A method and system for predicting the stability of an overpass bridge demolition structure

Abstract

The present application relates to the technical field of bridge demolition, and more particularly to a method and system for predicting the stability of an overpass bridge demolition structure, which obtains structural response data through construction monitoring and preprocessing, and extracts multi-source heterogeneous cross-modal structural response monitoring information; analyzes the cross-modal correlation between elevated support system degradation and cross-line construction interference, fuses time-space-structure characteristics, and constructs bridge demolition multi-modal fusion features; constructs a bridge demolition structure stability prediction model based on deep learning, inputs multi-modal fusion features to realize accurate prediction of the structure stability at the current demolition stage, and outputs stability prediction information; according to the prediction result, it is judged whether the structure state meets the safety threshold, if not, the demolition scheme is dynamically optimized, and intelligent auxiliary decision-making of the construction process is realized. The method improves the intelligent level of safety warning and decision-making of bridge demolition construction in complex environment.

Description

Technical Field

[0001] This invention relates to the field of bridge demolition technology, and in particular to a method and system for predicting the structural stability of overpass bridges during demolition. Background Technology

[0002] With the continuous upgrading and renovation of urban transportation infrastructure, a large number of existing overpasses are entering the demolition and reconstruction phase due to reasons such as the expiration of their service life, structural performance degradation, or adjustments to route planning. In actual engineering, overpasses often cross existing operating lines (such as highways, railways, and urban rail transit), and their demolition construction environment is complex and carries high safety risks. They are highly susceptible to the combined effects of vibration interference from the operation of the underlying lines and the performance degradation of the support system after long-term service. Traditional bridge demolition stability assessments rely heavily on static mechanical models and empirical judgments, making it difficult to reflect the actual stress state of the structure under the coupled action of dynamic loads and progressive damage in real time. Especially when facing latent defects such as stiffness degradation of the elevated support system, concrete cracking, and steel corrosion, there is a lack of effective fusion and intelligent analysis capabilities for multi-source heterogeneous monitoring data.

[0003] In recent years, although some studies have begun to introduce methods combining sensor monitoring and numerical simulation for construction safety early warning, existing technologies are generally limited to the analysis of single-modal data (such as displacement or strain), failing to fully explore the cross-modal correlation mechanism between the degradation characteristics of the support system and external construction disturbances, resulting in insufficient sensitivity and accuracy in identifying precursors of structural instability. Furthermore, achieving spatiotemporal alignment, noise suppression, and feature enhancement of multi-source monitoring information under complex working conditions remains a significant technical challenge. Simultaneously, existing stability prediction models mostly employ traditional machine learning methods, which are insufficient for effectively modeling the nonlinear and time-varying evolution of structural responses during demolition, and lack the intelligent decision support capability to dynamically optimize construction plans in conjunction with prediction results.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for predicting the structural stability of overpass demolition structures. This aims to solve the technical problems of existing overpass demolition stability prediction technologies, which are unable to effectively integrate multimodal monitoring data on support system degradation and cross-line construction interference, and lack the ability to accurately model and dynamically warn of the nonlinear time-varying response of structures.

[0006] To achieve the above objectives, the present invention provides a method for predicting the stability of a demolished overpass structure, the method comprising:

[0007] Structural response monitoring data was obtained by monitoring the demolition construction of the target overpass bridge. The obtained structural response monitoring data was preprocessed to obtain cross-modal structural response monitoring information.

[0008] Based on the cross-modal structural response monitoring information, the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data is analyzed, and multi-modal fusion features are constructed to obtain the multi-modal fusion features of bridge demolition.

[0009] A bridge demolition structural stability prediction model is constructed. The multimodal fusion features of the bridge demolition are input into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure in the current demolition stage, thereby obtaining bridge structural stability prediction information.

[0010] Based on the bridge structural stability prediction information, it is determined whether the structural stability of the current demolition stage meets the safety threshold. If it does not, the current demolition plan is optimized to assist in the demolition construction decision-making.

[0011] Optionally, the step of monitoring the demolition construction of the target overpass to obtain structural response monitoring data, and preprocessing the obtained structural response monitoring data to obtain cross-modal structural response monitoring information, specifically includes:

[0012] During the demolition of the target overpass bridge, a network of sensors was deployed to monitor the structural response of key parts of the bridge and obtain structural response monitoring data. The structural response monitoring data includes support system degradation monitoring data, cross-line construction interference monitoring data, and structural dynamic response data.

[0013] The structural response monitoring data is decomposed into wavelet packets using a preset Morlet wavelet basis function to obtain multi-scale sub-bands, and the Hilbert energy spectrum of each sub-band is calculated. After determining the characteristic frequency band based on the energy spectrum distribution, an improved Wiener filter is used for noise suppression.

[0014] Spatiotemporal alignment processing is performed on the filtered structural response monitoring data. Adaptive median filtering is used to eliminate impulse noise, and bilateral filtering is used to eliminate displacement ambiguity caused by structural vibration. A sliding window algorithm is introduced to calculate the root mean square value and kurtosis coefficient of the data within the window based on a preset window. Spline interpolation is then used to repair missing data.

[0015] Z-score standardization was performed on the repaired structural response monitoring data to convert the discrete monitoring data into a continuous probability distribution. K-means clustering was then used to perform load condition binning to generate cross-modal structural response monitoring information.

[0016] Optionally, the step of analyzing the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data based on the cross-modal structural response monitoring information, and constructing multi-modal fusion features to obtain the multi-modal fusion features of bridge demolition, specifically includes:

[0017] Obtain cross-modal structural response monitoring information, extract concrete crack image data from the support system degradation monitoring data, and import the concrete crack image data into the Lab color space to obtain chromaticity feature distribution;

[0018] The mean brightness of the crack region and the gradient variance of the a-channel are calculated by weighted fusion using the chromaticity feature distribution to form the damage index of the support system. Then, the boundary of the main connected region of the crack is extracted by the watershed algorithm, and the ratio of the boundary curvature to the crack area is calculated as the structural degradation sensitivity index. The first feature information is generated by combining the crack texture fractal dimension.

[0019] The construction vibration acceleration time history signal is extracted from the cross-line construction interference monitoring data, the time-frequency domain cross spectral density is calculated to generate the vibration interference feature spectrum, and the vibration interference feature spectrum is input into the pre-trained deep residual network to invert the construction load transfer path, generate the construction interference dynamic feature vector, and obtain the second feature information.

[0020] A cross-modal attention gating mechanism is introduced, using the first feature information as a state vector and the second feature information as an input vector and a hidden vector. The activation weights of the gating unit are calculated to perform feature interaction fusion and generate degradation-interference coupled feature information.

[0021] Displacement time history and stress spectrum are extracted from the structural dynamic response data using cross-modal structural response monitoring information. Combined with the degradation-interference coupling feature information, these are input into a spatiotemporal graph convolutional network for multimodal feature fusion to obtain the multimodal fusion features of bridge demolition.

[0022] Optionally, the step of extracting displacement time history and stress spectrum from the structural dynamic response data using cross-modal structural response monitoring information, and combining this with the degradation-interference coupling feature information, inputting it into a spatiotemporal graph convolutional network for multimodal feature fusion to obtain multimodal fusion features for bridge demolition, specifically includes:

[0023] In the time dimension, the data input to the spatiotemporal graph convolutional network is modeled as a sequence of spatiotemporal nodes, where each node contains the support damage state, construction vibration spectrum, beam displacement and stress components at the current moment;

[0024] The forward propagation layer captures the transient effects of support system degradation on the structural response through time-gated units, while the backward propagation layer uses memory decay gates to filter the cumulative effects of cross-line construction disturbances and generates time-varying dynamic characteristics.

[0025] In the spatial dimension, a spatial topology graph is constructed based on the bridge finite element model. Graph convolution is performed on the node features to obtain a spatial feature matrix. Global pooling with node degree weighting is then performed on the spatial feature matrix to generate a spatial importance weight vector.

[0026] Obtain the hidden states of nodes in the spatiotemporal graph convolutional network, perform tensor shrunk with the spatial importance weight vector, calculate the structural contribution score of each node in the dynamic response, and perform node weighting on the original spatial feature matrix to obtain spatially sensitive features.

[0027] The time-varying dynamic features and spatially sensitive features are input into a multilayer perceptron for feature stitching and nonlinear mapping to generate multimodal fusion features of bridge demolition during the demolition process of the target overpass bridge.

[0028] Optionally, the construction of the bridge demolition structural stability prediction model involves inputting the multimodal fusion features of the bridge demolition into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure at the current demolition stage, thereby obtaining bridge structural stability prediction information. Specifically, this includes:

[0029] Based on the historical bridge demolition database, historical demolition cases with different stability states are retrieved. Based on the historical demolition cases, the historical structural response characteristics during the demolition process of each case are extracted to obtain historical structural response characteristic information.

[0030] Extract the demolition process sequence corresponding to each historical demolition case, divide the historical structure response feature information into process partitions, and generate historical feature subsets corresponding to different demolition stages;

[0031] The historical structural response characteristics after process partitioning are associated with the corresponding structural stability level, and a process-structure response topology is constructed with demolition process as node label;

[0032] A bridge demolition structural stability prediction model is constructed based on gated recurrent units and graph attention networks. A subset of historical structural response features is input into the gated recurrent units to extract multi-scale process time-series features.

[0033] The adjacency matrix is ​​obtained through the topological graph. Node features are aggregated and node representations are updated based on the attention mechanism. Hierarchical pooling is performed on the node representations of the entire graph to obtain process space features. The multi-scale process temporal features and process space features are fused to generate process-structure spatiotemporal features.

[0034] The process-structure spatiotemporal features are input into a softmax classification layer for stability level prediction. The prediction results are verified by a confusion matrix and the hyperparameters are tuned by reserving a validation set. After iterative training, a bridge demolition structure stability prediction model that meets engineering accuracy is obtained.

[0035] The multimodal fusion features of bridge demolition are obtained and then input into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure in the current demolition stage, thereby obtaining bridge structural stability prediction information.

[0036] Optionally, the step of determining whether the structural stability of the current demolition stage meets the safety threshold based on the bridge structural stability prediction information, and if not, optimizing the current demolition plan to assist in demolition construction decision-making, specifically includes:

[0037] Obtain bridge structural stability prediction information, and generate a time-varying curve of structural safety at the current demolition stage based on the bridge structural stability prediction information. The time-varying curve of structural safety represents the safety reserve coefficient of key bridge sections at different demolition times.

