algorithm to obtain a candidate solution matrix;

[0110] S43. Based on the candidate solution matrix, the feasibility of the solution is verified by Lagrangian relaxation method to obtain a feasible solution set; based on the feasible solution set, the implementation cost of the solution is calculated by multi-objective optimization method to obtain a cost matrix; based on the feasible solution set and the cost matrix, the hierarchical analysis method is used for comprehensive sorting to generate sorting results;

[0111] S44. Based on the sorting results, a genetic algorithm is used for optimization and adjustment to obtain an optimization solution set; based on the optimization solution set, a dynamic programming algorithm is used to generate a solution implementation sequence to obtain an implementation sequence; based on the implementation sequence, a multi-agent reinforcement learning algorithm is used to optimize the synergy of the solutions to generate a decision-making solution sequence.

[0112] In one embodiment of the present application, based on the warning information package W_info(t) and the optimization parameter θnew, an evaluation index system is constructed: safety index: S(t) = ∑(αi * si(t)), where si(t) is the safety metric of each subsystem; reliability index: R(t) = exp(-∫λ(τ)dτ), where λ(τ) is the failure rate function; economic index: E(t) = Cp(t) + Cm(t) + Cr(t), where Cp is the prevention cost, Cm is the maintenance cost, and Cr is the repair cost; output comprehensive evaluation vector V(t) = [S(t), R(t), E(t)]. Based on the comprehensive evaluation vector V(t), a fuzzy cognitive relationship matrix is ​​constructed: FCM(t) = [c 11 (t)... c 1n (t)] [ ... ... ... ] [c n1 (t) ... c nn (t)], where c ij Represents the influence of concept i on concept j; calculates the decision state evolution: X(t+1) = f(FCM(t) * X(t)), where f is a nonlinear threshold function; outputs the decision state sequence D(t) = {X(t), X(t+1), ..., X(t+k)}. Based on the decision state sequence D(t), constructs the maintenance activity set: M(t) = {m1(t), m2(t), ..., m k (t)}; where m i (t) includes: maintenance type, execution time, resource requirements and expected results. Optimize the maintenance plan: Mopt(t) = argmin{∑(ωi * Ci(M(t)))}, where Ci is a cost function of different types; output the optimal maintenance plan Mopt(t). Based on the optimal maintenance plan Mopt(t), construct the resource requirement matrix: Q(t) = [q 11 (t) ... q 1m (t)] [ ... ... ... ] [q n1 (t) ... q nm (t)], where q ijRepresents the demand of activity i for resource j; calculates the optimal allocation plan: A*(t) = optimize(Q(t), B(t), C(t)), where B(t) is the resource constraint and C(t) is the cost constraint; outputs the resource allocation plan A*(t). Based on the resource allocation scheme A*(t) and maintenance implementation data, the implementation effect indicators are calculated: plan completion rate: PCR(t) = Nc / Np; resource utilization rate: RUR(t) = Ru / Ra; maintenance effect index: MEI(t) = ΔH(t) / ΔC(t); generate feedback optimization parameters: β(t+1) = β(t) + η▽J(β), where J(β) is the effect evaluation function; output the implementation evaluation report E_report(t) and optimization parameters β(t+1), where Nc is the number of tasks planned to be completed, Np is the total number of planned tasks, Ru is the actual amount of resources used, Ra is the total amount of available resources, ΔH(t) is the change in health status, ΔC(t) is the change in cost, β(t) is the parameter before optimization, and η is the learning rate.

[0113] In another embodiment of the present application, the process of the reinforcement learning algorithm is specifically: constructing an action value function Q(s, a, π) = r(s, a) +γ·E[V(s')]·exp(-β·c(s, a)); wherein V(s) = max{Q(s, a)}; r(s, a) is the immediate reward; c(s, a) is the constraint violation degree; s is the state vector; a is the action vector; π is the strategy function; γ is the discount factor; β is the constraint penalty coefficient; E is the expected operator; s' is the next state; state transition probability P(s'|s, a); action generation probability π(a|s) = exp(Q(s, a) / τ) / ∑exp(Q(s, a') / τ); τ is the temperature parameter.

