Support structure stress overrun early warning method

Abstract

The application discloses a kind of support structure stress over-limit early warning method, comprising: configuration multimodal stress perception array, and carry out time synchronization, generate multimodal stress original data matrix, configuration multimodal sensing response mapping model, and execute channel correction.Through constructing multimodal stress perception array, dynamic threshold neural network model, multiscale structure damage evolution discriminant model and cognitive atlas joint decision mechanism mechanism.The application relates to the field of building construction technology, realizes the whole cycle perception and intelligent early warning control of support structure stress state, and can time synchronization, channel correction and dynamic fusion are carried out to the multimodal stress original data matrix collected by fiber grating sensing unit, miniature piezoelectric acceleration sensing unit and acoustic emission array sensing unit under complex environmental conditions.

Description

Technical Field

[0001] This invention relates to the field of building construction technology, and in particular to a method for early warning of stress exceeding limits in support structures. Background Technology

[0002] In underground engineering, slope protection, foundation pit and tunnel construction, the support structure bears the long-term constraint of the surrounding rock and the transfer of loads. Its safety is directly related to the structural stability and construction safety. With the development of monitoring technology, the stress monitoring system of the support structure has gradually evolved from single sensor measurement to multimodal perception and data-driven analysis. Existing stress monitoring schemes usually acquire data by deploying strain gauges, accelerometers or acoustic emission sensors, and execute over-limit alarms based on static thresholds or empirical criteria. However, due to the coupling influence of multiple factors such as ambient temperature and humidity, pore water pressure and construction disturbances, the stress response of the support structure exhibits obvious time-varying and nonlinear characteristics. Traditional fixed threshold models are difficult to accurately reflect the true state of the structure, thus affecting the reliability and timeliness of the early warning.

[0003] Existing methods for stress monitoring and early warning of support structures have two technical problems. First, they lack unified analysis of multimodal data and dynamic threshold adaptive capability, making it impossible to adjust the stress judgment benchmark in real time in complex environments, resulting in insufficient accuracy in stress over-limit identification. Second, they lack a multi-spatial-scale structural damage evolution identification mechanism and a digital twin cascade analysis framework, making it impossible to achieve dynamic mapping of structural state and risk propagation path reasoning, and difficult to form an intelligent early warning and decision-making closed loop.

[0004] Therefore, there is a need to provide a method for early warning of stress exceeding limits in support structures, which can solve the above-mentioned technical problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for early warning of stress exceeding limits in support structures, which can solve the above-mentioned technical problems.

[0006] This invention is implemented as follows:

[0007] A method for early warning of stress exceeding limits in support structures includes the following steps:

[0008] Step S100: Configure the multimodal stress sensing array and perform time synchronization, generate the original multimodal stress data matrix, configure the multimodal sensing response mapping model, and perform channel correction.

[0009] Step S200: Extract the original multimodal stress data matrix, calculate the stress evolution rate and acoustic emission energy flux, determine the correlation, define the stress-energy coupling index, and set the sensitivity judgment range.

[0010] Step S300: Configure a dynamic threshold neural network model that integrates multiple input sources, outputs a time-varying dynamic stress threshold, sets an early warning criterion, and executes dynamic environmental factor correction and model stability constraint mechanisms.

[0011] Step S400: Configure the multi-scale stress input sequence, generate the spatiotemporal pyramid input tensor, the multi-scale structural damage evolution discrimination model, and output the structural damage evolution index. Set the structural damage evolution index and growth rate criteria.

[0012] Step S500: Configure edge computing nodes and cloud intelligent center, establish a multi-level collaborative computing architecture, configure real-time early warning triggering mechanism and event package generation mechanism, configure edge consistency judgment mechanism and cloud risk escalation response mechanism;

[0013] Step S600: Configure the digital twin model of the support structure, perform real-time state mapping, configure the cognitive graph structure and knowledge reasoning engine, configure the risk triplet output mechanism, and generate a structured early warning report.

[0014] Step S100 includes the following sub-steps:

[0015] Step S100.1: Configure the multimodal stress sensing array and perform time synchronization to generate the original multimodal stress data matrix;

[0016] Specifically, fiber optic grating sensing units, miniature piezoelectric accelerometer sensing units, and acoustic emission array sensing units are selected from the sensor spare parts library. Then, using installation fixtures and clamping components, the fiber optic grating sensing units, miniature piezoelectric accelerometer sensing units, and acoustic emission array sensing units are fixed at typical stress positions of reinforced concrete structural units, concrete structural units, and soil contact structural units to form a support structure stress sensing array.

[0017] Fiber grating sensing units are used to acquire the micro-strain response of steel reinforcement and concrete structural units, micro piezoelectric accelerometers are used to acquire the local dynamic acceleration response caused by load changes, and acoustic emission array sensing units are used to acquire the high-frequency acoustic signals generated by microcrack initiation and propagation. The point zoning is set according to the structural risk level.

[0018] An embedded time synchronization module is assembled from an embedded control board, and the signal acquisition channel of the support structure stress sensing array is connected to the acquisition port of the embedded time synchronization module to obtain the raw acquisition signals generated by the fiber optic grating sensing unit, the miniature piezoelectric acceleration sensing unit, and the acoustic emission array sensing unit.

[0019] The embedded time synchronization module performs time alignment processing on the raw acquisition signal with millisecond-level precision, and writes three-dimensional layout coordinate information and acquisition timestamp for each acquisition, which are aggregated to form a multimodal stress raw data matrix. The multimodal stress raw data matrix has spatial positioning attributes and temporal consistency attributes, and serves as the input data carrier for subsequent stress evolution rate calculation and dynamic early warning model training.

[0020] Step S100.2: Configure the multimodal sensing response mapping model and perform channel correction;

[0021] Specifically, the data acquisition program starts the self-calibration operation cycle in the first hour after the system is started. The self-calibration operation cycle includes a static load data acquisition stage and an unloaded data acquisition stage. The static load data acquisition stage uses the structure's self-weight and predetermined working conditions to obtain the static load response curve, while the unloaded data acquisition stage obtains the baseline response curve under the condition of no external disturbance.

[0022] After obtaining the static load response curve and the baseline response curve, the difference between the two types of curves is calculated and fitted to establish a multimodal sensing response mapping model covering the fiber optic grating sensing unit, the micro piezoelectric accelerometer sensing unit, and the acoustic emission array sensing unit. Based on the multimodal sensing response mapping model, channel gain parameter correction and zero bias parameter correction are performed.

[0023] Step S200 includes the following sub-steps:

[0024] Step S200.1: Extract the original multimodal stress data matrix, calculate the stress evolution rate and acoustic emission energy flux, and determine the correlation.