[0038] Obtain the standard safety threshold for the current demolition stage, calculate the safety deviation value at each time node based on the time-varying curve of structural safety, and compare the calculated deviation value with the preset risk tolerance range.

[0039] If the deviation value exceeds the preset risk tolerance range, it is determined that the current demolition control parameters do not meet the structural safety requirements, and then historical cases that meet the target bridge type and degradation status are retrieved from the preset historical demolition scheme library.

[0040] Extract the demolition control parameter set corresponding to historical cases, introduce a constrained multi-objective particle swarm optimization algorithm, construct a dual objective function with the goal of maximizing safety and construction efficiency, and set displacement limit and stress threshold as constraints.

[0041] Initialize the particle swarm position as the historical demolition parameter vector, calculate the non-dominated solution front and evaluate the individual crowding distance, select the safest particle in the Pareto front as the guide particle, and update the velocity-position equation of the following particles.

[0042] By iteratively optimizing the output set of feasible demolition parameters, the feasibility of construction is verified using the set of feasible demolition parameters. Based on the verification results, the optimal demolition optimization scheme is selected to generate a dynamic demolition instruction set, and the demolition construction is assisted in decision-making through a digital twin platform.

[0043] Furthermore, to achieve the above objectives, the present invention also provides a system for predicting the structural stability of an overpass bridge during demolition, the system comprising:

[0044] The monitoring and preprocessing module is used to monitor the demolition construction of the target overpass bridge, obtain structural response monitoring data, and preprocess the obtained structural response monitoring data to obtain cross-modal structural response monitoring information.

[0045] The feature fusion module is used to analyze the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data based on the cross-modal structural response monitoring information, and to construct multi-modal fusion features to obtain the multi-modal fusion features of bridge demolition.

[0046] The stability prediction module is used to construct a stability prediction model for bridge demolition structures. The multimodal fusion features of bridge demolition are input into the trained bridge demolition structure stability prediction model to predict the stability of the bridge structure in the current demolition stage and obtain bridge structure stability prediction information.

[0047] The decision optimization module is used to determine whether the structural stability of the current demolition stage meets the safety threshold based on the bridge structural stability prediction information. If it does not meet the safety threshold, the module optimizes the current demolition plan to make auxiliary decisions for demolition construction.

[0048] In addition, to achieve the above objectives, the present invention also provides a device for predicting the stability of an overpass demolition structure. The device includes: a memory, a processor, and an overpass demolition structure stability prediction program stored in the memory and executable on the processor. The overpass demolition structure stability prediction program is configured to implement the steps of the overpass demolition structure stability prediction method as described in any one of the above descriptions.

[0049] In addition, to achieve the above objectives, the present invention also provides a medium storing a stability prediction program for the demolition structure of an overpass, wherein when the overpass demolition structure stability prediction program is executed by a processor, the program implements the steps of the stability prediction method for the demolition structure of an overpass as described above.

[0050] In addition, to achieve the above objectives, the present invention also provides a computer program product, including a stability prediction program for the demolition structure of an overpass, wherein when the overpass demolition structure stability prediction program is executed by a processor, it implements the steps of the stability prediction method for the demolition structure of an overpass as described above.

[0051] This invention provides a method for predicting the structural stability of overpass demolition structures. The method acquires structural response monitoring data during the demolition process using multi-source sensors. Preprocessing techniques such as wavelet packet-Hilbert joint filtering and spatiotemporal alignment are used to extract cross-modal structural response monitoring information, effectively improving data quality and spatiotemporal consistency in complex environments. Furthermore, a cross-modal attention mechanism and a spatiotemporal graph convolutional network are introduced to deeply integrate the coupling relationship between the degradation characteristics of the elevated support system and the interference from cross-line construction, constructing a multi-modal fusion feature with physical meaning and dynamic representation capabilities. A stability prediction model with process-structure response topology awareness is constructed by fusing gated recurrent units and graph attention networks, achieving accurate temporal prediction of the structural stability state during demolition. Combined with safety criteria and constrained multi-objective optimization algorithms, a closed loop of perception, prediction, and decision-making is formed, supporting dynamic adjustment of construction plans and intelligent auxiliary decision-making. The overall method overcomes the limitations of traditional methods relying on experience-based judgment or single data source analysis, significantly improving the safety, prediction accuracy, and intelligence level of overpass demolition construction under complex interference and structural degradation coupling conditions, providing reliable technical support for high-risk bridge demolition projects in densely populated urban areas. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating an embodiment of the method for predicting the stability of a bridge demolition structure according to the present invention.

[0053] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0054] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0055] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the method for predicting the stability of a demolished overpass structure according to the present invention. An embodiment of the method for predicting the stability of a demolished overpass structure according to the present invention is presented.

[0056] In one embodiment, the method for predicting the stability of the overpass demolition structure includes:

[0057] Step S100: Conduct demolition construction monitoring of the target overpass bridge to obtain structural response monitoring data, and preprocess the obtained structural response monitoring data to obtain cross-modal structural response monitoring information.

[0058] The target overpass can be an existing overpass requiring demolition, typically spanning highways, railways, or urban rail transit lines. It serves as the main structure throughout the demolition process, and its stability directly impacts the safety and efficiency of the demolition work. Structural response monitoring data can be real-time data on the bridge's structural state collected by multi-source sensors (such as displacement gauges and strain gauges) installed on the bridge. This provides firsthand information on the dynamic changes of the bridge during demolition, offering fundamental information for subsequent analysis and prediction. For example, structural response monitoring data can be achieved by continuously monitoring key parameters at different parts of the bridge using various types of sensors and transmitting this data to a central processing system. Cross-modal structural response monitoring information can be pre-processed structural response monitoring data, possessing higher quality and spatiotemporal consistency. This improves data quality, making subsequent analysis more accurate and reliable.

[0059] Furthermore, cross-modal structural response monitoring information can be obtained by applying signal processing and spatiotemporal alignment techniques to the raw monitoring data to remove noise and enhance useful signals. For monitoring the demolition of a target overpass bridge, structural response monitoring data can be obtained by deploying various sensors at key locations on the bridge to periodically or continuously collect changes in relevant physical quantities. Furthermore, this can be achieved by manually reading instrument readings and automatically uploading them to a cloud server, ensuring sufficient raw data to support subsequent steps. Preprocessing the acquired structural response monitoring data to obtain cross-modal structural response monitoring information can involve using signal processing algorithms to remove noise and correct biases. Furthermore, this preprocessing can be achieved through offline batch processing and online real-time processing, thereby improving data availability.

[0060] Step S200: Based on the cross-modal structural response monitoring information, analyze the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data, and construct multi-modal fusion features to obtain the multi-modal fusion features of bridge demolition.

[0061] Among these, the elevated support system degradation data can reflect information on the aging and damage of the bridge support structure due to long-term use. This data can help identify potential safety hazards and provide a basis for developing reasonable demolition plans. For example, elevated support system degradation data can be obtained through a comprehensive evaluation combining historical maintenance records and on-site inspection results. Cross-line construction disturbance data can be records of external vibrations or other forms of disturbance caused to the bridge being demolished by the operation of the underpass. This data can help understand how external factors affect the changing trends of bridge structural stability. For example, cross-line construction disturbance data can be obtained by capturing such disturbances using equipment such as vibration sensors placed near the bridge. The multimodal fusion feature of bridge demolition can be a comprehensive feature vector constructed based on cross-modal structural response monitoring information, capable of simultaneously reflecting the relationships between different types of monitoring data. This provides strong support for a more comprehensive understanding of the current state of the bridge and its evolution over time.

[0062] Furthermore, the multimodal fusion features of bridge demolition can be achieved by employing cross-modal attention mechanisms and spatiotemporal graph convolutional networks to mine and integrate relevant information from multiple perspectives. Analyzing the cross-modal correlation between elevated support system degradation data and cross-line construction interference data based on cross-modal structural response monitoring information, and constructing multimodal fusion features, can utilize machine learning techniques to explore the intrinsic connections between data from different sources. Furthermore, analyzing the cross-modal correlation between elevated support system degradation data and cross-line construction interference data based on cross-modal structural response monitoring information, and constructing multimodal fusion features, can be achieved through rule-based methods and statistical methods, thereby revealing hidden patterns.

[0063] Step S300: Construct a bridge demolition structural stability prediction model. Input the multimodal fusion features of bridge demolition into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure in the current demolition stage and obtain bridge structural stability prediction information.

[0064] The stability prediction information for bridge demolition structures can be the output of a specially designed prediction model based on input multimodal fusion features, indicating the potential stability of the bridge structure over a future period. This information can guide operators to take timely measures to prevent accidents. For example, the stability prediction information can be achieved by modeling complex nonlinear dynamic processes using a combination of gated recurrent units and graph attention networks. Constructing the bridge demolition structure stability prediction model involves selecting an appropriate mathematical framework to describe the problem and adjusting parameters using a training set. Furthermore, the model can be built using supervised learning or reinforcement learning, thereby generating prediction tools. Inputting the multimodal fusion features of bridge demolition into the trained model to predict the stability of the bridge structure at the current demolition stage can yield stability prediction information by performing forward propagation calculations. Further, inputting the multimodal fusion features into the trained model to predict the stability of the bridge structure at the current demolition stage can also yield stability prediction information through single-shot or rolling prediction, providing forward-looking insights.

[0065] Step S400: Based on the bridge structure stability prediction information, determine whether the structural stability of the current demolition stage meets the safety threshold. If it does not meet the threshold, optimize the current demolition plan to make auxiliary decisions for demolition construction.

[0066] The method for determining whether the structural stability of the current demolition stage meets the safety threshold based on bridge structural stability prediction information can be achieved by comparing the prediction results with preset standards. Furthermore, this determination can be implemented through manual review or automated decision-making, allowing for rapid risk identification. If the stability does not meet the threshold, the current demolition plan is optimized for auxiliary decision-making, such as adjusting the construction sequence or methods. Further, if the stability does not meet the threshold, optimization of the current demolition plan for auxiliary decision-making can be achieved through expert system suggestions or optimization algorithms, thereby reducing the probability of accidents. For example, in a scenario with heavy urban traffic, the structural stability prediction method for overpass demolition in this embodiment can be as follows: During the demolition process, multi-source sensors are deployed at key locations on the bridge to monitor the bridge structural response data in real time. After preprocessing, machine learning techniques are used to analyze the correlation between the degradation data of the elevated support system and the interference data of the cross-line construction, constructing multimodal fusion features. These features are then input into a pre-trained stability prediction model to predict the structural stability of the bridge at the current demolition stage. If the prediction results show that the structural stability does not meet the safety threshold, the demolition plan is optimized to ensure construction safety.