[0114] The decision tree algorithm is as follows: G(S, A) = H(S)-∑(v∈V)[|Sv| / |S|·H(Sv)]·P(A); H(S) = -∑(c∈C)[p(c)·log(p(c))]; P(A) = exp[-λ·|▽ 2 A| / max(|▽ 2 A|)]; where G(S, A) is the gain function; H(S) is the information entropy; S is the sample set; A is the segmentation attribute; Sv is the subset segmented by attribute A; V is the attribute value set; C is the category set; p(c) is the probability of category c; P(A) is the physical constraint penalty term; λ is the penalty coefficient; ▽ 2 is a second-order difference operator; |S| represents the cardinality of the set S; exp(·) is an exponential function; log(·) is a natural logarithm; and feature selection is optimized by introducing a physical constraint penalty term.

[0115] The dynamic programming algorithm is as follows: V(s, t) = min[c(s, a) + γ·∑P(s'|s, a)·V(s', t+1)]; c(s, a) = ∑(i=1→m)[wi·ci(s, a)]·R(s, a); R(s, a) = exp[-λ·||▽ 2 s|| 2 ]; where V(s, t) is the value function; c(s, a) is the comprehensive cost function; ci(s, a) is the cost of the i-th target; P(s'|s, a) is the state transition probability; s, s' are state variables; a is the action variable; wi is the target weight; R(s, a) is the physical constraint term; γ is the discount factor; λ is the constraint coefficient; ▽ 2 is the Laplace operator; t is the time step; m is the number of targets; exp(·) is the exponential function; and multi-objective decision making is optimized through physical constraints.

[0116] This embodiment constructs a complete multi-level decision support system. The principal component analysis method extracts key warning indicators and reduces the complexity of decision-making; the mapping relationship between parameters and decision rules established by the deep neural network has strong generalization ability; the hierarchical decision rules constructed by the decision tree algorithm are clear and intuitive; the fuzzy reasoning method provides a flexible rule parsing mechanism, and the cosine similarity calculation ensures consistency with historical decisions; the candidate solutions generated by the reinforcement learning algorithm have strong adaptability; the Lagrangian relaxation method verifies the feasibility of the solution, the multi-objective optimization calculates the implementation cost, and the hierarchical analysis method sorts the solutions, forming a complete solution evaluation system; the genetic algorithm optimizes the high-ranking solution, the dynamic programming algorithm generates the implementation sequence, and the multi-agent reinforcement learning optimizes the synergistic effect of the solution. These optimization methods jointly ensure the feasibility and optimality of the decision-making solution. This embodiment can effectively integrate multi-source information and generate scientific and reasonable decision-making solutions.

[0117] According to one aspect of the present application, step S41 is further:

[0118] S411, input the early warning indicator matrix W into the feature importance module, calculate the indicator contribution by the random forest algorithm, and obtain the importance vector X1; input the importance vector X1 into the correlation analysis module, eliminate redundant features by the maximum information coefficient, and obtain the screening indicator X2; input the screening indicator X2 into the principal component extraction module, extract nonlinear features by the kernel principal component analysis method, and generate the key indicator set K1;

[0119] S412, inputting the optimized parameter vector θ into the parameter encoding module, compressing the parameter space through the automatic encoder, and obtaining the encoding feature Y1; inputting the encoding feature Y1 into the neural mapping module, establishing nonlinear mapping through the multi-layer perceptron, and obtaining the hidden layer feature Y2; inputting the hidden layer feature Y2 into the parameter decoupling module, separating independent features through variational inference, and generating the mapping feature K2;

[0120] S413, input the key indicator set K1 and the mapping feature K2 into the feature fusion module, calculate the feature weight through the attention mechanism, and obtain the fusion feature Z1; input the fusion feature Z1 into the threshold learning module, determine the decision boundary through fuzzy clustering, and obtain the threshold vector Z2; input the threshold vector Z2 into the rule optimization module, optimize the rule structure through the genetic algorithm, and generate the initial rule Z3;