[0025] Specifically, stress time series data of reinforced concrete structural units, concrete structural units, and soil contact structural units are extracted from the multimodal stress raw data matrix, and time intervals are set. For any time stress value Compared to the previous moment stress value Differential calculations were performed to obtain the stress evolution rate. The specific calculation formula is as follows:

[0026]

[0027] In the formula: For stress evolution rate, For a moment The stress value, For a moment The stress value, The time interval between adjacent sampling;

[0028] Extracting the acoustic emission signal waveform from the acoustic emission array sensing unit Set the energy calculation window length Calculate time points using integral operations acoustic emission energy flux The specific calculation formula is as follows:

[0029]

[0030] In the formula: For time points Acoustic emission energy flux, For energy calculation window length, For acoustic emission array sensing units in time The instantaneous amplitude signal at a given moment;

[0031] Constructing time synchronization sequences The correlation between the two was calculated using the Pearson correlation coefficient. The specific formula for calculating the Pearson correlation coefficient is as follows:

[0032]

[0033] In the formula: The Pearson correlation coefficient is used. The covariance between stress evolution rate and acoustic emission energy flux. Let V be the variance of the stress evolution rate. Let V be the variance of acoustic emission energy flux;

[0034] The correlation threshold is set to 0.85. When r>0.85, the corresponding spatial point is determined to have entered the microcrack initiation stage and is marked as a crack evolution early warning area.

[0035] Step S200.2: Define the stress-energy coupling index and set the sensitivity judgment range;

[0036] Specifically, a stress-energy coupling index is defined for spatial locations where the correlation determination results meet the conditions. The stress response sensitivity, used to characterize the stress response sensitivity per unit energy release, is calculated using the following formula:

[0037]

[0038] In the formula: As a stress-energy coupling index, For stress evolution rate, For acoustic emission energy flux, the normal fluctuation range of the stress-energy coupling index is set as follows: ,when When the corresponding spatial location is determined to have nonlinear stress release characteristics, the determination result is input into the dynamic threshold self-learning and environmental correction mechanism for subsequent adjustment of risk weight and response level.

[0039] Step S300 includes the following sub-steps:

[0040] Step S300.1: Configure a dynamic threshold neural network model that integrates multiple input sources, outputs a time-varying dynamic stress threshold, and sets an early warning criterion;

[0041] Specifically, configure the input feature sequence, including stress evolution rate. Acoustic emission energy flux Ambient temperature relative humidity Groundwater level With soil pore water pressure A dynamic threshold neural network model with a hybrid structure of long short-term memory neural network and gated recurrent unit is constructed to complete the internal state reset of the gated structure;

[0042] The input features undergo max-min normalization before entering the dynamic threshold neural network model, with the normalization interval set to [value missing]. This is used to unify the feature scale across dimensions and improve the model's convergence stability.

[0043] The time points are output through the trained dynamic threshold neural network model. Corresponding dynamic stress threshold The output function expression is:

[0044]

[0045] In the formula: It is a point in time. The dynamic stress threshold, It is a point in time. The rate of stress evolution, It is a point in time. Acoustic emission energy flux, It is a point in time. Ambient temperature correction factor, It is a point in time. The soil pore water pressure weighting parameter;

[0046] The warning criterion for setting the dynamic stress threshold is: when the measured stress value satisfy: When the following conditions are met, the system enters an early warning preparation state: When the stress exceeds the limit, the system triggers a stress over-limit warning state to avoid repeated state switching caused by threshold fluctuations. A hysteresis recovery mechanism is set: when the stress value drops to meet the threshold... When this happens, the system status changes from early warning preparation or over-limit early warning to normal status;

[0047] Step S300.2: Execute dynamic environmental factor correction and dynamic threshold neural network model stability constraint mechanism;

[0048] Specifically, configure an environmental data normalization processing module to process environmental temperature. air humidity Groundwater level and soil pore water pressure Perform a maximum-minimum normalization operation to normalize all environmental factor input values ​​to the minimum value. The interval is the dynamic stress threshold output by the dynamic threshold neural network model. Set a continuity constraint to limit the threshold change rate between adjacent time points to no more than 10%, with the following judgment condition:

[0049]

[0050] In the formula: It is the current time point. The dynamic stress threshold, It is the dynamic stress threshold at the previous time point. It is the sampling interval between adjacent time points, and 0.10 is the maximum allowable rate of change threshold;

[0051] When the output of the dynamic threshold neural network model fails to meet the above conditions three times consecutively, the system pauses the weight update of the dynamic threshold neural network model and triggers the dynamic threshold neural network model steady-state self-checking mechanism, marking the current state of the dynamic threshold neural network model as requiring external intervention.

[0052] Step S400 includes the following sub-steps:

[0053] Step S400.1: Configure the multi-scale stress input sequence, generate the spatiotemporal pyramid input tensor, the multi-scale structural damage evolution discrimination model, and output the structural damage evolution index.

[0054] Specifically, by using fiber optic grating sensing units, micro piezoelectric accelerometer sensing units, and acoustic emission array sensing units deployed in the support structure, a multimodal stress raw data matrix is ​​collected. Based on the hierarchical topological characteristics of the support structure, the measurement points of the fiber optic grating sensing units, micro piezoelectric accelerometer sensing units, and acoustic emission array sensing units are reconstructed to form stress time series at the steel structure unit level, key node level, component assembly level, and overall structure level, respectively, forming a set of stress sequences at four spatial scales with hierarchical topological relationships.

[0055] A sliding window sampling mechanism is used to uniformly sample the stress sequence set at four spatial scales. The stress time series data at each level are aligned and integrated according to the time order and spatial hierarchy to generate a unified spatiotemporal pyramid input tensor. This spatiotemporal pyramid input tensor is expanded into a time series in the time dimension and maintains the four-level hierarchical structure of steel structure unit level, key node level, component combination level and overall structure level in the spatial dimension.

[0056] Based on the established spatiotemporal pyramid input tensor, a multi-scale structural damage evolution discrimination model is configured. This multi-scale structural damage evolution discrimination model adopts a spatiotemporal pyramid neural network structure with spatiotemporal feature fusion capability. The spatiotemporal pyramid neural network consists of four scale input channels, corresponding to the steel structure unit level, key node level, component combination level, and overall structure level, respectively. Each scale input channel contains a three-layer one-dimensional convolutional neural network, which is used to extract the local time series evolution features of the corresponding level. Cross-level feature association and weighted fusion are realized through a unified fusion attention module. Finally, the structural damage evolution index is output through a fully connected output layer.

[0057] Structural damage evolution index Used to characterize time points The overall damage state at each spatial level is defined as follows: ,in, This indicates that the support structure is in an undamaged state. This indicates that the support structure has reached a critical failure state;

[0058] Step S400.2: Set the structural damage evolution index and growth rate criterion;

[0059] Specifically, the structural damage evolution index at each spatial level Time series differencing is performed to calculate the structural damage evolution growth rate, which is defined as:

[0060]

[0061] In the formula: For time points The rate of evolution and growth of structural damage. The structural damage evolution index at the current time point. The structural damage evolution index at the previous time point. The sampling interval;

[0062] The risk identification threshold for the structural damage evolution index is set as follows: When the structural damage evolution index at any spatial level When the above risk identification threshold is met, the system automatically determines that the spatial level is a potential over-limit risk area and writes the unique structural level identifier and its structural damage evolution time series corresponding to the level into the early warning index queue.