[0067] In one embodiment, structural response monitoring data is obtained by monitoring the demolition of the target overpass bridge. The obtained structural response monitoring data is then preprocessed to obtain cross-modal structural response monitoring information, specifically including:

[0068] During the demolition of the target overpass bridge, a network of sensors was deployed to monitor the structural response of key parts of the bridge and obtain structural response monitoring data. The structural response monitoring data included support system degradation monitoring data, cross-line construction interference monitoring data, and structural dynamic response data.

[0069] The sensor network, composed of multiple sensor nodes, is used to monitor the structural response of critical bridge components in real time. This can be achieved by installing various types of sensors (such as displacement gauges and strain gauges) at key locations on the bridge and transmitting data to a central processing system via wired or wireless means. Furthermore, the sensor network can provide real-time, comprehensive structural response data, supporting subsequent data analysis and prediction. Support system degradation monitoring data reflects the aging and damage to the bridge's support structure due to long-term use. This data can be obtained by continuously monitoring changes in relevant parameters using sensors installed on the bridge's support system. Furthermore, support system degradation monitoring data can help identify potential safety hazards, providing a basis for developing reasonable demolition plans.

[0070] Cross-line construction disturbance monitoring data can be records of external vibrations or other forms of disturbance caused to the bridge being demolished by the operation of the underlying track. It is understood that such disturbances can be captured by devices such as vibration sensors installed near the bridge. Furthermore, cross-line construction disturbance monitoring data helps to understand how external factors affect the changing trends of bridge structural stability. Structural dynamic response data can reflect the bridge's response to various dynamic loads during demolition. It is understood that structural dynamic response data can be collected in real time by sensors such as accelerometers and displacement gauges installed at key locations on the bridge. Furthermore, structural dynamic response data provides information on the dynamic behavior of the bridge during demolition, supporting stability assessment. Structural response monitoring data can be obtained by deploying various types of sensors at key locations on the bridge to periodically or continuously collect changes in relevant physical quantities. Furthermore, structural response monitoring data can be obtained by manually reading instrument readings and automatically uploading them to a cloud server, ensuring sufficient raw data to support subsequent steps.

[0071] The structural response monitoring data is decomposed into wavelet packets using a preset Morlet wavelet basis function to obtain multi-scale sub-bands. The Hilbert energy spectrum of each sub-band is calculated. After determining the characteristic frequency band based on the energy spectrum distribution, an improved Wiener filter is used for noise suppression.

[0072] The Morlet wavelet basis function is a commonly used wavelet basis function with good time-frequency localization properties. It's understood that the Morlet wavelet basis function is pre-defined in signal processing algorithms as the basis for wavelet packet decomposition. Furthermore, the Morlet wavelet basis function is used to perform multi-scale decomposition on structural response monitoring data to extract different frequency components. Multi-scale sub-bands can be sub-bands within different frequency ranges obtained by wavelet packet decomposition of the original signal. It's understood that multi-scale sub-bands utilize the Morlet wavelet basis function to perform wavelet packet decomposition on the structural response monitoring data. Furthermore, multi-scale sub-bands provide information on different frequency components, facilitating subsequent energy spectrum calculations and characteristic frequency band determination.

[0073] The Hilbert energy spectrum can be an energy distribution obtained by performing a Hilbert transform on multi-scale sub-bands. It can be understood that the Hilbert energy spectrum is calculated by performing a Hilbert transform on each sub-band. Furthermore, the Hilbert energy spectrum is used to determine characteristic frequency bands, providing a basis for noise suppression. The characteristic frequency band can be a frequency range with significant energy contributions determined based on the Hilbert energy spectrum distribution. It can be understood that the characteristic frequency band is selected based on the distribution of the Hilbert energy spectrum, choosing frequency bands with higher energy. Furthermore, the characteristic frequency band is used to guide the noise suppression process of improved Wiener filtering, improving data quality. Improved Wiener filtering can be an enhanced Wiener filtering algorithm used to remove noise from a signal. It can be understood that improved Wiener filtering is based on the characteristic frequency band, designing and applying an improved Wiener filter. Furthermore, improved Wiener filtering effectively suppresses noise and improves the signal-to-noise ratio.

[0074] Multi-scale subbands are obtained by wavelet packet decomposition of structural response monitoring data using a preset Morlet wavelet basis function. This process can be implemented through offline batch processing or online real-time processing, allowing for the extraction of different frequency components for subsequent energy spectrum calculations. The Hilbert energy spectrum of each subband is calculated by performing a Hilbert transform on each subband. This calculation can also be achieved through offline batch processing or online real-time processing, providing energy distributions for different frequency components and supporting the determination of characteristic frequency bands. After determining the characteristic frequency bands based on the energy spectrum distribution, noise suppression is performed using an improved Wiener filter. This can be done by selecting higher-energy frequency bands based on the Hilbert energy spectrum distribution and applying an improved Wiener filter. Furthermore, after determining the characteristic frequency band based on the energy spectrum distribution, noise suppression using improved Wiener filtering can be achieved through offline batch processing or online real-time processing, thereby effectively suppressing noise and improving the signal-to-noise ratio of the data.

[0075] Spatiotemporal alignment processing is performed on the filtered structural response monitoring data. Adaptive median filtering is used to eliminate impulse noise, and bilateral filtering is used to eliminate displacement ambiguity caused by structural vibration. A sliding window algorithm is introduced to calculate the root mean square value and kurtosis coefficient of the data within the window based on a preset window. Spline interpolation is then used to repair missing data.

[0076] Adaptive median filtering is a median filtering method that automatically adjusts the filter window size based on data characteristics. It dynamically adjusts the filter window size according to the statistical characteristics of the data to eliminate impulse noise. Furthermore, adaptive median filtering effectively removes impulse noise and improves data consistency. Bilateral filtering is a nonlinear filtering method that can smooth images while preserving edges. It smooths data by combining spatial distance and grayscale differences. Furthermore, bilateral filtering eliminates displacement blur caused by structural vibrations, improving data clarity. The sliding window algorithm is a data processing method based on a fixed window size, used to calculate the statistical characteristics of data within the window. It sets a fixed-size window, slides it across the data sequence, and calculates the root mean square value and kurtosis coefficient within the window. Furthermore, the sliding window algorithm provides statistical characteristics of local data, supporting data repair and anomaly detection.

[0077] The root mean square (RMS) value is the square root of the average of squared data points, often used to measure data volatility. It's understood that the RMS value is calculated using a sliding window algorithm to determine the root mean square of the data within a window. Furthermore, the RMS value reflects data volatility and supports data repair and anomaly detection. The kurtosis coefficient is a statistic describing the sharpness of a data distribution, used to measure the degree to which data deviates from a normal distribution. It's understood that the kurtosis coefficient is calculated using a sliding window algorithm to determine the kurtosis of the data within a window. Furthermore, the kurtosis coefficient reflects the distribution characteristics of the data and supports data repair and anomaly detection. Spline interpolation is an interpolation method based on spline functions used to fill in missing data. It's understood that spline interpolation uses known data points and fits a spline function to generate missing data points. Furthermore, spline interpolation repairs missing data, improving data integrity and continuity.

[0078] Spatiotemporal alignment of filtered structural response monitoring data can be performed by synchronizing the time and calibrating the space of the filtered data. Furthermore, spatiotemporal alignment of filtered structural response monitoring data can be achieved through offline batch processing or online real-time processing, thereby improving the temporal and spatial consistency of the data.

[0079] Adaptive median filtering is used to eliminate impulse noise by dynamically adjusting the filter window size based on the statistical characteristics of the data. Furthermore, this elimination can be achieved through offline batch processing and online real-time processing, effectively removing impulse noise and improving data consistency. Bilateral filtering is used to eliminate displacement ambiguity caused by structural vibration by combining spatial distance and grayscale differences to smooth the data and eliminate displacement ambiguity. This can also be achieved through offline batch processing and online real-time processing, improving data clarity and reducing displacement ambiguity. A sliding window algorithm is introduced to calculate the root mean square (RMS) and kurtosis coefficient of the data within a preset window. This can be achieved by setting a fixed-size window that slides across the data sequence, calculating the RMS and kurtosis coefficients within the window. This algorithm can be implemented through offline batch processing and online real-time processing, providing statistical characteristics of local data and supporting data repair and anomaly detection. Missing data repair using spline interpolation can be achieved by using known data points and fitting a spline function to generate missing data points. Furthermore, missing data repair using spline interpolation can be implemented through offline batch processing or online real-time processing, thereby repairing missing data and improving data integrity and continuity.

[0080] Z-score standardization was performed on the repaired structural response monitoring data to convert the discrete monitoring data into a continuous probability distribution. K-means clustering was then used to perform load condition binning to generate cross-modal structural response monitoring information.

[0081] Z-score standardization is a data standardization method that transforms data into a standard normal distribution. Essentially, Z-score standardization converts data to a standard normal distribution by subtracting the mean and dividing by the standard deviation. Furthermore, Z-score standardization ensures that the data has a uniform scale, facilitating subsequent data processing and model training.

[0082] K-means clustering can be a supervised learning method used to divide data into multiple clusters. It can be understood that K-means clustering iteratively optimizes the distribution of data points to the nearest cluster centers. Furthermore, K-means clustering bins the data, facilitating subsequent data analysis and feature extraction. Similarly, load-condition binning divides data under different operating conditions into different bins for easier subsequent processing. It can be understood that load-condition binning uses K-means clustering to divide the data into multiple bins. Furthermore, load-condition binning distinguishes data under different operating conditions, supporting more refined data analysis.

[0083] Z-score standardization of the repaired structural response monitoring data can be performed by subtracting the mean and dividing by the standard deviation to convert the data into a standard normal distribution. Furthermore, Z-score standardization can be implemented through offline batch processing or online real-time processing, ensuring the data has a uniform scale, facilitating subsequent data processing and model training. Converting discrete monitoring data into a continuous probability distribution can also be done through Z-score standardization. This can be achieved through offline batch processing or online real-time processing, ensuring the data has a uniform scale, facilitating subsequent data processing and model training. Binning based on operating conditions using K-means clustering can be achieved through iterative optimization, assigning data points to the nearest cluster centers. This binning can also be achieved through offline batch processing or online real-time processing, facilitating subsequent data analysis and feature extraction.