[0121] S414, input the initial rule Z3 into the rule refinement module, construct a multi-layer rule through decision tree ensemble learning, and obtain a hierarchical rule U1; input the hierarchical rule U1 into the rule pruning module, optimize the rule complexity through the cost complexity criterion, and obtain a simplified rule U2; input the simplified rule U2 into the rule combination module, combine the sub-rules through logistic regression, and generate a combined rule U3;

[0122] S415. Input the combined rule U3 into the rule verification module, evaluate the rule performance through cross-validation, and obtain the verification result V1; input the verification result V1 into the rule tuning module, adjust the rule parameters through Bayesian optimization, and obtain the optimized rule V2; input the optimized rule V2 into the rule encapsulation module, build the rule knowledge base through semantic coding, and generate the decision rule set R.

[0123] In one embodiment of the present application, the random forest algorithm is specifically: P(y|x) = ∑(l=1→L)[αl·∑(t=1→T)[wlt·pt(y|x)]]; wlt = softmax(vlt); vlt = f(hlt); hlt = g(Wl·x + bl); wherein P(y|x) is the predicted probability; pt(y|x) is the base classifier probability; αl is the layer weight; wlt is the tree weight; vlt is the attention score; hlt is the hidden layer feature; Wl is the weight matrix; bl is the bias vector; f(·) is the activation function; g(·) is the feature transformation function; x is the input feature; y is the category label; L is the number of layers; T is the number of trees per layer; softmax is the normalization function; deep features and ensemble learning are combined to improve classification performance.

[0124] The principal component analysis method is as follows: construct a feature dimension reduction function PCA(X, Ψ) = (U T X)·exp(-μ·‖ΨX-X‖ 2 / ‖X‖ 2), where X is the input matrix, U is the eigenvector matrix, UΛU T = X'X; Λ is the eigenvalue diagonal matrix; Ψ is the physical reconstruction operator; μ is the reconstruction error penalty factor; X' is the transpose of X; ‖·‖ represents the Frobenius norm; exp is the exponential function; the eigenvalue λi satisfies λ1≥λ2≥...≥λn.

[0125] This embodiment establishes an efficient decision rule generation mechanism. The dimension reduction of early warning indicators by principal component analysis method not only reduces the dimension of decision space, but also retains the most representative indicator combination, thus improving decision efficiency; the mapping relationship between parameters and decision rules established by deep neural network has nonlinear fitting ability, which can accurately capture complex parameter-rule correspondence; the hierarchical rule system constructed by decision tree algorithm has a clear logical structure, which improves the interpretability of rules; the introduction of fuzzy reasoning mechanism makes the rules have soft boundaries, which enhances the robustness of decision-making; the historical decision similarity is calculated by cosine similarity, which ensures the consistency of newly generated rules with historical experience; the reinforcement learning algorithm can continuously optimize decision rules according to environmental feedback, thus improving the adaptability of rules. This embodiment not only ensures the scientific nature of decision rules, but also improves the practicality of rules.

[0126] According to one aspect of the present application, step S43 is further:

[0127] S431, input the candidate solution matrix A into the constraint decomposition module, construct the constraint function by the interior point method, and obtain the constraint vector X1; input the constraint vector X1 into the dual transformation module, construct the dual problem by the Lagrange multiplier method, and obtain the dual variable X2; input the dual variable X2 into the relaxation optimization module, solve the relaxation problem by the augmented Lagrangian method, and generate the feasible solution set B1;

[0128] S432, input the feasible solution set B1 into the resource evaluation module, quantify the resource consumption by fuzzy comprehensive evaluation method, and obtain the resource vector Y1; input the resource vector Y1 into the time estimation module, calculate the implementation period by critical path method, and obtain the time vector Y2; input the resource vector Y1 and the time vector Y2 into the cost integration module, balance the multi-objective constraints by Pareto optimization, and generate the cost matrix B2;

[0129] S433, input the feasible solution set B1 into the criterion construction module, determine the evaluation criteria by the Delphi method, and obtain the criterion vector Z1; input the criterion vector Z1 into the weight calculation module, calculate the criterion weight by the entropy weight method, and obtain the weight vector Z2; input the feasible solution set B1 and the weight vector Z2 into the judgment matrix module, construct the judgment matrix by fuzzy hierarchical analysis, and generate the evaluation matrix Z3;