[0063] Step S500 includes the following sub-steps:

[0064] Step S500.1: Configure edge computing nodes and cloud intelligent center, establish a multi-level collaborative computing architecture, and configure real-time early warning triggering mechanism and event package generation mechanism;

[0065] Specifically, a distributed computing architecture consisting of edge computing nodes and a cloud-based intelligent center is configured. The edge computing nodes are used to perform real-time calculation tasks of the multimodal stress raw data matrix, while the cloud-based intelligent center is used to perform global trend analysis and model iteration training tasks. The edge computing nodes integrate embedded edge processing chips and 5G communication modules and are deployed in the near-field position of the stress sensing array of the support structure. The cloud-based intelligent center centrally manages and integrates the stress evolution rate, acoustic emission energy flux, structural damage evolution index, and dynamic stress threshold uploaded by multiple edge computing nodes through a unified data platform. The cache period of the edge computing nodes is set, and a sliding cache mechanism is used to segment and encapsulate the continuously acquired multimodal stress raw data matrix.

[0066] Set the stress evolution rate of edge computing nodes Acoustic emission energy flux With dynamic stress threshold Real-time monitoring of the time series is performed, and the measured stress value satisfies: If the event is detected, it is determined to be an over-limit warning state, and the event package generation and reporting process is executed within 1 second. The event package includes three-dimensional layout coordinates, detection timestamp, stress evolution rate sequence, acoustic emission energy flux sequence, structural damage evolution index sequence, and dynamic stress threshold estimation sequence.

[0067] The event packets are encapsulated using a standardized data structure and transmitted encrypted via a 5G private network to the cloud-based intelligent center for subsequent fusion analysis and instruction distribution.

[0068] Step S500.2: Configure the edge consistency judgment mechanism and the cloud risk escalation response mechanism;

[0069] Specifically, an edge consistency judgment mechanism is configured to determine the consistency of multiple consecutive early warning results from edge computing nodes. The judgment criterion is that the stress judgment results within three consecutive sampling periods all meet the over-limit early warning conditions. When the consistency conditions are met, the edge computing node outputs a level-one risk confirmation signal. The cloud intelligent center is configured with an event fusion analysis module to perform fusion analysis on the event packets uploaded by multiple edge computing nodes and the level-one risk confirmation signal. Based on the fusion results, the cloud intelligent center performs structural damage evolution trend fitting and stress accumulation effect assessment, generates enhanced response instructions, and triggers a level-one risk alarm.

[0070] Step S600 includes the following sub-steps:

[0071] Step S600.1: Configure the digital twin model of the support structure and perform real-time state mapping; configure the cognitive graph structure and knowledge reasoning engine.

[0072] Specifically, a digital twin model of the support structure is configured. The digital twin model of the support structure takes the original multimodal stress data matrix collected by fiber optic grating sensing units, micro piezoelectric accelerometer sensing units and acoustic emission array sensing units as input to establish a digital equivalent structural model of the real-time stress state and deformation state of the support structure monitoring points. The digital twin model of the support structure uses a structural topology mapping algorithm to topologically bind the location of each monitoring point to the type of support structure component, generating a set of digitally mapped entities with topological consistency and response synchronization.

[0073] The digital twin model of the support structure updates the set of state parameters with a fixed refresh cycle. The refresh cycle is set, and the stress state and deformation state of the monitoring points are continuously updated through a real-time refresh mechanism. The state change trajectory obtained from each refresh is written into the dynamic state buffer and used as the input feature set of the cognitive graph knowledge reasoning engine.

[0074] Configure a cognitive graph structure. The cognitive graph structure takes the historical operation data of the support structure and the structural design parameters as input to establish a semantic association network for risk identification and trend prediction. The cognitive graph structure contains a set of semantic nodes: component type semantic nodes, material property semantic nodes, stress threshold semantic nodes, damage evolution path semantic nodes, and risk level semantic nodes.

[0075] Each semantic node is connected through structural association semantic edges, response evolution semantic edges, and risk propagation semantic edges to form a directed graph structure with multi-layered semantic relationships, which is used to represent the stress evolution logic and risk evolution path of the support structure during long-term operation.

[0076] Configure a knowledge reasoning engine module. The knowledge reasoning engine module calculates the similarity between the set of state parameters from the digital twin model of the support structure and the historical semantic path in the cognitive graph structure. It uses a vectorized graph matching algorithm to obtain the similarity "Sim" between the current state path and the historical semantic path, and performs a judgment: "Sim>0.90". When the similarity value meets the judgment condition, the knowledge reasoning engine module determines the risk level semantic node and the response strategy semantic node corresponding to the high similarity path, and generates a risk path backtracking result and response measure recommendation set.

[0077] Step S600.2: Configure the risk triplet output mechanism and generate a structured early warning report;

[0078] Specifically, a risk triplet output mechanism is configured. The risk triplet output mechanism takes the risk path backtracking results and the set of recommended countermeasures output by the knowledge reasoning engine module as input, and generates a risk triplet structure consisting of three elements: risk source semantic node, propagation path semantic node, and countermeasure semantic node.

[0079] The risk source semantic node corresponds to the unique identifier of the support structure component or monitoring node; the propagation path semantic node corresponds to the set of intermediate states in the risk propagation path; and the response measure semantic node corresponds to the valid response operation stored in the historical knowledge graph.

[0080] The risk triplet structure is encapsulated in a standardized data format to form a structured early warning report. The structured early warning report includes a unique event number, trigger timestamp, similarity score, and a set of recommended operation instructions. The structured early warning report is transmitted through the early warning notification module and written to the log database for risk event tracing and decision optimization. The structured early warning report generated by the cognitive graph decision engine serves as the core data basis for classifying the early warning level of support structure stress over-limit, scheduling response instructions, and human-computer interaction feedback, constituting the final intelligent decision output unit of the support structure stress over-limit early warning method.

[0081] Compared with the prior art, the present invention has the following advantages:

[0082] This invention achieves full-cycle perception and intelligent early warning control of the stress state of support structures by constructing a multimodal stress sensing array, a dynamic threshold neural network model, a multi-scale structural damage evolution discrimination model, and a cognitive map joint decision-making mechanism. It can perform time synchronization, channel correction, and dynamic fusion of the multimodal stress raw data matrix collected by fiber optic grating sensing units, micro piezoelectric accelerometer sensing units, and acoustic emission array sensing units under complex environmental conditions. Based on the dynamic threshold neural network model, it realizes self-learning and environmental factor correction of stress thresholds, dynamically identifies stress anomaly trends, and outputs time-varying threshold results. It also combines the spatiotemporal pyramid input tensor with multi-scale structural damage... The damage evolution discrimination model enables multi-scale damage identification and growth rate determination from the steel structure unit level to the overall structure level, and can adjust the stress judgment benchmark in real time in complex environments. At the same time, it further completes real-time event packet generation and consistency determination through the collaborative computing mechanism of edge computing nodes and cloud intelligent center, realizing the collaborative response of rapid edge identification and cloud global analysis. Finally, based on the digital twin model of the support structure and the cognitive graph decision engine, a risk triplet output mechanism is constructed to output a structured early warning report containing risk source, propagation path and countermeasures, thereby realizing high-precision identification of stress over-limit state of support structure, risk trend prediction and intelligent decision-making closed loop. Attached Figure Description

[0083] Figure 1 This is a flowchart of the method for early warning of stress exceeding limits in the support structure according to the present invention. Detailed Implementation

[0084] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0085] Please see the appendix Figure 1 A method for early warning of stress exceeding limits in support structures includes the following steps:

[0086] Step S100: Configure the multimodal stress sensing array and perform time synchronization to generate the original multimodal stress data matrix, configure the multimodal sensing response mapping model, and perform channel correction.