[0084] Generating cross-modal structural response monitoring information can be achieved by combining the above preprocessing steps to generate high-quality cross-modal structural response monitoring information. Furthermore, the generation of cross-modal structural response monitoring information can be achieved through offline batch processing or online real-time processing, thereby providing a high-quality data foundation to support subsequent multimodal fusion feature construction and stability prediction models.

[0085] For example, in the scenario of demolishing an urban overpass, the structural stability prediction method for overpass demolition in this embodiment can be as follows: During the demolition construction of the urban overpass, a sensor network deployed at key parts of the bridge is used to monitor the structural response of the bridge in real time, acquiring various data including support system degradation, cross-line construction interference, and structural dynamic response. This data undergoes wavelet packet-Hilbert joint filtering to effectively extract multi-scale sub-bands and calculate their Hilbert energy spectra, thereby determining characteristic frequency bands and suppressing noise through improved Wiener filtering, thus improving data quality. Subsequently, through spatiotemporal alignment processing, adaptive median filtering and bilateral filtering are used to eliminate impulse noise and displacement ambiguity, and data repair is performed using a sliding window algorithm and spline interpolation, further enhancing data consistency and integrity. Finally, Z-score normalization and K-means clustering are used for load condition binning to generate cross-modal structural response monitoring information, providing a high-quality data foundation for subsequent multimodal fusion feature construction and stability prediction models.

[0086] In one embodiment, the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data is analyzed based on cross-modal structural response monitoring information, and multi-modal fusion features are constructed to obtain the multi-modal fusion features of bridge demolition, specifically including:

[0087] Obtain cross-modal structural response monitoring information, extract concrete crack image data from the support system degradation monitoring data, and import the concrete crack image data into the Lab color space to obtain chromaticity feature distribution;

[0088] The concrete crack image data can be images of concrete cracks in the elevated support system captured by image sensors, which can be used to analyze and assess the morphology, distribution, and severity of the concrete cracks. Furthermore, concrete crack image data can be obtained by using high-resolution cameras or human-machine-mounted cameras to photograph key parts of the bridge. The Lab color space can be a color model that represents color as luminance (L) and two color components (a and b), facilitating the extraction of chromaticity features from images, which is helpful for crack detection and analysis. The chromaticity feature distribution can be the color distribution of the crack area in the Lab color space, which can be used to calculate the mean luminance and the gradient variance of the a-channel in the crack area, thereby assessing the degree of crack damage. Obtaining cross-modal structural response monitoring information can be achieved by collecting structural response data of the bridge during the demolition process using multi-source sensors. Furthermore, obtaining cross-modal structural response monitoring information can be achieved by manually reading instrument display values ​​and automatically uploading them to a cloud server, thus ensuring sufficient raw data to support subsequent steps.

[0089] Extracting concrete crack image data from the support system degradation monitoring data can be achieved by extracting concrete crack images from the elevated support system degradation data. Furthermore, the extraction of concrete crack image data from the support system degradation monitoring data can be automated using image processing software; manual screening and annotation can also be implemented, thus obtaining detailed image data for analyzing concrete cracks. Importing concrete crack image data into the Lab color space to obtain chromaticity feature distribution can be achieved by converting the image data to the Lab color space and extracting chromaticity features. Furthermore, importing concrete crack image data into the Lab color space to obtain chromaticity feature distribution can be performed using image processing libraries (such as OpenCV); custom algorithms can also be implemented, facilitating the extraction of chromaticity features from the images, which is helpful for crack detection and analysis.

[0090] The mean brightness of the crack region and the gradient variance of the a-channel are calculated by weighted fusion using the chromaticity feature distribution to form the damage index of the support system. Then, the boundary of the main connected region of the crack is extracted by the watershed algorithm, and the ratio of the boundary curvature to the crack area is calculated as the structural degradation sensitivity index. The first feature information is generated by combining the crack texture fractal dimension.

[0091] The support system damage index, calculated based on chromaticity feature distribution, is a quantitative indicator reflecting the degree of damage to the support system. It provides a quantitative damage assessment standard to help identify potential safety hazards. Furthermore, the support system damage index can be obtained by weighted fusion of the mean brightness of the crack region and the gradient variance of the α-channel. The watershed algorithm, an image segmentation technique, separates different regions in an image and can be used to extract the boundaries of the main connected regions of cracks, facilitating subsequent crack feature analysis. The structural degradation sensitivity index, calculated by the ratio of crack boundary curvature to crack area, reflects the sensitivity to structural degradation and can be used to assess the impact of cracks on the overall stability of the structure. The crack texture fractal dimension is a mathematical parameter describing the complexity of crack texture. It can be combined with other feature information to generate primary feature information, more comprehensively reflecting crack characteristics. The primary feature information can be a feature vector generated by combining the support system damage index, structural degradation sensitivity index, and crack texture fractal dimension. It can serve as the state vector of a cross-modal attention gating mechanism, participating in feature interaction fusion.

[0092] The damage index of the support system can be constructed by weighted fusion of the mean brightness and a-channel gradient variance of the crack region using chromaticity feature distribution. This can be achieved by calculating the mean brightness and a-channel gradient variance of the crack region and then weighting and fusing them. Furthermore, this weighted fusion can be implemented using statistical methods or rule-based approaches, providing a quantitative damage assessment standard to help identify potential safety hazards. The watershed algorithm can be used to extract the boundaries of the main connected regions of the crack by segmenting the crack image. This can be further implemented using image processing libraries (such as OpenCV) or custom algorithms, accurately extracting the boundaries of the main connected regions of the crack for subsequent crack feature analysis. The ratio of boundary curvature to crack area can be used as a structural degradation sensitivity index. Furthermore, calculating the ratio of boundary curvature to crack area as a structural degradation sensitivity index can be achieved using geometric calculation methods or rule-based methods, thereby assessing the impact of cracks on the overall structural stability. Generating the first feature information by combining the crack texture fractal dimension can be achieved by integrating the support system damage index, structural degradation sensitivity index, and crack texture fractal dimension. Furthermore, generating the first feature information by combining the crack texture fractal dimension can be achieved using feature engineering methods or rule-based methods, thereby generating first feature information containing multiple features for subsequent feature fusion.

[0093] The construction vibration acceleration time history signal is extracted from the cross-line construction interference monitoring data, the time-frequency domain cross spectral density is calculated to generate the vibration interference feature spectrum, the vibration interference feature spectrum is input into the pre-trained deep residual network to invert the construction load transfer path, generate the construction interference dynamic feature vector, and obtain the second feature information.

[0094] The construction vibration acceleration time history signal can be data on the change of vibration acceleration over time during construction, collected by accelerometers, and can be used to analyze the impact of construction disturbances on the bridge structure. Furthermore, the construction vibration acceleration time history signal can be obtained by real-time monitoring and recording vibration data using accelerometers installed near the bridge. The vibration disturbance feature spectrum can be a feature spectrum obtained by calculating the cross-spectral density in the time-frequency domain of the construction vibration acceleration time history signal, and can be used to invert the construction load transfer path and generate a construction disturbance dynamic feature vector. The pre-trained deep residual network can be a deep neural network trained on a large amount of data, capable of effectively extracting and processing complex features, and can be used to invert the construction load transfer path and generate a construction disturbance dynamic feature vector. The construction disturbance dynamic feature vector can be a feature vector generated after processing the vibration disturbance feature spectrum through the pre-trained deep residual network, and can be used as secondary feature information to participate in the feature interaction fusion of the cross-modal attention gating mechanism. The secondary feature information can be a feature vector composed of construction disturbance dynamic feature vectors, and can be used as the input vector and hidden vector of the cross-modal attention gating mechanism to participate in feature interaction fusion.

[0095] Extracting construction vibration acceleration time history signals from cross-line construction interference monitoring data can be achieved by extracting these signals from the data. Furthermore, this extraction can be accomplished through real-time monitoring and recording of data using acceleration sensors, and offline processing of historical data, thus obtaining vibration data for analyzing the impact of construction interference on the bridge structure. Generating a vibration interference characteristic spectrum by calculating the time-frequency domain cross-spectral density can be achieved by performing time-frequency domain cross-spectral density calculations on the construction vibration acceleration time history signals. Further, this calculation can be implemented using signal processing libraries (such as SciPy) or custom algorithms, thereby generating a characteristic spectrum reflecting the characteristics of construction interference for subsequent analysis.

[0096] Inverting the construction load transfer path by inputting the vibration disturbance feature spectrum into a pre-trained deep residual network can be achieved by using a deep learning framework (such as TensorFlow or PyTorch) and implementing a custom algorithm to generate a construction disturbance dynamic feature vector for subsequent feature fusion. Generating the construction disturbance dynamic feature vector and obtaining the second feature information can also be achieved by processing the vibration disturbance feature spectrum through a pre-trained deep residual network. Furthermore, generating the construction disturbance dynamic feature vector and obtaining the second feature information can be achieved by using a deep learning framework (such as TensorFlow or PyTorch) and implementing a custom algorithm to generate second feature information containing construction disturbance characteristics for subsequent feature fusion.

[0097] A cross-modal attention gating mechanism is introduced, using the first feature information as the state vector and the second feature information as the input vector and hidden vector. The activation weights of the gating unit are calculated to perform feature interaction fusion and generate degradation-interference coupled feature information.

[0098] The cross-modal attention gating mechanism can be a mechanism for multimodal feature fusion. It achieves feature interaction fusion by calculating the activation weights of the gating units, fusing the first and second feature information to generate degenerate-interference coupled feature information. The state vector can be a feature vector representing the current state in the cross-modal attention gating mechanism; in this scheme, the first feature information serves as the state vector in feature interaction fusion. The input vector can be a feature vector representing the input information in the cross-modal attention gating mechanism; in this scheme, the second feature information serves as the input vector in feature interaction fusion. The hidden vector can be a feature vector representing the hidden layer information in the cross-modal attention gating mechanism; in this scheme, the second feature information also serves as a hidden vector in feature interaction fusion. The gating unit activation weights can be weights used to control the feature interaction fusion process in the cross-modal attention gating mechanism; by calculating the gating unit activation weights, weighted fusion of different feature information can be achieved.