[0130] S434, input the evaluation matrix Z3 and the cost matrix B2 into the utility calculation module, calculate the utility value of the solution by the TOPSIS method, and obtain the utility vector U1; input the utility vector U1 into the consistency verification module, verify the rationality of the judgment by the consistency ratio, and obtain the correction vector U2; input the correction vector U2 into the sorting optimization module, optimize the sorting result by the order relationship analysis method, and generate the initial sorting U3;

[0131] S435. Input the initial ranking U3 into the sensitivity analysis module, evaluate the ranking stability through Monte Carlo simulation, and obtain the stability index V1; input the stability index V1 into the ranking calibration module, optimize the ranking result through grey correlation analysis, and obtain the calibration ranking V2; input the calibration ranking V2 into the result encapsulation module, integrate the ranking information through the decision support system, and generate the ranking result B3.

[0132] In one embodiment of the present application, the interior point method is specifically: min f(x); st g(x)≤0, h(x)=0; L(x, λ, μ) = f(x) + λ T g(x) + μ T h(x) + ρ·B(x); B(x) = -∑log(-gi(x)); where f(x) is the objective function; g(x) is the inequality constraint; h(x) is the equality constraint; L(·) is the augmented Lagrangian function; λ, μ are Lagrangian multipliers; B(x) is the barrier function; ρ is the barrier parameter; x is the optimization variable; log(·) is the natural logarithm; i is the constraint index; the convergence is improved by the adaptive barrier function.

[0133] The specific implementation process of the TOPSIS method is: construct a utility scoring function U(d, w, ρ) = (d - ·w) / (d + ·w +d - ·w)·exp(-ρ·v); where d + is the positive ideal solution distance vector, d + = (∑(xi-x + i) 2 ) 1 / 2 ;d - is the negative ideal solution distance vector, d - = (∑(xi-x - i) 2 ) 1 / 2 ; w is the weight vector; ρ is the penalty factor; v is the constraint violation degree, v = max(0, g(x)); xi is the i-th evaluation index value; x + i is the i-th positive ideal solution; x - i is the i-th negative ideal solution; g(x) is the constraint function; exp is the exponential function.

[0134] The specific method of grey relational analysis is: γ(t)=[1+|S0(t)-Si(t)|+ζ·|S0'(t)-Si'(t)|] / [1+|S0(t)-Si(t)|+ζ·|S0'(t)-Si'(t)|+ρ·|S0(t)-Si(t)|]; S'(t)= dS(t) / dt; ζ=exp(-λ·V(t)); where γ(t) is the dynamic correlation degree; S0(t) is the reference sequence; Si(t) is the comparison sequence; S'(t) is the sequence derivative; ζ is the dynamic weight; ρ is the resolution coefficient; V(t) is the sequence volatility; λ is the sensitivity parameter; t is the time variable; exp(·) is the exponential function; the accuracy of relational analysis is improved by introducing dynamic weights.

[0135] This embodiment realizes complex multi-objective optimization decision-making. The Lagrangian relaxation method verifies the feasibility of the scheme and ensures that the optimization results meet various constraints; the multi-objective optimization algorithm comprehensively considers multiple objectives such as implementation cost, construction period, risk, etc., and can find multiple non-dominated solutions; the scheme evaluation system established by the hierarchical analysis method is systematic and comprehensive, which improves the rationality of the scheme ranking; the genetic algorithm optimizes the high-ranking scheme and explores a better solution space through operations such as crossover and mutation; the implementation sequence generated by the dynamic programming algorithm takes into account resource constraints and ensures implementation efficiency; the synergistic effect of the multi-agent reinforcement learning optimization scheme improves the overall effect of the decision-making scheme. This embodiment not only ensures the feasibility of the decision-making scheme, but also realizes the balanced optimization of multiple objectives.