[0087] Step S100 includes the following sub-steps:

[0088] Step S100.1: Configure the multimodal stress sensing array and perform time synchronization to generate the original multimodal stress data matrix.

[0089] Specifically, fiber optic grating sensing units, miniature piezoelectric accelerometers, and acoustic emission array sensing units are selected from the sensor spare parts library. Then, using installation fixtures and clamping components, the fiber optic grating sensing units, miniature piezoelectric accelerometers, and acoustic emission array sensing units are fixed at typical stress locations of the reinforced concrete structure unit, concrete structure unit, and soil contact structure unit to form a support structure stress sensing array.

[0090] Fiber optic grating sensing units are used to acquire the micro-strain response of steel and concrete structural units, micro piezoelectric accelerometers are used to acquire the local dynamic acceleration response caused by load changes, and acoustic emission array sensing units are used to acquire the high-frequency acoustic signals generated by the initiation and propagation of microcracks. The point zoning is set according to the structural risk level, and the spacing between sensing points of the stress sensing array in a single area is set to 1.5 meters to 4 meters.

[0091] An embedded time synchronization module is assembled from the embedded control board, and the signal acquisition channel of the support structure stress sensing array is connected to the acquisition port of the embedded time synchronization module to obtain the raw acquisition signals generated by the fiber optic grating sensing unit, the miniature piezoelectric acceleration sensing unit, and the acoustic emission array sensing unit.

[0092] The embedded time synchronization module performs time alignment processing on the raw acquisition signal with millisecond-level precision, and writes three-dimensional layout coordinate information and acquisition timestamp for each acquisition, which are aggregated to form a multimodal stress raw data matrix. The multimodal stress raw data matrix has spatial positioning attributes and temporal consistency attributes, and serves as the input data carrier for subsequent stress evolution rate calculation and dynamic early warning model training.

[0093] Step S100.2: Configure the multimodal sensing response mapping model and perform channel correction.

[0094] Specifically, the data acquisition program initiates a self-calibration operation cycle within the first hour after system startup. The self-calibration operation cycle includes a static load data acquisition phase and an unloaded data acquisition phase. The static load data acquisition phase utilizes the structure's self-weight and predetermined working conditions to obtain the static load response curve, while the unloaded data acquisition phase obtains the baseline response curve under conditions without external disturbance.

[0095] After obtaining the static load response curve and the baseline response curve, the difference between the two types of curves is calculated and fitted to establish a multimodal sensing response mapping model covering the fiber optic grating sensing unit, the micro piezoelectric accelerometer sensing unit, and the acoustic emission array sensing unit. Based on the multimodal sensing response mapping model, channel gain parameter correction and zero bias parameter correction are performed to control the output error of the corresponding channel within the accuracy error range of ±0.2%.

[0096] Step S200: Extract the original multimodal stress data matrix, calculate the stress evolution rate and acoustic emission energy flux, determine the correlation, define the stress energy coupling index, and set the sensitivity judgment range.

[0097] Step S200 includes the following sub-steps:

[0098] Step S200.1: Extract the original multimodal stress data matrix, calculate the stress evolution rate and acoustic emission energy flux, and determine the correlation.

[0099] Specifically, stress time series data of reinforced concrete structural units, concrete structural units, and soil contact structural units are extracted from the multimodal stress raw data matrix, and time intervals are set. For any time stress value Compared to the previous moment stress value Differential calculations were performed to obtain the stress evolution rate. The specific calculation formula is as follows:

[0100]

[0101] In the formula: For stress evolution rate, For a moment The stress value, For a moment The stress value, The time interval between adjacent sampling.

[0102] Stress evolution rate Used to characterize the rate of change of local stress over time, serving as a key quantitative parameter for identifying the stress concentration trend in support structures.

[0103] Extracting the acoustic emission signal waveform from the acoustic emission array sensing unit Set the energy calculation window length Calculate time points using integral operations acoustic emission energy flux The specific calculation formula is as follows:

[0104]

[0105] In the formula: For time points Acoustic emission energy flux, For energy calculation window length, For acoustic emission array sensing units in time The instantaneous amplitude signal at a given moment.

[0106] Constructing time synchronization sequences The correlation between the two was calculated using the Pearson correlation coefficient. The specific formula for calculating the Pearson correlation coefficient is as follows:

[0107]

[0108] In the formula: The Pearson correlation coefficient is used. The covariance between stress evolution rate and acoustic emission energy flux. Let V be the variance of the stress evolution rate. Let V be the variance of the acoustic emission energy flux.

[0109] Set the correlation threshold as follows When satisfied When the corresponding spatial location is identified as entering the microcrack initiation stage, it is marked as a crack evolution early warning area.

[0110] Step S200.2: Define the stress-energy coupling index and set the sensitivity judgment range.

[0111] Specifically, a stress-energy coupling index is defined for spatial locations where the correlation determination results meet the conditions. The stress response sensitivity, used to characterize the stress response sensitivity per unit energy release, is calculated using the following formula:

[0112]

[0113] In the formula: As a stress-energy coupling index, For stress evolution rate, For acoustic emission energy flux, the normal fluctuation range of the stress-energy coupling index is set as follows: ,when When the corresponding spatial location is determined to have nonlinear stress release characteristics, the determination result is input into the dynamic threshold self-learning and environmental correction mechanism for subsequent adjustment of risk weight and response level.

[0114] Step S300: Configure a dynamic threshold neural network model that integrates multiple input sources, outputs a time-varying dynamic stress threshold, sets an early warning criterion, and executes dynamic environmental factor correction and model stability constraint mechanisms.

[0115] Step S300 includes the following sub-steps:

[0116] Step S300.1: Configure a dynamic threshold neural network model that integrates multiple input sources, outputs a time-varying dynamic stress threshold, and sets early warning criteria.

[0117] Specifically, configure the input feature sequence, including stress evolution rate. Acoustic emission energy flux Ambient temperature relative humidity Groundwater level With soil pore water pressure A dynamic threshold neural network model with a hybrid structure of long short-term memory neural network and gated recurrent unit was constructed. The input feature dimension was set to 6. The training sequence was constructed using a sliding window method with a sliding window length of 3600 seconds and a step size of 60 seconds.

[0118] The dynamic threshold neural network model consists of a long short-term memory neural unit connected in series with a gated recurrent unit. The number of nodes in each layer is set to 32. The activation function is the ReLU function, the loss function is the mean squared error, and the optimizer uses the Adam algorithm for parameter iteration. The weights are updated every 24 hours. Online learning is performed using newly added multi-source incremental data, and the internal state of the gated structure is reset.

[0119] The input features undergo max-min normalization before entering the dynamic threshold neural network model, with the normalization interval set to [value missing]. This is used to unify the feature scale of each dimension and improve the convergence stability of the model.