[0099] The degradation-interference coupling feature information can be a feature vector generated by fusing the first and second feature information through a cross-modal attention gating mechanism. This vector can be combined with displacement time history and stress spectrum data from structural dynamic response data to perform multimodal feature fusion, generating multimodal fused features for bridge demolition. Introducing a cross-modal attention gating mechanism, using the first feature information as a state vector and the second feature information as both input and hidden vectors, can be achieved using deep learning frameworks (such as TensorFlow and PyTorch) or custom algorithms, thus enabling the interactive fusion of features from different modalities.

[0100] Calculating the activation weights of gating units to perform feature interaction fusion and generate degenerate-interference coupled feature information can be achieved by calculating the activation weights of gating units and fusing the first and second feature information to generate degenerate-interference coupled feature information. Furthermore, calculating the activation weights of gating units to perform feature interaction fusion and generate degenerate-interference coupled feature information can be implemented using deep learning frameworks (such as TensorFlow, PyTorch); a custom algorithm can be implemented to generate coupled feature information containing both degenerate and interference characteristics for subsequent multimodal feature fusion.

[0101] Displacement time history and stress spectrum are extracted from the structural dynamic response data using cross-modal structural response monitoring information. Combined with degradation-interference coupling feature information, these are input into a spatiotemporal graph convolutional network for multimodal feature fusion to obtain the multimodal fusion features of bridge demolition.

[0102] The structural dynamic response data can be the dynamic response data of the bridge during the demolition process collected by sensors, including displacement time history and stress spectrum, which can be used to analyze the dynamic response characteristics of the bridge during demolition. Furthermore, the structural dynamic response data can be obtained through real-time monitoring and recording of data by various sensors installed on the bridge (such as displacement gauges, strain gauges, etc.). The displacement time history can be the data on the displacement changes of key points of the bridge over time during demolition, which can be used to analyze the displacement changes of the bridge during demolition. The stress spectrum can be the data on the stress changes of key points of the bridge over frequency during demolition, which can be used to analyze the stress distribution of the bridge during demolition. Spatiotemporal graph convolutional networks can be a deep learning model for processing spatiotemporal data, effectively fusing multimodal features. They can be used to fuse degradation-interference coupling feature information with structural dynamic response data to generate multimodal fused features for bridge demolition.

[0103] Extracting displacement time history and stress spectrum from structural dynamic response data using cross-modal structural response monitoring information can be achieved by extracting these parameters from the monitoring data. Furthermore, this extraction can be accomplished through real-time monitoring and recording of data using sensors, and offline processing of historical data, thereby obtaining displacement and stress data for analyzing the dynamic response characteristics of bridges.

[0104] By combining degradation-interference coupling feature information with structural dynamic response data and inputting it into a spatiotemporal graph convolutional network for multimodal feature fusion, the multimodal fused features for bridge demolition can be obtained. This can be achieved using deep learning frameworks (such as TensorFlow and PyTorch) or custom algorithms, thereby generating multimodal fused features for bridge demolition that contain multiple features, improving the sensitivity and accuracy of identifying precursors to structural instability.

[0105] For example, in the scenario of overpass demolition, the structural stability prediction method for overpass demolition in this embodiment can improve data quality and spatiotemporal consistency by preprocessing data collected from multiple sources of sensors. When analyzing the degradation data of the support system, the concrete crack image data is converted to the Lab color space and the chromaticity feature distribution is extracted. The damage index of the support system is calculated, and the first feature information is generated by combining the watershed algorithm and the fractal dimension of the crack texture. Simultaneously, by analyzing the cross-line construction interference data, the construction vibration acceleration time history signal is extracted to generate a vibration interference feature spectrum. The construction load transfer path is then inverted using a deep residual network to generate the second feature information. Next, a cross-modal attention gating mechanism is used to fuse the first and second feature information to generate degradation-interference coupled feature information. Finally, by combining the displacement time history and stress spectrum in the structural dynamic response data, a spatiotemporal graph convolutional network is used to perform multimodal feature fusion to obtain the bridge demolition multimodal fusion features. These multimodal fusion features can more comprehensively reflect the true stress state of the structure in complex environments, thereby improving the sensitivity and accuracy of identifying precursors to structural instability. This method enables accurate time-series prediction of structural stability during demolition, supports dynamic adjustment of construction plans and intelligent decision support, and significantly improves the safety, prediction accuracy and intelligence level of overpass demolition construction.

[0106] In one embodiment, displacement time history and stress spectrum are extracted from the structural dynamic response data using cross-modal structural response monitoring information. These are then combined with degradation-interference coupling feature information and input into a spatiotemporal graph convolutional network for multimodal feature fusion to obtain the bridge demolition multimodal fusion features, specifically including:

[0107] In the time dimension, the data input to the spatiotemporal graph convolutional network is modeled as a sequence of spatiotemporal nodes, where each node contains the support damage state, construction vibration spectrum, beam displacement, and stress components at the current moment.

[0108] The cross-modal structural response monitoring information can be pre-processed structural response monitoring data, possessing higher quality and spatiotemporal consistency, which can improve data quality and make subsequent analysis more accurate and reliable. In this embodiment, the cross-modal structural response monitoring information can remove noise and enhance useful signals by applying signal processing and spatiotemporal alignment techniques to the original monitoring data. The spatiotemporal node sequence can be a sequence of nodes in the time dimension, modeling the data input to the spatiotemporal graph convolutional network as a sequence containing the support damage state, construction vibration spectrum, beam displacement, and stress components at the current moment. This can be used to provide input data in the time dimension for the spatiotemporal graph convolutional network, facilitating the capture of dynamic changes in structural response. In an exemplary embodiment, the spatiotemporal node sequence can be achieved by organizing the cross-modal structural response monitoring information and degradation-interference coupling feature information into a node sequence in chronological order. The support damage state can be state information describing the degree of damage to the bridge support system at the current moment, which can be used to reflect the health status of the support system and provide key parameters for subsequent analysis. For example, the support damage state can be extracted and quantified from the support system degradation monitoring data.

[0109] Construction vibration spectrum can describe the frequency variation of vibration acceleration experienced by the bridge during construction. It can be used to analyze vibration disturbances during construction and assess their impact on the bridge structure. Furthermore, the construction vibration spectrum can be acquired using vibration sensors and subjected to spectral analysis. Beam displacement and stress components can describe the displacement and stress components of the bridge beam at the current moment. They can be used to analyze displacement and stress changes during bridge demolition and assess structural stability. In this embodiment, beam displacement and stress components can be acquired and recorded in real time using sensors such as displacement sensors and strain gauges.

[0110] The forward propagation layer captures the transient effects of support system degradation on the structural response through time-gated units, while the backward propagation layer uses memory decay gates to filter the cumulative effects of cross-line construction disturbances, generating time-varying dynamic characteristics.

[0111] The time-gating unit can be a neural network unit used to capture transient effects in time-series data, such as the transient effects of support system degradation on the structural response, generating time-varying dynamic features. In this embodiment, the time-gating unit can be implemented using a time-gating unit in the forward propagation layer. The memory decay gate can be a neural network unit used to filter cumulative effects in time-series data, such as the cumulative effects of cross-line construction disturbances, generating time-varying dynamic features. In a specific embodiment, the memory decay gate can be implemented using a memory decay gate in the backward propagation layer. The time-varying dynamic features can be generated by combining the time-gating unit and the memory decay gate, reflecting the dynamic characteristics of the structural response changing over time. This can comprehensively reflect the dynamic changes of the structural response, providing key information for subsequent analysis. Furthermore, the time-varying dynamic features can be generated through calculations in the forward and backward propagation layers.

[0112] In the spatial dimension, a spatial topology graph is constructed based on the bridge finite element model. Graph convolution is performed on the node features to obtain the spatial feature matrix. Global pooling with node degree weighting is then performed on the spatial feature matrix to generate a spatial importance weight vector.

[0113] The bridge finite element model can be a bridge structural model established using the finite element method, used to simulate the bridge's behavior under various working conditions. It can provide a basis for constructing a spatial topology map and assist in analyzing the mechanical properties of the bridge structure. For example, the bridge finite element model can be established and verified using finite element software. The spatial topology map can be constructed based on the bridge finite element model, describing the graph structure between different parts of the bridge. It can provide a basis for graph convolution operations, facilitating the capture of the spatial features of the bridge structure. In this embodiment, the spatial topology map can be generated based on the bridge finite element model. The spatial feature matrix can be a feature matrix obtained by performing graph convolution operations on the node features in the spatial topology map. It can be used to comprehensively reflect the spatial features of the bridge structure, providing a basis for subsequent analysis. Further, the spatial feature matrix can be generated through graph convolution operations. The spatial importance weight vector can be a weight vector generated by global pooling of the spatial feature matrix with node degree weights. It can be used to reflect the importance of each node in the spatial features, providing key information for subsequent analysis. In an exemplary embodiment, the spatial importance weight vector can be generated by global pooling of the spatial feature matrix.

[0114] The hidden states of nodes in the spatiotemporal graph convolutional network are obtained and tensor-folded with the spatial importance weight vector. The structural contribution score of each node in the dynamic response is calculated, and the original spatial feature matrix is ​​weighted by nodes to obtain spatially sensitive features.

[0115] The hidden states of nodes can be the hidden states of each node in a spatiotemporal graph convolutional network, containing the node's feature information, which can be used to provide a foundation for subsequent tensor shrinking and feature concatenation. In this embodiment, the hidden states of nodes can be generated through forward propagation computation of the spatiotemporal graph convolutional network. Spatial sensitive features can be generated by node weighting of the original spatial feature matrix, reflecting the structural contribution of each node in the dynamic response, and can be used to comprehensively reflect the contribution of each node in the dynamic response, providing key information for subsequent analysis. Furthermore, spatial sensitive features can be generated by node weighting of the original spatial feature matrix.

[0116] Time-varying dynamic features and spatially sensitive features are input into a multilayer perceptron for feature stitching and nonlinear mapping to generate multimodal fusion features of bridge demolition during the demolition process of the target overpass bridge.