[0136] The present invention directly embeds physical parameters such as mass matrix and stiffness matrix into the neural network structure through Hamilton quantity encoding and physical parameter mapping, ensuring that the feature extraction and learning process strictly follow physical laws such as conservation of energy and conservation of momentum; this deep fusion mechanism solves the problem of features violating physical laws in traditional methods and improves the reliability and interpretability of evaluation results; at the same time, the introduction of physical constraints also provides prior knowledge for the neural network, reduces data requirements, and improves the generalization ability of the model. The multi-level data fusion architecture is constructed, taking into account the spatial distribution and physical topological relationship of sensors. The weights of different data sources are adaptively adjusted through the attention mechanism, which not only maintains the correlation between local monitoring data, but also improves the expression ability of fusion features. The combination of tensor decomposition and autoencoder realizes efficient feature compression and reconstruction, significantly reducing the computational complexity while retaining key information. The time series feature learning mechanism based on recursive neural network and dynamic time warping algorithm can accurately capture the dynamic evolution law of structural state. The probabilistic representation provided by variational autoencoder and the long-term dependency modeling of multi-head attention mechanism make the state prediction have high accuracy and reliable uncertainty estimation. The dynamic threshold boundary generated by fuzzy mapping has strong adaptive ability, which improves the accuracy and timeliness of early warning. The multi-objective optimization framework comprehensively considers the coupling relationship of physical parameters and actual engineering constraints. Through the combined optimization of genetic algorithm and dynamic programming algorithm, it can generate decision-making schemes that meet both physical feasibility and engineering practicality. The introduction of multi-agent reinforcement learning improves the synergistic effect of the scheme, and the hierarchical analysis method ensures the scientificity and comprehensiveness of the scheme evaluation.

[0137] According to one aspect of the present application, a building health monitoring and assessment system based on a physical neural network includes:

[0138] at least one processor; and,

[0139] a memory communicatively connected to at least one of the processors; wherein,

[0140] 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 building health monitoring and assessment method based on physical neural network described in any of the above embodiments.

[0141] The present invention constructs a complete physical neural network building health monitoring and evaluation system. By introducing physical constraints, multi-source data fusion, dynamic feature learning and multi-level decision support, it effectively solves the problems existing in traditional methods, such as feature extraction not conforming to physical laws, poor data fusion effect, large deviation of diagnostic results, poor model interpretability, etc. The system has the following advantages: the introduction of physical constraints ensures that the feature extraction and learning process conform to basic physical laws, and improves the reliability of the evaluation results; the physical topological relationship of the spatial distribution of sensors is considered, and the correlation between local monitoring data is retained; the structural dynamics equation constraints are embedded in the diagnostic model, so that the diagnostic results conform to the actual physical mechanism; the embedding of physical parameters improves the interpretability of the neural network model; the physical knowledge in the historical decision-making experience is fully utilized to improve the rationality of the decision-making plan; the coupling relationship between physical parameters is considered to ensure that the optimization results meet the actual needs of the project. The present invention realizes the intelligence of the entire process from data acquisition, feature extraction to state evaluation and decision support, and has strong practical value.