[0120] The time points are output through the trained dynamic threshold neural network model. Corresponding dynamic stress threshold The output function expression is:

[0121]

[0122] In the formula: It is a point in time. The dynamic stress threshold, It is a point in time. The rate of stress evolution, It is a point in time. Acoustic emission energy flux, It is a point in time. Ambient temperature correction factor, It is a point in time. The soil pore water pressure weighting parameter.

[0123] The warning criterion for setting the dynamic stress threshold is: when the measured stress value satisfy: When the following conditions are met, the system enters an early warning preparation state: When the stress exceeds the limit, the system triggers a stress over-limit warning state to avoid repeated state switching caused by threshold fluctuations. A hysteresis recovery mechanism is set: when the stress value drops to meet the threshold... When this happens, the system status returns from the early warning preparation or over-limit early warning state to the normal state.

[0124] Step S300.2: Execute dynamic environmental factor correction and dynamic threshold neural network model stability constraint mechanism.

[0125] Specifically, configure an environmental data normalization processing module to process environmental temperature. air humidity Groundwater level and soil pore water pressure Perform a maximum-minimum normalization operation to normalize all environmental factor input values ​​to the minimum value. The interval is the dynamic stress threshold output by the dynamic threshold neural network model. Set a continuity constraint to limit the threshold change rate between adjacent time points to no more than 10%, with the following judgment condition:

[0126]

[0127] In the formula: It is the current time point. The dynamic stress threshold, It is the dynamic stress threshold at the previous time point. 0.10 is the sampling interval between adjacent time points, and 0.10 is the maximum allowable rate of change threshold.

[0128] When the output of the dynamic threshold neural network model fails to meet the above conditions three times consecutively, the system pauses the weight update of the dynamic threshold neural network model and triggers the dynamic threshold neural network model steady-state self-checking mechanism, marking the current state of the dynamic threshold neural network model as requiring external intervention.

[0129] Step S400: Configure the multi-spatial-scale stress input sequence, generate the spatiotemporal pyramid input tensor, the multi-scale structural damage evolution discrimination model, and output the structural damage evolution index. Set the structural damage evolution index and growth rate criteria.

[0130] Step S400 includes the following sub-steps:

[0131] Step S400.1: Configure the multi-spatial-scale stress input sequence, generate the spatiotemporal pyramid input tensor, the multi-scale structural damage evolution discrimination model, and output the structural damage evolution index.

[0132] Specifically, multimodal stress raw data matrices are collected by fiber optic grating sensing units, micro piezoelectric accelerometer sensing units, and acoustic emission array sensing units deployed in the support structure. Based on the hierarchical topological characteristics of the support structure, the measurement points of the fiber optic grating sensing units, micro piezoelectric accelerometer sensing units, and acoustic emission array sensing units are reconstructed to form stress time series at the steel structure unit level, key node level, component assembly level, and overall structure level, respectively, forming a set of stress sequences at four spatial scales with hierarchical topological relationships.

[0133] A sliding window sampling mechanism was used to uniformly sample the stress sequence set at four spatial scales. The sampling window length was set to 3600 seconds and the sampling step size was set to 120 seconds. Through the sampling window mechanism, the stress time series data at each level were aligned and integrated according to the time order and spatial hierarchy to generate a unified spatiotemporal pyramid input tensor. This spatiotemporal pyramid input tensor is expanded into a time series of length 30 in the time dimension and maintains a four-level hierarchical structure in the spatial dimension, namely, steel structure unit level, key node level, component combination level and overall structure level.

[0134] Based on the established spatiotemporal pyramid input tensor, a multi-scale structural damage evolution discrimination model is configured. This multi-scale structural damage evolution discrimination model adopts a spatiotemporal pyramid neural network structure with spatiotemporal feature fusion capability. The spatiotemporal pyramid neural network consists of four scale input channels, corresponding to the steel structure unit level, key node level, component combination level, and overall structure level, respectively. Each scale input channel contains a three-layer one-dimensional convolutional neural network, which is used to extract the local time series evolution features of the corresponding level. Cross-level feature association and weighted fusion are realized through a unified fusion attention module. Finally, the structural damage evolution index is output through a fully connected output layer.

[0135] Structural damage evolution index Used to characterize time points The overall damage state at each spatial level is defined as follows: ,in, This indicates that the support structure is in an undamaged state. This indicates that the support structure has reached a critical failure state.

[0136] Step S400.2: Set the structural damage evolution index and growth rate criterion.

[0137] Specifically, the structural damage evolution index at each spatial level Time series differencing is performed to calculate the structural damage evolution growth rate, which is defined as:

[0138]

[0139] In the formula: For time points The rate of evolution and growth of structural damage. The structural damage evolution index at the current time point. This represents the structural damage evolution index at the previous time point. The sampling interval is set to 1 hour.

[0140] The risk identification threshold for the structural damage evolution index is set as follows: When the structural damage evolution index at any spatial level When the above risk identification threshold is met, the system automatically determines that the spatial level is a potential over-limit risk area and writes the unique structural level identifier and its structural damage evolution time series corresponding to the level into the early warning index queue.

[0141] Dynamic identification of precursors to progressive instability of support structures can be achieved without relying on fixed thresholds, effectively reducing false alarms and missed alarms caused by static judgments and improving the timeliness and reliability of structural safety assessments.

[0142] Step S500: Configure edge computing nodes and cloud intelligent center, establish a multi-level collaborative computing architecture, configure real-time early warning triggering mechanism and event package generation mechanism, configure edge consistency judgment mechanism and cloud risk escalation response mechanism.

[0143] Step S500 includes the following sub-steps:

[0144] Step S500.1: Configure edge computing nodes and cloud intelligent center, establish a multi-level collaborative computing architecture, and configure real-time early warning triggering mechanism and event package generation mechanism.

[0145] Specifically, a distributed computing architecture consisting of edge computing nodes and a cloud-based intelligent center is configured. The edge computing nodes are used to perform real-time calculation tasks of the multimodal stress raw data matrix, while the cloud-based intelligent center is used to perform global trend analysis and model iterative training tasks. The edge computing nodes integrate embedded edge processing chips and 5G communication modules and are deployed in the near-field position of the stress sensing array of the support structure. They have data acquisition, compression, normalization, and dynamic threshold determination functions. The cloud-based intelligent center centrally manages and integrates data such as stress evolution rate, acoustic emission energy flux, structural damage evolution index, and dynamic stress threshold uploaded by multiple edge computing nodes through a unified data platform. The buffer period of the edge computing nodes is set to 10 seconds, and a sliding buffer mechanism is used to segment and encapsulate the continuously acquired multimodal stress raw data matrix.

[0146] Set the stress evolution rate of edge computing nodes Acoustic emission energy flux With dynamic stress threshold Real-time monitoring of the time series is performed, and the measured stress value satisfies: If the event is detected, it is determined to be an over-limit warning state, and the event packet generation and reporting process is executed within 1 second. The event packet includes three-dimensional layout coordinates, detection timestamp, stress evolution rate sequence, acoustic emission energy flux sequence, structural damage evolution index sequence, and dynamic stress threshold estimation sequence.

[0147] The event packets are encapsulated using a standardized data structure and transmitted encrypted via a 5G private network to the cloud-based intelligent center for subsequent fusion analysis and instruction distribution.