[0117] The multimodal fusion feature for bridge demolition can be generated by a spatiotemporal graph convolutional network based on cross-modal structural response monitoring information and degradation-interference coupling feature information. This multimodal fusion feature can comprehensively reflect the structural response, degradation characteristics, and construction interference during bridge demolition, providing high-quality input for stability prediction models. In this embodiment, the multimodal fusion feature for bridge demolition can be generated by fusing cross-modal structural response monitoring information and degradation-interference coupling feature information through a spatiotemporal graph convolutional network. Taking the demolition of elevated bridges in densely populated urban areas as an example, the structural stability prediction method for the demolition of overpass bridges in this embodiment can be as follows: First, dynamic response data such as displacement and stress of the bridge during the demolition process are continuously monitored by multiple sensors. Then, displacement time history and stress spectrum are extracted using cross-modal structural response monitoring information, and combined with degradation-interference coupling feature information, and input into a spatiotemporal graph convolutional network for multimodal feature fusion. Next, in the time dimension, the transient effects of support system degradation and the cumulative effects of cross-line construction interference are captured by time gating units and memory decay gates, respectively, to generate time-varying dynamic features. In the spatial dimension, based on the spatial topology graph constructed by the bridge finite element model, spatial importance weight vectors are generated by performing graph convolution operations and global pooling on node features. Finally, the time-varying dynamic features and spatially sensitive features are input into a multilayer perceptron for feature splicing and nonlinear mapping to generate multimodal fusion features of bridge demolition with physical meaning and dynamic representation capabilities, thereby achieving accurate temporal prediction of the structural stability state during the demolition process.

[0118] In one embodiment, a bridge demolition structural stability prediction model is constructed. The multimodal fusion features of bridge demolition are input into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure at the current demolition stage, obtaining bridge structural stability prediction information, specifically including:

[0119] Based on the historical bridge demolition database, historical demolition cases with different stability states are retrieved. Based on the historical demolition cases, the historical structural response characteristics during the demolition process of each case are extracted to obtain historical structural response characteristic information.

[0120] The historical bridge demolition database can be a collection storing a large number of completed bridge demolition cases and related data. It can provide rich historical data support for model training, helping to improve the accuracy and generalization ability of the prediction model. Furthermore, the historical bridge demolition database can be established by collecting and organizing monitoring data, construction records, and other information from past actual projects. For example, the historical bridge demolition database can be implemented using SQL queries or database API calls. Historical demolition cases can be detailed records of past successful or failed bridge demolition projects retrieved and extracted from the historical bridge demolition database. These can be used as training samples for feature extraction and model building. Furthermore, historical demolition cases can be retrieved and extracted from the historical bridge demolition database. Historical structural response feature information can be response features of the bridge structure during the demolition process extracted from historical demolition cases. This can be used to provide input data for model training, helping the model learn structural response patterns at different demolition stages.

[0121] Furthermore, historical structural response characteristic information can be obtained by analyzing and processing data from historical demolition cases. Retrieving historical demolition cases with different stability states from a historical bridge demolition database can be achieved by querying and extracting such cases. This operation can be further implemented through SQL queries, database API calls, etc., thus providing diverse training samples covering different stability states. Extracting historical structural response characteristics from each historical demolition case during the demolition process can yield historical structural response characteristic information by processing and analyzing monitoring data from historical demolition cases to extract key features. This operation can be further implemented through feature selection algorithms, feature engineering, etc., thus providing high-quality input data for model training.

[0122] Extract the demolition process sequence corresponding to each historical demolition case, divide the historical structural response feature information into process partitions, and generate historical feature subsets corresponding to different demolition stages;

[0123] The demolition sequence can be the sequential arrangement of each step in the bridge demolition process. It can be used to partition historical structural response feature information by process, facilitating subsequent analysis. Furthermore, the demolition sequence can be derived from construction records in historical demolition cases. Historical feature subsets can be subsets of historical structural response feature information corresponding to different demolition stages. They can be used to provide phased feature data for model training, helping to capture feature changes at different stages. Furthermore, historical feature subsets can be generated by partitioning historical structural response feature information by process. Extracting the demolition sequence corresponding to each historical demolition case can be achieved by extracting detailed demolition process arrangements from historical demolition cases. Furthermore, this operation can be achieved through text parsing; structured data extraction, etc., thus providing a foundation for subsequent process partitioning. Partitioning historical structural response feature information by process to generate historical feature subsets corresponding to different demolition stages can be achieved by dividing historical structural response feature information into different subsets based on the demolition sequence. Furthermore, this operation can be achieved through timestamp-based partitioning; process identifier-based partitioning, etc., thus facilitating the model's capture of feature changes at different demolition stages.

[0124] The historical structural response characteristics after process partitioning are associated with the corresponding structural stability level, and a process-structure response topology is constructed with demolition process as node label;

[0125] The process-structure response topology graph can be a graphical structure representing the relationship between structural response features between different demolition stages, with demolition processes as node labels. It can help the model understand the mutual influence between different demolition stages, improving prediction accuracy. Furthermore, the process-structure response topology graph can be constructed based on historical structural response feature information and demolition process sequences. Associating the historical structural response feature information after process partitioning with the corresponding structural stability level can be achieved by labeling each subset of historical features with its corresponding structural stability level. This operation can be implemented through manual labeling or automatic labeling, thus providing labeled data for model training. Constructing the process-structure response topology graph with demolition processes as node labels can be achieved by using demolition processes as nodes and structural response features as edges. This operation can be implemented through graph theory algorithms or network construction tools, thus helping the model understand the relationship between different demolition stages.

[0126] A bridge demolition structure stability prediction model is constructed based on gated recurrent units and graph attention networks. A subset of historical structural response features is input into the gated recurrent units to extract multi-scale process time-series features.

[0127] The multi-scale process temporal features can be features with different time scales extracted from a subset of historical features through a gated recurrent unit (GRU). These features can be used to capture structural response patterns at different time scales, enhancing the model's temporal dynamics. Furthermore, multi-scale process temporal features can be extracted by inputting a subset of historical features into a GRU. Constructing a bridge demolition structural stability prediction model based on a GRU and graph attention network can involve designing and implementing a deep learning model combining GRU and GAT. This operation can be further implemented using TensorFlow / Keras or PyTorch, enabling the construction of prediction models capable of handling multimodal data and considering process temporal characteristics. Extracting multi-scale process temporal features by inputting a subset of historical structural response features into a GRU can also be achieved by inputting the historical feature subset into a GRU. This operation can be further implemented using a single-layer GRU or multi-layer GRU stacking, thus capturing structural response patterns at different time scales.

[0128] The adjacency matrix is ​​obtained through the topological graph. Node features are aggregated and node representations are updated based on the attention mechanism. Hierarchical pooling is performed on the node representations of the entire graph to obtain process space features. Multi-scale process temporal features and process space features are fused to generate process-structure spatiotemporal features.

[0129] The adjacency matrix can be a matrix representing the relationships between nodes in the process-structure response topology graph. It can be used for node feature aggregation and updating in graph attention networks. Furthermore, the adjacency matrix can be extracted from the process-structure response topology graph. Node features can be the feature vectors represented by each node in the process-structure response topology graph. They can be used for feature aggregation and updating in graph attention networks. Furthermore, node features can be extracted from historical structural response feature information. The full graph node representation can be a comprehensive representation of all nodes in the entire topology graph after processing by the graph attention network. It can be used for further feature fusion and classification. Furthermore, the full graph node representation can be obtained by aggregating and updating node features through the graph attention network. Process spatial features can be features obtained by hierarchical pooling of the full graph node representation. They can be used to capture spatial features at different demolition stages, enhancing the model's spatial expressive power. Furthermore, process spatial features can be achieved by hierarchical pooling of the full graph node representation. Process-structure spatiotemporal features can be comprehensive features generated by fusing multi-scale process temporal features with process spatial features. They can be used to provide comprehensive spatiotemporal features for final stability level prediction.

[0130] Furthermore, the spatiotemporal features of the process-structure can be achieved by combining multi-scale process temporal features with process spatial features through feature fusion techniques. Obtaining the adjacency matrix through the topological graph can be achieved by extracting the adjacency matrix from the process-structure response topological graph. This operation can be further implemented using graph theory algorithms or graph data processing libraries, thus providing node connection information for the graph attention network. Aggregating node features and updating node representations based on attention mechanisms can be achieved by using attention mechanisms to weighted aggregate node features and update node representations. This operation can be further implemented using self-attention mechanisms or multi-head attention mechanisms, thereby enhancing the model's focus on important features. Obtaining process spatial features by performing hierarchical pooling on the full graph node representation can be achieved by performing hierarchical pooling on the full graph node representation to extract spatial features. This operation can be further implemented using max pooling or average pooling, thereby capturing spatial features at different demolition stages. Fusing multi-scale process temporal features with process spatial features to generate process-structure spatiotemporal features can be achieved by fusing multi-scale process temporal features with process spatial features. Furthermore, this operation can be achieved through feature concatenation, feature weighted fusion, and other methods, thereby generating comprehensive spatiotemporal features and improving prediction accuracy.

[0131] The process-structure spatiotemporal features are input into the softmax classification layer for stability level prediction. The prediction results are verified by confusion matrix and hyperparameters are tuned by reserving a validation set. After iterative training, a bridge demolition structure stability prediction model that meets engineering accuracy is obtained.

[0132] The softmax classification layer can be used to map the spatiotemporal features of the process-structure relationship to probability distributions for various stability levels. It can be used to output bridge structure stability prediction results. Furthermore, the softmax classification layer can be implemented as the last layer of a neural network, using the softmax function for classification. The confusion matrix can be a tool for evaluating the model's prediction performance, showing the matching between the actual and predicted categories. It can help evaluate the model's classification performance and guide hyperparameter tuning. Furthermore, the confusion matrix can be generated by comparing the actual labels and predicted labels on the validation set. Hyperparameters are parameters that need to be manually set during model training, such as learning rate and batch size. They can affect the model's training effect and convergence speed and need to be tuned using the validation set.

[0133] Furthermore, hyperparameters can be determined empirically or through automatic search methods. Inputting the process-structure spatiotemporal features into a softmax classification layer for stability level prediction can be achieved by inputting the process-structure spatiotemporal features into the softmax classification layer and outputting stability level probabilities. This operation can be further implemented using single-layer softmax, multi-layer softmax, etc., to output bridge structure stability prediction results. Validating the prediction results using a confusion matrix and tuning hyperparameters with a reserved validation set can be done by evaluating model performance and adjusting hyperparameters using a reserved validation set. This operation can be further implemented using grid search, random search, etc., to optimize model performance and improve prediction accuracy. A bridge demolition structure stability prediction model that meets engineering accuracy is obtained through iterative training, which can be achieved by gradually optimizing model parameters through multiple iterations of training. This operation can be further implemented using batch gradient descent, mini-batch gradient descent, etc., to obtain a high-precision prediction model.