[0142] It should be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A building health monitoring and assessment method based on physical neural network, characterized in that: The steps include: S1, obtaining the original sensor data set, and performing data preprocessing to generate a standardized data set; performing feature extraction on the standardized data set to generate a feature vector set; Fusing the feature vector set and the pre-stored historical feature data to generate a fused feature tensor; S2, performing physical constraint processing on the fused feature tensor and the pre-stored physical model parameter set to generate a physical constraint feature matrix; performing time series decomposition on the physical constraint feature matrix, and using a dynamic time warping algorithm to perform feature alignment to obtain a time-varying feature matrix; performing feature mapping processing on the time-varying feature matrix to generate a structural state feature vector; S3, performing state prediction based on the structural state feature vector and the pre-stored historical state database to obtain a predicted state sequence; performing early warning analysis based on the predicted state sequence and the pre-stored threshold parameter set to obtain an early warning indicator matrix; Optimize and analyze the early warning indicator matrix to generate an optimized parameter vector; S4, read and extract key indicators and decision rules according to the early warning indicator matrix and the optimization parameter vector, and generate a decision rule set; Generate solutions based on the decision rule set and the pre-stored historical decision library to obtain a candidate solution matrix; perform evaluation and optimization based on the candidate solution matrix to generate a decision solution sequence; Step S1 is further as follows: S11, collecting original sensor data sets through acceleration sensors, strain sensors, displacement sensors and temperature sensors arranged in the building, and using wavelet denoising method to eliminate high-frequency noise in the original sensor data sets to obtain a denoised data set; Interpolate missing values ​​in the denoised data set to obtain a complete data set; Normalize the complete data set to obtain a standardized data set; S12, using a sliding time window method, dividing the standardized data set into predetermined time series segments to obtain a time series segment set; calculating time domain features and frequency domain features for each segment in the time series segment set to obtain a first feature subset; An autoencoder is used to perform dimensionality reduction processing on the time series segment set, and potential feature representations are extracted to obtain a second feature subset; the first feature subset and the second feature subset are combined to form a feature vector set; The time domain features include mean, variance and peak value; the frequency domain features include power spectrum density and main frequency component; S13, using an attention mechanism to calculate the correlation weights of each feature in the feature vector set and the historical features in the pre-stored historical feature data to obtain a first weight matrix; based on the first weight matrix, performing weighted fusion on the feature vector set and the historical feature data to generate an initial fused feature; The tensor decomposition method is used to reduce the rank of the initial fusion features, extract the feature patterns, and finally obtain the fusion feature tensor; Step S2 is further as follows: S21. Use Hamilton quantity encoding method to map the features in the fused feature tensor to the physical space to obtain physical space features; input the pre-stored physical model parameter set into the parameter mapping module to obtain physical parameter mapping; based on the physical space features and physical parameter mapping, construct constraint conditions through Lagrange multiplier method to generate a physical constraint feature matrix; The physical model parameter set includes mass matrix, stiffness matrix and damping matrix; S22, using a recursive neural network to perform time series decomposition on the physical constraint feature matrix to obtain a time series feature sequence; Apply the dynamic time warping algorithm to the time series feature sequence to align the features of different time scales and obtain the aligned feature sequence; Align the feature sequence and input it into the attention layer, calculate the importance weight of the time series feature, and generate a weighted feature sequence; perform time series aggregation on the weighted feature sequence to obtain a time-varying feature matrix; S23, extracting high-order features in the time-varying feature matrix through a deep residual network; performing dimensionality reduction and normalization processing on the high-order features to obtain normalized features; mapping the normalized features to the structural state space to generate a structural state feature vector; Step S3 is further as follows: S31, using an empirical mode decomposition method to decompose the structural state characteristic vector into a predetermined number of intrinsic mode functions to obtain a mode sequence; based on the mode sequence, extracting the time-varying trend of each mode through wavelet transform to obtain trend characteristics; Based on the pre-stored historical state database, the dynamic time warping algorithm is used to match the trend characteristics and generate a matching feature sequence; S32, based on the matching feature sequence, using a variational autoencoder to extract potential feature representation to obtain encoding features; Based on the encoding features, the feature dependencies at different time scales are calculated to obtain the attention features; Based on the attention features, a multi-step prediction result is generated through a long short-term memory network to obtain a prediction state sequence; S33, based on the predicted state sequence, calculating the statistical characteristics of each state indicator to obtain a statistical characteristic matrix; based on the pre-stored threshold parameter set, generating a dynamic threshold boundary through fuzzy mapping to obtain a threshold boundary set; based on the statistical characteristic matrix and the threshold boundary set, calculating the warning probability through a probability density estimation model to obtain a warning indicator matrix; S34. Based on the early warning indicator matrix, the early warning model parameters are adjusted using the Bayesian optimization method to obtain an optimized parameter set; Based on the optimized parameter set, the parameter performance is evaluated through cross-validation to obtain the validation results; Based on the verification results, an adaptive learning algorithm is used to update the model parameters and generate an optimized parameter vector; Step S4 is further as follows: S41. Based on the early warning indicator matrix, the principal component analysis method is used to extract key indicators to obtain a key indicator set; Based on the optimized parameter vector, a mapping relationship between parameters and decision rules is constructed through a deep neural network to obtain mapping features; Based on the key indicator set and mapping features, the decision tree algorithm is used to construct hierarchical decision rules and generate a decision rule set; S42, based on the decision rule set, using fuzzy reasoning method to analyze the decision rules and obtain reasoning results; based on the pre-stored historical decision library, using cosine similarity to calculate the degree of matching with the reasoning results, to obtain a similarity vector; based on the reasoning results and the similarity vector, using reinforcement learning algorithm to obtain a candidate solution matrix; S43, based on the candidate solution matrix, verify the feasibility of the solution through Lagrangian relaxation method to obtain a feasible solution set; Based on the feasible solution set, the multi-objective optimization method is used to calculate the implementation cost of the solution and obtain the cost matrix; Based on the feasible solution set and cost matrix, the hierarchical analysis method is used to perform comprehensive sorting and generate sorting results; S44, based on the sorting results, using a genetic algorithm to perform optimization adjustment to obtain an optimization solution set; Based on the optimization plan set, the implementation timing of the plan is generated through the dynamic programming algorithm to obtain the implementation sequence; based on the implementation sequence, the multi-agent reinforcement learning algorithm is used to optimize the synergy of the plan and generate a decision plan sequence.