[0148] Step S500.2: Configure the edge consistency judgment mechanism and the cloud risk escalation response mechanism.

[0149] Specifically, an edge consistency judgment mechanism is configured to determine the consistency of multiple consecutive early warning results from edge computing nodes. The judgment criterion is that the stress judgment results within three consecutive sampling periods all meet the over-limit early warning conditions. When the consistency conditions are met, the edge computing node outputs a level-one risk confirmation signal. The cloud intelligent center is configured with an event fusion analysis module to perform fusion analysis on the event packets uploaded by multiple edge computing nodes and the level-one risk confirmation signal. Based on the fusion results, the cloud intelligent center performs structural damage evolution trend fitting and stress accumulation effect assessment, generates enhanced response instructions, and triggers a level-one risk alarm.

[0150] The cloud-based intelligent center distributes risk instructions, pushes real-time early warnings, schedules regional inspections, and records execution trajectories based on risk levels, forming an edge-cloud integrated closed-loop response mechanism for stress risks in support structures.

[0151] Step S600: Configure the digital twin model of the support structure, perform real-time state mapping, configure the cognitive graph structure and knowledge reasoning engine, configure the risk triplet output mechanism, and generate a structured early warning report.

[0152] Step S600 includes the following sub-steps:

[0153] Step S600.1: Configure the digital twin model of the support structure and perform real-time state mapping, and configure the cognitive graph structure and knowledge reasoning engine.

[0154] Specifically, a digital twin model of the support structure is configured. The digital twin model of the support structure takes the original multimodal stress data matrix collected by fiber optic grating sensing units, micro piezoelectric accelerometer sensing units and acoustic emission array sensing units as input to establish a digital equivalent structural model of the real-time stress state and deformation state of the support structure monitoring points. The digital twin model of the support structure uses a structural topology mapping algorithm to topologically bind the location of each monitoring point to the type of support structure component, generating a set of digitally mapped entities with topological consistency and response synchronization.

[0155] The digital twin model of the support structure updates the set of state parameters with a fixed refresh cycle of 10 seconds. The stress and deformation states of the monitoring points are continuously updated through a real-time refresh mechanism. The state change trajectory obtained from each refresh is written into the dynamic state buffer and used as the input feature set of the cognitive graph knowledge reasoning engine.

[0156] Configure a cognitive graph structure. The cognitive graph structure takes the historical operation data of the support structure and the structural design parameters as input to establish a semantic association network for risk identification and trend prediction. The cognitive graph structure contains a set of semantic nodes: component type semantic nodes, material property semantic nodes, stress threshold semantic nodes, damage evolution path semantic nodes, and risk level semantic nodes.

[0157] The semantic nodes are connected through structural association semantic edges, response evolution semantic edges, and risk propagation semantic edges to form a directed graph structure with multi-layered semantic relationships, which is used to represent the stress evolution logic and risk evolution path of the support structure during long-term operation.

[0158] The cognitive graph structure is stored and queried in the form of a graph database, and has efficient graph traversal and feature matching capabilities.

[0159] The knowledge reasoning engine module is configured to perform similarity calculations on the set of state parameters from the digital twin model of the support structure and the historical semantic paths in the cognitive graph structure. It uses a vectorized graph matching algorithm to obtain the similarity "Sim" between the current state path and the historical semantic path, and performs a judgment: "Sim>0.90". When the similarity value meets the judgment condition, the knowledge reasoning engine module determines the risk level semantic node and the response strategy semantic node corresponding to the high similarity path, and generates a risk path backtracking result and response measure recommendation set.

[0160] Step S600.2: Configure the risk triplet output mechanism and generate a structured early warning report.

[0161] Specifically, a risk triplet output mechanism is configured. The risk triplet output mechanism takes the risk path backtracking results and the set of recommended countermeasures output by the knowledge reasoning engine module as input to generate a risk triplet structure consisting of three elements: risk source semantic node, propagation path semantic node, and countermeasure semantic node.

[0162] The risk source semantic node is a unique identifier for the support structure component or monitoring node; the propagation path semantic node is a set of intermediate states in the risk propagation path; and the response measure semantic node is a valid response operation stored in the historical knowledge graph.

[0163] The risk triplet structure is encapsulated in a standardized data format to form a structured early warning report. The structured early warning report includes a unique event number, trigger timestamp, similarity score, and a set of recommended operation instructions. The structured early warning report is transmitted through the early warning notification module and written to the log database for risk event tracing and decision optimization. The structured early warning report generated by the cognitive graph decision engine serves as the core data basis for classifying the early warning level of support structure stress over-limit, scheduling response instructions, and human-computer interaction feedback, constituting the final intelligent decision output unit of the support structure stress over-limit early warning method.

[0164] This invention aims to achieve time synchronization and channel correction of multi-source sensor data by configuring a multimodal stress sensing array and a multimodal sensing response mapping model, ensuring data consistency and comparability. By establishing a dynamic threshold neural network model that integrates stress evolution rate and acoustic emission energy flux, it achieves self-learning and environmental correction of dynamic stress thresholds, significantly improving the time-varying adaptability of early warning criteria. By constructing a multi-spatial-scale spatiotemporal pyramid input tensor and a multi-scale structural damage evolution discrimination model, it achieves damage evolution identification from single-point monitoring to the overall structure. Furthermore, by combining the collaborative mechanism of edge computing nodes and cloud intelligent centers, it achieves multi-level real-time early warning triggering and event packet reporting. Finally, it introduces a digital twin model of the support structure and a cognitive graph decision engine, and realizes the linkage reasoning and report generation of risk sources, propagation paths, and countermeasures through a risk triplet output mechanism, thereby achieving full-cycle perception, dynamic early warning, and intelligent decision support for the stress state of the support structure.

[0165] Example 1: First, fiber optic grating sensing units, micro piezoelectric accelerometers, and acoustic emission array sensing units are fixedly installed at typical stress locations in the reinforced concrete structural units, concrete structural units, and soil contact structural units of the support structure to form a stress sensing array for the support structure. The spacing between each sensing point is set to 1.5 meters to 4 meters. The fiber optic grating sensing units are used to acquire micro-strain responses, the micro piezoelectric accelerometers are used to acquire local dynamic acceleration responses, and the acoustic emission array sensing units are used to acquire high-frequency acoustic signals. The signal channels of all sensing units are connected to an embedded time synchronization module with a time synchronization accuracy of milliseconds. The synchronization module writes three-dimensional coordinate information and a timestamp for each acquisition, forming a multimodal stress raw data matrix with spatial positioning attributes and temporal consistency, which is used as the input data carrier for subsequent stress evolution rate calculation and dynamic early warning model training.

[0166] Next, a self-calibration cycle is initiated within the first hour of system operation. The self-calibration cycle includes a static load data acquisition phase and an unloaded data acquisition phase. In the static load phase, the static load response curve is obtained under the structural self-weight condition. In the unloaded phase, the baseline response curve is obtained under a undisturbed environment. The difference results of the two types of curves are calculated and fitted to establish a multimodal sensing response mapping model. The gain parameters and zero bias parameters of each channel are corrected through the mapping model to ensure that the channel output error is controlled within ±0.2%, thereby achieving a unified reference for the amplitude and phase of multi-source signals and improving signal reliability and subsequent feature extraction accuracy.