[0134] The multimodal fusion features of bridge demolition are obtained and then input into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure in the current demolition stage, thus obtaining bridge structural stability prediction information.

[0135] The acquisition of multimodal fusion features for bridge demolition can be achieved by extracting these features from preprocessed cross-modal structural response monitoring information. This provides high-quality input data for model prediction. Furthermore, this operation can be implemented through feature selection and feature engineering. Inputting the multimodal fusion features into a trained bridge demolition structural stability prediction model to predict the stability of the bridge structure at the current demolition stage can also be done by inputting the multimodal fusion features into the trained model for stability prediction. This operation can be implemented through single-shot prediction and rolling prediction, thus providing real-time bridge structural stability prediction. Obtaining bridge structural stability prediction information can be achieved by the model outputting the bridge structural stability prediction results for the current demolition stage. This operation can be implemented through direct output or output via API interface, thus providing a basis for construction decisions.

[0136] For example, in the scenario of urban bridge demolition, the structural stability prediction method for overpass bridge demolition in this embodiment can be as follows: First, retrieve historical demolition cases with different stability states from the historical bridge demolition database and extract the historical structural response features from these cases; then, partition these features according to the demolition process sequence to generate historical feature subsets for different demolition stages; next, associate these feature subsets with the corresponding structural stability levels and construct a process-structure response topology graph; then, construct a prediction model using gated recurrent units and graph attention networks, extract multi-scale process temporal features and process spatial features, and fuse them to generate process-structure spatiotemporal features; finally, input these features into a softmax classification layer for stability level prediction, and fine-tune the model using a reserved validation set to obtain a prediction model that meets engineering accuracy requirements. In the actual demolition process, by acquiring the multimodal fusion features of the current demolition stage and inputting them into the trained model, the stability of the bridge structure can be predicted in real time, providing a basis for construction decisions.

[0137] In one embodiment, based on bridge structural stability prediction information, it is determined whether the structural stability of the current demolition stage meets the safety threshold. If it does not, the current demolition plan is optimized to provide auxiliary decision-making for demolition construction, specifically including:

[0138] Obtain bridge structural stability prediction information, and generate a time-varying curve of structural safety at the current demolition stage based on the bridge structural stability prediction information. The time-varying curve of structural safety represents the safety reserve coefficient of key bridge sections at different demolition times.

[0139] The time-varying curve of structural safety can be a curve characterizing the safety reserve coefficient of a bridge's key sections under different demolition sequences. It can be used to help intuitively understand and assess the bridge's safety during demolition. In this embodiment, the time-varying curve of structural safety can be generated using bridge structural stability prediction information, reflecting the trend of safety changes during the bridge's demolition process.

[0140] Obtain the standard safety threshold for the current demolition stage, calculate the safety deviation value at each time node by combining the time-varying curve of structural safety, and compare the calculated deviation value with the preset risk tolerance range.

[0141] The standard safety threshold can be the minimum safety requirement set according to relevant specifications and standards during bridge demolition construction. It can be used as a benchmark to determine whether the current demolition stage meets the safety requirements. In this embodiment, the standard safety threshold can be determined according to national or industry standards, design specifications, and other documents. The safety deviation value can be the difference between the actual safety reserve coefficient at each time node on the time-varying curve of structural safety and the standard safety threshold. It can be used to quantify the degree of safety deviation in the current demolition stage. In this embodiment, the safety deviation value can be obtained by calculating the difference between the data points on the time-varying curve of structural safety and the standard safety threshold. The preset risk tolerance range can be a pre-set acceptable range for the safety deviation value. It can be used as a basis for determining whether the demolition plan needs to be adjusted. In this embodiment, the preset risk tolerance range can be set according to engineering experience and safety requirements.

[0142] If the deviation value exceeds the preset risk tolerance range, it is determined that the current demolition control parameters do not meet the structural safety requirements, and then historical cases that meet the target bridge type and degradation status are retrieved from the preset historical demolition scheme library.

[0143] The historical demolition plan database can be a collection of data storing similar bridge demolition cases and their corresponding demolition control parameters. It can be used to provide a reference and optimization basis for the current demolition plan. In this embodiment, the historical demolition plan database can be established by collecting and organizing data from historical demolition projects. The demolition control parameter set can contain specific parameter settings for various operations during the demolition process, such as demolition sequence, speed, and force. It can be used to guide the actual demolition construction process. In this embodiment, the demolition control parameter set can be extracted from the historical demolition plan database or set through expert experience.

[0144] Extract the demolition control parameter set corresponding to historical cases, introduce a constrained multi-objective particle swarm optimization algorithm, construct a dual objective function with the goal of maximizing safety and construction efficiency, and set displacement limit and stress threshold as constraints.

[0145] The displacement limit can be the maximum allowable displacement during demolition. It can be used to ensure the structural stability and safety during demolition. In this embodiment, the displacement limit can be set according to the bridge structural characteristics and construction safety requirements. The stress threshold can be the maximum allowable stress value during demolition. It can be used to prevent structural damage due to excessive stress during demolition. In this embodiment, the stress threshold can be set according to the material mechanical properties and structural design requirements.

[0146] Initialize the particle swarm position as the historical demolition parameter vector, calculate the non-dominated solution front and evaluate the individual crowding distance, select the safest particle in the Pareto front as the guide particle, and update the velocity-position equation of the following particles.

[0147] In a multi-objective optimization problem, the non-dominated solution front can be defined as a set of solutions for which no other solution outperforms it across all objectives. It can be used to provide multiple feasible optimization options. In this embodiment, the non-dominated solution front can be calculated using a multi-objective optimization algorithm. Individual crowding distance, in multi-objective optimization, measures the distance between adjacent individuals on the non-dominated solution front. It can be used to evaluate the distribution of solutions and avoid local optima. In this embodiment, the individual crowding distance can be obtained by calculating the Euclidean distance or other distance metrics between adjacent individuals on the non-dominated solution front. The Pareto front, also in a multi-objective optimization problem, is a set of solutions for which no other solution outperforms it across all objectives. It can be used to provide multiple feasible optimization options. In this embodiment, the Pareto front can be calculated using a multi-objective optimization algorithm.

[0148] By iteratively optimizing the output set of feasible demolition parameters, the feasibility of construction is verified using the set of feasible demolition parameters. Based on the verification results, the optimal demolition optimization scheme is selected to generate a dynamic demolition instruction set, and the demolition construction is assisted in decision-making through a digital twin platform.

[0149] The optimal demolition optimization scheme can be the best demolition scheme after multi-objective optimization, comprehensively considering safety and construction efficiency. It can be used to ensure that the demolition process is both safe and efficient. In this embodiment, the optimal demolition optimization scheme can be obtained through iterative optimization using a constrained multi-objective particle swarm optimization algorithm. The dynamic demolition instruction set can be a series of specific demolition operation instructions generated based on the optimal demolition optimization scheme. It can be used to guide the actual demolition construction process. In this embodiment, the dynamic demolition instruction set can be generated after the optimization algorithm outputs and the construction feasibility is verified. The digital twin platform can be a virtual simulation technology that uses real-time data to drive the simulation and visualization of physical systems. It can be used to provide visualization and auxiliary decision support for the demolition construction process. In this embodiment, the digital twin platform can be constructed by combining sensor data, models, and algorithms.

[0150] Taking the demolition of an urban overpass as an example, the structural stability prediction method for the demolition of an overpass in this embodiment can be as follows: First, obtain the structural stability prediction information of the bridge and generate a time-varying curve of structural safety. Then, obtain the standard safety threshold, calculate the safety deviation value at each time node, and compare it with the preset risk tolerance range. If the deviation value exceeds the preset risk tolerance range, retrieve historical cases from the historical demolition scheme library that meet the target bridge type and degradation state. Next, extract the demolition control parameter set corresponding to the historical cases, introduce a constrained multi-objective particle swarm optimization algorithm, construct a dual objective function with maximizing safety and maximizing construction efficiency, and set displacement limit and stress threshold as constraints. Initialize the particle swarm position as the historical demolition parameter vector, calculate the non-dominated solution front and evaluate the individual congestion distance, select the safest particle in the Pareto front as the guiding particle, and update the velocity-position equation of the following particles. Output the feasible demolition parameter set through iterative optimization, verify the construction feasibility, select the optimal demolition optimization scheme to generate a dynamic demolition instruction set, and use a digital twin platform for demolition construction auxiliary decision-making.

[0151] Furthermore, to achieve the above objectives, the present invention also provides a system for predicting the structural stability of an overpass bridge during demolition, the system comprising:

[0152] The monitoring and preprocessing module is used to monitor the demolition construction of the target overpass bridge, obtain structural response monitoring data, and preprocess the obtained structural response monitoring data to obtain cross-modal structural response monitoring information.

[0153] The feature fusion module is used to analyze the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data based on the cross-modal structural response monitoring information, and to construct multi-modal fusion features to obtain the multi-modal fusion features of bridge demolition.

[0154] The stability prediction module is used to construct a stability prediction model for bridge demolition structures. The multimodal fusion features of bridge demolition are input into the trained bridge demolition structure stability prediction model to predict the stability of the bridge structure in the current demolition stage and obtain bridge structure stability prediction information.

[0155] The decision optimization module is used to determine whether the structural stability of the current demolition stage meets the safety threshold based on the bridge structural stability prediction information. If it does not meet the safety threshold, the module optimizes the current demolition plan to make auxiliary decisions for demolition construction.

[0156] Other embodiments or specific implementations of the overpass bridge demolition structure stability prediction system of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0157] In addition, to achieve the above objectives, the present invention also provides a device for predicting the stability of an overpass demolition structure. The device includes: a memory, a processor, and an overpass demolition structure stability prediction program stored in the memory and executable on the processor. The overpass demolition structure stability prediction program is configured to implement the steps of the overpass demolition structure stability prediction method as described in any one of the above descriptions.

[0158] In addition, to achieve the above objectives, the present invention also provides a medium storing a stability prediction program for the demolition structure of an overpass, wherein when the overpass demolition structure stability prediction program is executed by a processor, the program implements the steps of the stability prediction method for the demolition structure of an overpass as described above.

[0159] In addition, to achieve the above objectives, the present invention also provides a computer program product, including a stability prediction program for the demolition structure of an overpass, wherein when the overpass demolition structure stability prediction program is executed by a processor, it implements the steps of the stability prediction method for the demolition structure of an overpass as described above.