2. The building health monitoring and assessment method based on physical neural network according to claim 1 is characterized in that: Step S11 is further as follows: S111, acquiring acceleration sensor data, performing time synchronization through the GPS timing system, and obtaining synchronized acceleration data; acquiring strain sensor data, achieving a unified sampling rate through piecewise linear interpolation, and obtaining resampled strain data; acquiring displacement sensor data and temperature sensor data, and packaging the data through a time series compression algorithm to obtain a compressed data packet; Based on the synchronized acceleration data, resampled strain data and compressed data packets, a time window segmentation algorithm is used to generate the original sensor data set; S112, based on the original sensor data set, decomposing it through db4 wavelet basis function to obtain wavelet coefficients; Based on the wavelet coefficients, the soft threshold function is used for screening to obtain the noise reduction coefficient; Based on the noise reduction coefficient, the outliers are identified through the isolation forest algorithm to obtain the outlier marks; Based on the noise reduction coefficients and abnormal markers, the signal is reconstructed by inverse wavelet transform to generate a noise reduction data set; S113, based on the denoised data set, using a sliding window algorithm to locate data gaps and obtain missing locations; Based on the denoised data set, the correlation between sensors is calculated using the Pearson correlation coefficient to obtain the correlation matrix; Based on the denoised dataset, missing locations and correlation matrix, the low-rank structure of the data is extracted through Tucker decomposition to obtain the core tensor. Based on the core tensor, the cubic spline interpolation algorithm is used to fill the missing values ​​and generate a complete dataset. S114, based on the complete data set, calculate the data distribution parameters through adaptive moment estimation; based on the data distribution parameters, map the data to the [-1, 1] interval through maximum and minimum normalization to obtain normalized data; Based on the normalized data, the dynamic time warping algorithm is used to align the timing features of different sensors to obtain the aligned features; Based on the alignment features, the tensor encapsulation algorithm is used to integrate multi-source data and generate a standardized dataset.

3. The building health monitoring and assessment method based on physical neural network according to claim 1 is characterized in that: Step S12 is further as follows: S121, based on the standardized data set, performing signal segmentation by using a Hamming window function to obtain signal segments; based on the signal segments, using an adaptive window algorithm to adjust the window length and overlap rate to obtain optimized segments; Based on the optimized fragments, a three-dimensional data structure is constructed using a tensor rearrangement algorithm to generate a set of time-series fragments; S122, based on the time series segment set, calculating the spectral components by fast Fourier transform to obtain a spectral matrix; based on the spectral matrix, using an optimal basis selection algorithm to extract frequency band energy features; Based on the frequency band energy characteristics, the peak detection algorithm is used to extract the main frequency component and obtain the frequency domain characteristics; S123, calculating the signal analysis form based on the time series segment set to obtain an analysis signal; Based on the analytical signal, the variational mode decomposition algorithm is used to extract the intrinsic mode function and obtain the modal characteristics; Based on the modal features, the time-frequency features are extracted through Hilbert spectrum analysis to obtain the joint features; S124. Based on the time series segment set, extract local features through a one-dimensional convolutional neural network to obtain convolution features; based on the convolution features, use a self-attention algorithm to calculate feature weights to obtain weighted features; based on the weighted features, extract deep feature representations through latent variable encoding to obtain deep features; S125. Based on the frequency domain features, joint features and deep features, the Lasso regression algorithm is used to screen the key features to obtain the screening features; Based on the screening features, the feature importance is calculated through Fisher discriminant analysis to obtain important features; Based on important features, an adaptive weighting algorithm is used to integrate multi-source features to generate a feature vector set.