[0167] Subsequently, stress time series data of reinforced concrete structural units, concrete structural units, and contact soil structural units were extracted from the original multimodal stress data matrix. The stress evolution rate was calculated to characterize the local stress change trend. At the same time, acoustic emission signals were extracted to calculate acoustic emission energy flux to characterize crack propagation energy release. The Pearson correlation coefficient was used to determine the correlation between the two, and the correlation threshold was set to 0.85. When the correlation coefficient between stress evolution rate and acoustic emission energy flux is greater than 0.85, the corresponding point is determined to have entered the crack initiation stage and is marked as a crack evolution warning area. A stress-energy coupling index was defined to describe the stress sensitivity under unit energy release. The normal fluctuation range of the index was set to 0.8 to 1.2. When it exceeds this range, nonlinear stress release characteristics are determined to exist.

[0168] Subsequently, a dynamic threshold neural network model was constructed that integrates stress evolution rate, acoustic emission energy flux, ambient temperature, relative humidity, groundwater level, and pore water pressure in the soil. This model adopts a hybrid structure of long short-term memory neural network and gated recurrent unit, with 32 nodes, a sliding window length of 3600 seconds, and a step size of 60 seconds. The model outputs the dynamic stress threshold that varies over time and sets early warning criteria. When the measured stress value exceeds 1.05 times the dynamic stress threshold, it enters the early warning preparation state; when it exceeds 1.15 times, it triggers the over-limit early warning state. An environmental correction mechanism is configured to normalize the ambient temperature, humidity, groundwater level, and pore water pressure to ensure input consistency. Continuity constraints are set to limit the threshold change rate between adjacent time points to no more than 10%. If the condition is violated three times consecutively, the system triggers the model steady-state self-check mechanism and suspends weight updates.

[0169] Then, based on the hierarchical topological characteristics of the support structure, the stress time series of each level is divided into four spatial scales: steel structure unit level, key node level, component combination level, and overall structure level. A sliding window mechanism is used to generate a spatiotemporal pyramid input tensor with a length of 3600 seconds and a step size of 120 seconds. A multi-scale structural damage evolution discrimination model is configured, and the evolutionary features between levels are extracted through the spatiotemporal pyramid neural network to output the structural damage evolution index. The risk identification threshold is set to 0.75. When the index exceeds this threshold, it is marked as a potential risk area. An edge computing node and a cloud intelligent center are deployed to form a distributed collaborative computing architecture. The edge node cache period is 10 seconds. Dynamic judgment is performed on real-time data and event packets are generated. When the judgment results of three consecutive sampling periods meet the over-limit warning conditions, the cloud risk escalation response mechanism is triggered.

[0170] Finally, a digital twin model of the support structure is constructed. Using the original multimodal stress data matrix as input, real-time mapping of the structural state is performed with a refresh cycle of 10 seconds. The model outputs the dynamic trajectories of stress and deformation states at monitoring points. A cognitive graph structure is established based on historical operational data and design parameters, containing five types of semantic nodes: component type, material properties, stress threshold, damage evolution path, and risk level. A risk evolution path network is formed through semantic edges. A knowledge reasoning engine is configured to calculate the similarity between the current state path and historical semantic paths. When the similarity is higher than 0.90, risk level semantic nodes and response strategy semantic nodes are output. Based on the reasoning results, risk triples are generated, including risk source nodes, propagation path nodes, and response measure nodes, forming a structured early warning report. The report includes a unique event number, trigger timestamp, similarity score, and recommended operation instructions for risk level classification and response scheduling.

[0171] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of stress exceeding limits in support structures, characterized in that: Includes the following steps: Step S100: Configure the multimodal stress sensing array and perform time synchronization, generate the original multimodal stress data matrix, configure the multimodal sensing response mapping model, and perform channel correction. Step S200: Extract the original multimodal stress data matrix, calculate the stress evolution rate and acoustic emission energy flux, determine the correlation, define the stress-energy coupling index, and set the sensitivity judgment range. Step S300: Configure a dynamic threshold neural network model that integrates multiple input sources, outputs a time-varying dynamic stress threshold, sets an early warning criterion, and executes dynamic environmental factor correction and model stability constraint mechanisms. Step S400: Configure the multi-scale stress input sequence, generate the spatiotemporal pyramid input tensor, the multi-scale structural damage evolution discrimination model, and output the structural damage evolution index. Set the structural damage evolution index and growth rate criteria. Step S500: Configure edge computing nodes and cloud intelligent center, establish a multi-level collaborative computing architecture, configure real-time early warning triggering mechanism and event package generation mechanism, configure edge consistency judgment mechanism and cloud risk escalation response mechanism; Step S600: Configure the digital twin model of the support structure, perform real-time state mapping, configure the cognitive graph structure and knowledge reasoning engine, configure the risk triplet output mechanism, and generate a structured early warning report.

2. The method for early warning of stress exceeding limits in support structures according to claim 1, characterized in that: Step S100 includes the following sub-steps: Step S100.1: Configure the multimodal stress sensing array and perform time synchronization to generate the original multimodal stress data matrix; Specifically, fiber optic grating sensing units, miniature piezoelectric accelerometer sensing units, and acoustic emission array sensing units are selected from the sensor spare parts library. Then, using installation fixtures and clamping components, the fiber optic grating sensing units, miniature piezoelectric accelerometer sensing units, and acoustic emission array sensing units are fixed at typical stress positions of reinforced concrete structural units, concrete structural units, and soil contact structural units to form a support structure stress sensing array. Fiber grating sensing units are used to acquire the micro-strain response of steel reinforcement and concrete structural units, micro piezoelectric accelerometers are used to acquire the local dynamic acceleration response caused by load changes, and acoustic emission array sensing units are used to acquire the high-frequency acoustic signals generated by microcrack initiation and propagation. The point zoning is set according to the structural risk level. An embedded time synchronization module is assembled from an embedded control board, and the signal acquisition channel of the support structure stress sensing array is connected to the acquisition port of the embedded time synchronization module to obtain the raw acquisition signals generated by the fiber optic grating sensing unit, the miniature piezoelectric acceleration sensing unit, and the acoustic emission array sensing unit. The embedded time synchronization module performs time alignment processing on the raw acquisition signal with millisecond-level precision, and writes three-dimensional layout coordinate information and acquisition timestamp for each acquisition, which are aggregated to form a multimodal stress raw data matrix. The multimodal stress raw data matrix has spatial positioning attributes and temporal consistency attributes, and serves as the input data carrier for subsequent stress evolution rate calculation and dynamic early warning model training. Step S100.2: Configure the multimodal sensing response mapping model and perform channel correction; Specifically, the data acquisition program starts the self-calibration operation cycle in the first hour after the system is started. The self-calibration operation cycle includes a static load data acquisition stage and an unloaded data acquisition stage. The static load data acquisition stage uses the structure's self-weight and predetermined working conditions to obtain the static load response curve, while the unloaded data acquisition stage obtains the baseline response curve under the condition of no external disturbance. After obtaining the static load response curve and the baseline response curve, the difference between the two types of curves is calculated and fitted to establish a multimodal sensing response mapping model covering the fiber optic grating sensing unit, the micro piezoelectric accelerometer sensing unit, and the acoustic emission array sensing unit. Based on the multimodal sensing response mapping model, channel gain parameter correction and zero bias parameter correction are performed.