[0160] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting the stability of a demolished overpass structure, characterized in that, The method includes: Structural response monitoring data was obtained by monitoring the demolition construction of the target overpass bridge. The obtained structural response monitoring data was preprocessed to obtain cross-modal structural response monitoring information. Based on the cross-modal structural response monitoring information, the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data is analyzed, and multi-modal fusion features are constructed to obtain the multi-modal fusion features of bridge demolition. A bridge demolition structural stability prediction model is constructed. The multimodal fusion features of the bridge demolition are input into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure in the current demolition stage, thereby obtaining bridge structural stability prediction information. Based on the bridge structural stability prediction information, it is determined whether the structural stability of the current demolition stage meets the safety threshold. If it does not, the current demolition plan is optimized to assist in the demolition construction decision-making.

2. The method for predicting the structural stability of an overpass bridge as described in claim 1, characterized in that, The process involves monitoring the demolition of the target overpass bridge to obtain structural response monitoring data, followed by preprocessing of the acquired structural response monitoring data to obtain cross-modal structural response monitoring information. Specifically, this includes: During the demolition of the target overpass bridge, a network of sensors was deployed to monitor the structural response of key parts of the bridge and obtain structural response monitoring data. The structural response monitoring data includes support system degradation monitoring data, cross-line construction interference monitoring data, and structural dynamic response data. The structural response monitoring data is decomposed into wavelet packets using a preset Morlet wavelet basis function to obtain multi-scale sub-bands, and the Hilbert energy spectrum of each sub-band is calculated. After determining the characteristic frequency band based on the energy spectrum distribution, an improved Wiener filter is used for noise suppression. Spatiotemporal alignment processing is performed on the filtered structural response monitoring data. Adaptive median filtering is used to eliminate impulse noise, and bilateral filtering is used to eliminate displacement ambiguity caused by structural vibration. A sliding window algorithm is introduced to calculate the root mean square value and kurtosis coefficient of the data within the window based on a preset window. Spline interpolation is then used to repair missing data. Z-score standardization was performed on the repaired structural response monitoring data to convert the discrete monitoring data into a continuous probability distribution. K-means clustering was then used to perform load condition binning to generate cross-modal structural response monitoring information.

3. The method for predicting the structural stability of an overpass bridge as described in claim 1, characterized in that, The analysis of the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data based on the cross-modal structural response monitoring information, and the construction of multi-modal fusion features to obtain the multi-modal fusion features of bridge demolition, specifically includes: Obtain cross-modal structural response monitoring information, extract concrete crack image data from the support system degradation monitoring data, and import the concrete crack image data into the Lab color space to obtain chromaticity feature distribution; The mean brightness of the crack region and the gradient variance of the a-channel are calculated by weighted fusion using the chromaticity feature distribution to form the damage index of the support system. Then, the boundary of the main connected region of the crack is extracted by the watershed algorithm, and the ratio of the boundary curvature to the crack area is calculated as the structural degradation sensitivity index. The first feature information is generated by combining the crack texture fractal dimension. The construction vibration acceleration time history signal is extracted from the cross-line construction interference monitoring data, the time-frequency domain cross spectral density is calculated to generate the vibration interference feature spectrum, and the vibration interference feature spectrum is input into the pre-trained deep residual network to invert the construction load transfer path, generate the construction interference dynamic feature vector, and obtain the second feature information. A cross-modal attention gating mechanism is introduced, using the first feature information as a state vector and the second feature information as an input vector and a hidden vector. The activation weights of the gating unit are calculated to perform feature interaction fusion and generate degradation-interference coupled feature information. Displacement time history and stress spectrum are extracted from the structural dynamic response data using cross-modal structural response monitoring information. Combined with the degradation-interference coupling feature information, these are input into a spatiotemporal graph convolutional network for multimodal feature fusion to obtain the multimodal fusion features of bridge demolition.

4. The method for predicting the stability of a bridge structure after demolition as described in claim 3, characterized in that, The displacement time history and stress spectrum in the structural dynamic response data are extracted using cross-modal structural response monitoring information. These are then combined with the degradation-interference coupling feature information and input into a spatiotemporal graph convolutional network for multimodal feature fusion to obtain the bridge demolition multimodal fusion features, specifically including: In the time dimension, the data input to the spatiotemporal graph convolutional network is modeled as a sequence of spatiotemporal nodes, where each node contains the support damage state, construction vibration spectrum, beam displacement and stress components at the current moment; The forward propagation layer captures the transient effects of support system degradation on the structural response through time-gated units, while the backward propagation layer uses memory decay gates to filter the cumulative effects of cross-line construction disturbances and generates time-varying dynamic characteristics. In the spatial dimension, a spatial topology graph is constructed based on the bridge finite element model. Graph convolution is performed on the node features to obtain a spatial feature matrix. Global pooling with node degree weighting is then performed on the spatial feature matrix to generate a spatial importance weight vector. Obtain the hidden states of nodes in the spatiotemporal graph convolutional network, perform tensor shrunk with the spatial importance weight vector, calculate the structural contribution score of each node in the dynamic response, and perform node weighting on the original spatial feature matrix to obtain spatially sensitive features. The time-varying dynamic features and spatially sensitive features are input into a multilayer perceptron for feature stitching and nonlinear mapping to generate multimodal fusion features of bridge demolition during the demolition process of the target overpass bridge.

5. The method for predicting the stability of a bridge structure after demolition as described in claim 1, characterized in that, The construction of the bridge demolition structural stability prediction model involves inputting the multimodal fusion features of the bridge demolition into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure at the current demolition stage, thereby obtaining bridge structural stability prediction information. Specifically, this includes: Based on the historical bridge demolition database, historical demolition cases with different stability states are retrieved. Based on the historical demolition cases, the historical structural response characteristics during the demolition process of each case are extracted to obtain historical structural response characteristic information. Extract the demolition process sequence corresponding to each historical demolition case, divide the historical structure response feature information into process partitions, and generate historical feature subsets corresponding to different demolition stages; The historical structural response characteristics after process partitioning are associated with the corresponding structural stability level, and a process-structure response topology is constructed with demolition process as node label; A bridge demolition structural stability prediction model is constructed based on gated recurrent units and graph attention networks. A subset of historical structural response features is input into the gated recurrent units to extract multi-scale process time-series features. The adjacency matrix is ​​obtained through the topological graph. Node features are aggregated and node representations are updated based on the attention mechanism. Hierarchical pooling is performed on the node representations of the entire graph to obtain process space features. The multi-scale process temporal features and process space features are fused to generate process-structure spatiotemporal features. The process-structure spatiotemporal features are input into a softmax classification layer for stability level prediction. The prediction results are verified by a confusion matrix and the hyperparameters are tuned by reserving a validation set. After iterative training, a bridge demolition structure stability prediction model that meets engineering accuracy is obtained. The multimodal fusion features of bridge demolition are obtained and then input into the trained bridge demolition structural stability prediction model to predict the stability of the bridge structure in the current demolition stage, thereby obtaining bridge structural stability prediction information.

6. The method for predicting the stability of a bridge structure after demolition as described in claim 1, characterized in that, The process involves determining whether the structural stability of the current demolition stage meets the safety threshold based on the bridge structural stability prediction information. If it does not, the current demolition plan is optimized to assist in demolition construction decision-making. This specifically includes: Obtain bridge structural stability prediction information, and generate a time-varying curve of structural safety at the current demolition stage based on the bridge structural stability prediction information. The time-varying curve of structural safety represents the safety reserve coefficient of key bridge sections at different demolition times. Obtain the standard safety threshold for the current demolition stage, calculate the safety deviation value at each time node based on the time-varying curve of structural safety, and compare the calculated deviation value with the preset risk tolerance range. If the deviation value exceeds the preset risk tolerance range, it is determined that the current demolition control parameters do not meet the structural safety requirements, and then historical cases that meet the target bridge type and degradation status are retrieved from the preset historical demolition scheme library. Extract the demolition control parameter set corresponding to historical cases, introduce a constrained multi-objective particle swarm optimization algorithm, construct a dual objective function with the goal of maximizing safety and construction efficiency, and set displacement limit and stress threshold as constraints. Initialize the particle swarm position as the historical demolition parameter vector, calculate the non-dominated solution front and evaluate the individual crowding distance, select the safest particle in the Pareto front as the guide particle, and update the velocity-position equation of the following particles. By iteratively optimizing the output set of feasible demolition parameters, the feasibility of construction is verified using the set of feasible demolition parameters. Based on the verification results, the optimal demolition optimization scheme is selected to generate a dynamic demolition instruction set, and the demolition construction is assisted in decision-making through a digital twin platform.

7. A system for predicting the structural stability of an overpass bridge during demolition, characterized in that, The system includes: The monitoring and preprocessing module is used to monitor the demolition construction of the target overpass bridge, obtain structural response monitoring data, and preprocess the obtained structural response monitoring data to obtain cross-modal structural response monitoring information. The feature fusion module is used to analyze the cross-modal correlation between the degradation data of the elevated support system and the cross-line construction interference data based on the cross-modal structural response monitoring information, and to construct multi-modal fusion features to obtain the multi-modal fusion features of bridge demolition. The stability prediction module is used to construct a stability prediction model for bridge demolition structures. The multimodal fusion features of bridge demolition are input into the trained bridge demolition structure stability prediction model to predict the stability of the bridge structure in the current demolition stage and obtain bridge structure stability prediction information. The decision optimization module is used to determine whether the structural stability of the current demolition stage meets the safety threshold based on the bridge structural stability prediction information. If it does not meet the safety threshold, the module optimizes the current demolition plan to make auxiliary decisions for demolition construction.

8. A device for predicting the structural stability of an overpass bridge during demolition, characterized in that, The device includes: a memory, a processor, and an overpass demolition structure stability prediction program stored in the memory and executable on the processor, the overpass demolition structure stability prediction program being configured to implement the steps of the overpass demolition structure stability prediction method as described in any one of claims 1 to 6.

9. A medium, characterized in that, The medium stores a stability prediction program for the demolition structure of an overpass, and when the processor executes the stability prediction program for the demolition structure of an overpass, it implements the steps of the stability prediction method for the demolition structure of an overpass as described in any one of claims 1 to 6.

10. A computer program product, comprising a stability prediction program for the demolition structure of an overpass, characterized in that, When the overpass demolition structure stability prediction program is executed by the processor, it implements the steps of the overpass demolition structure stability prediction method as described in any one of claims 1 to 6.

Images