4. The building health monitoring and assessment method based on physical neural network according to claim 1 is characterized in that: Step S21 is further as follows: S211. Based on the fused feature tensor, the energy conservation characteristics are maintained through symplectic geometry to obtain energy characteristics; based on the energy characteristics, the motion equation is constructed through the variational principle to obtain dynamic characteristics; based on the dynamic characteristics, the physical system invariants are maintained through symplectic transformation to generate physical space characteristics; S212, based on the pre-stored physical model parameter set, maintaining the positive definiteness of the parameters through Lie algebra to obtain a first embedded feature; Based on the first embedded feature, the parameter symmetry is encoded by geometric algebra to obtain the constraint feature; based on the constraint feature, the parameter correlation is constructed by the energy function to generate the physical parameter mapping; S213, based on physical space characteristics and physical parameter mapping, numerical stability is maintained through a symplectic integral network to obtain consistent characteristics; Based on the consistency feature, the constraint conditions are constructed through variational inequalities; based on the constraint conditions, the augmented Lagrangian method is used to solve the optimal multiplier and generate the constraint multiplier; S214, based on the constraint multiplier, the activation function is modified to satisfy the energy dissipation law to obtain the neuron characteristics; Based on neuron features, the physical topological structure is maintained through graph neural network to obtain topological features; Topological features, through wave equation modeling feature propagation law, generate propagation features; S215. Based on the propagation characteristics, physical constraint loss is constructed through energy functional to obtain loss characteristics; based on the loss characteristics, constraint parameters are optimized through KKT conditions to obtain optimization characteristics; Based on the optimized features, the physical features are reconstructed through the Hamiltonian neural network to generate a physical constraint feature matrix.

5. The building health monitoring and assessment method based on physical neural network according to claim 1 is characterized in that: Step S32 is further as follows: S321, based on the matching feature sequence, estimate the feature distribution through the Gaussian mixture model to obtain distribution parameters; based on the distribution parameters, generate latent variables through random sampling to obtain sampling features; Based on the sampling features, the encoding representation is optimized by KL divergence to generate encoding features; S322, constructing a query matrix through linear projection based on the encoding features to obtain query features; Based on the query features, the attention weights are calculated by dot product operation to obtain the second weight matrix; Based on the second weight matrix, multi-scale dependencies are captured through a parallel attention mechanism to generate attention features; S323, based on the attention feature, injecting sequence order information through position encoding to obtain a second embedded feature; Based on the second embedded feature, the information flow is controlled through the GRU unit to obtain the updated feature; Based on the updated features, long-term dependencies are stored through an external memory network to generate enhanced features; S324, generating an initial prediction through a transformer decoder based on the enhanced features; Based on the initial prediction, the prediction accuracy is optimized through residual learning to obtain a revised prediction; Based on the modified prediction, a multi-step sequence is generated through cyclic prediction to obtain a predicted state sequence; S325. Based on the prediction state sequence, the prediction confidence interval is estimated by Monte Carlo sampling to obtain the confidence boundary; based on the confidence boundary, the prediction stability is evaluated by the Lyapunov index to obtain the stability index; based on the prediction state sequence and the stability index, the prediction results are optimized by Bayesian posterior inference to obtain the optimized prediction state sequence.

6. Building health monitoring and assessment system based on physical neural network, 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 that can be executed by the processor, and the instructions are used to be executed by the processor to implement the building health monitoring and assessment method based on physical neural network as described in any one of claims 1 to 5.

Images