3. The method for early warning of stress exceeding limits in support structures according to claim 1, characterized in that: Step S200 includes the following sub-steps: Step S200.1: Extract the original multimodal stress data matrix, calculate the stress evolution rate and acoustic emission energy flux, and determine the correlation. Specifically, stress time series data of reinforced concrete structural units, concrete structural units, and soil contact structural units are extracted from the multimodal stress raw data matrix, and time intervals are set. For any time stress value Compared to the previous moment stress value Differential calculations were performed to obtain the stress evolution rate. The specific calculation formula is as follows: ； In the formula: For stress evolution rate, For a moment The stress value, For a moment The stress value, The time interval between adjacent sampling; Extracting the acoustic emission signal waveform from the acoustic emission array sensing unit Set the energy calculation window length Calculate time points using integral operations acoustic emission energy flux The specific calculation formula is as follows: ； In the formula: For time points Acoustic emission energy flux, For energy calculation window length, For acoustic emission array sensing units in time The instantaneous amplitude signal at a given moment; Constructing time synchronization sequences The correlation between the two was calculated using the Pearson correlation coefficient. The specific formula for calculating the Pearson correlation coefficient is as follows: ； In the formula: The Pearson correlation coefficient is used. The covariance between stress evolution rate and acoustic emission energy flux. Let V be the variance of the stress evolution rate. Let V be the variance of acoustic emission energy flux; The correlation threshold is set to 0.

85. When r>0.85, the corresponding spatial point is determined to have entered the microcrack initiation stage and is marked as a crack evolution early warning area. Step S200.2: Define the stress-energy coupling index and set the sensitivity judgment range; Specifically, a stress-energy coupling index is defined for spatial locations where the correlation determination results meet the conditions. The stress response sensitivity, used to characterize the stress response sensitivity per unit energy release, is calculated using the following formula: ； In the formula: As a stress-energy coupling index, For stress evolution rate, For acoustic emission energy flux, the normal fluctuation range of the stress-energy coupling index is set as follows: ,when When the corresponding spatial location is determined to have nonlinear stress release characteristics, the determination result is input into the dynamic threshold self-learning and environmental correction mechanism for subsequent adjustment of risk weight and response level.

4. The method for early warning of stress exceeding limits in support structures according to claim 1, characterized in that: Step S300 includes the following sub-steps: Step S300.1: Configure a dynamic threshold neural network model that integrates multiple input sources, outputs a time-varying dynamic stress threshold, and sets an early warning criterion; Specifically, configure the input feature sequence, including stress evolution rate. Acoustic emission energy flux Ambient temperature relative humidity Groundwater level With soil pore water pressure A dynamic threshold neural network model with a hybrid structure of long short-term memory neural network and gated recurrent unit is constructed to complete the internal state reset of the gated structure; The input features undergo max-min normalization before entering the dynamic threshold neural network model, with the normalization interval set to [value missing]. This is used to unify the feature scale across dimensions and improve the model's convergence stability. The time points are output through the trained dynamic threshold neural network model. Corresponding dynamic stress threshold The output function expression is: ； In the formula: It is a point in time. The dynamic stress threshold, It is a point in time. The rate of stress evolution, It is a point in time. Acoustic emission energy flux, It is a point in time. Ambient temperature correction factor, It is a point in time. The soil pore water pressure weighting parameter; The warning criterion for setting the dynamic stress threshold is: when the measured stress value satisfy: When the following conditions are met, the system enters an early warning preparation state: When the stress exceeds the limit, the system triggers a stress over-limit warning state to avoid repeated state switching caused by threshold fluctuations. A hysteresis recovery mechanism is set: when the stress value drops to meet the threshold... When this happens, the system status changes from early warning preparation or over-limit early warning to normal status; Step S300.2: Execute dynamic environmental factor correction and dynamic threshold neural network model stability constraint mechanism; Specifically, configure an environmental data normalization processing module to process environmental temperature. air humidity Groundwater level and soil pore water pressure Perform a maximum-minimum normalization operation to normalize all environmental factor input values ​​to the minimum value. The interval is the dynamic stress threshold output by the dynamic threshold neural network model. Set a continuity constraint to limit the threshold change rate between adjacent time points to no more than 10%, with the following judgment condition: ； In the formula: It is the current time point. The dynamic stress threshold, It is the dynamic stress threshold at the previous time point. It is the sampling interval between adjacent time points, and 0.10 is the maximum allowable rate of change threshold; When the output of the dynamic threshold neural network model fails to meet the above conditions three times consecutively, the system pauses the weight update of the dynamic threshold neural network model and triggers the dynamic threshold neural network model steady-state self-checking mechanism, marking the current state of the dynamic threshold neural network model as requiring external intervention.

5. The method for early warning of stress exceeding limits in support structures according to claim 1, characterized in that: Step S400 includes the following sub-steps: Step S400.1: Configure the multi-scale stress input sequence, generate the spatiotemporal pyramid input tensor, the multi-scale structural damage evolution discrimination model, and output the structural damage evolution index. Specifically, by using fiber optic grating sensing units, micro piezoelectric accelerometer sensing units, and acoustic emission array sensing units deployed in the support structure, a multimodal stress raw data matrix is ​​collected. Based on the hierarchical topological characteristics of the support structure, the measurement points of the fiber optic grating sensing units, micro piezoelectric accelerometer sensing units, and acoustic emission array sensing units are reconstructed to form stress time series at the steel structure unit level, key node level, component assembly level, and overall structure level, respectively, forming a set of stress sequences at four spatial scales with hierarchical topological relationships. A sliding window sampling mechanism is used to uniformly sample the stress sequence set at four spatial scales. The stress time series data at each level are aligned and integrated according to the time order and spatial hierarchy to generate a unified spatiotemporal pyramid input tensor. This spatiotemporal pyramid input tensor is expanded into a time series in the time dimension and maintains the four-level hierarchical structure of steel structure unit level, key node level, component combination level and overall structure level in the spatial dimension. Based on the established spatiotemporal pyramid input tensor, a multi-scale structural damage evolution discrimination model is configured. This multi-scale structural damage evolution discrimination model adopts a spatiotemporal pyramid neural network structure with spatiotemporal feature fusion capability. The spatiotemporal pyramid neural network consists of four scale input channels, corresponding to the steel structure unit level, key node level, component combination level, and overall structure level, respectively. Each scale input channel contains a three-layer one-dimensional convolutional neural network, which is used to extract the local time series evolution features of the corresponding level. Cross-level feature association and weighted fusion are realized through a unified fusion attention module. Finally, the structural damage evolution index is output through a fully connected output layer. Structural damage evolution index Used to characterize time points The overall damage state at each spatial level is defined as follows: ,in, This indicates that the support structure is in an undamaged state. This indicates that the support structure has reached a critical failure state; Step S400.2: Set the structural damage evolution index and growth rate criterion; Specifically, the structural damage evolution index at each spatial level Time series differencing is performed to calculate the structural damage evolution growth rate, which is defined as: ； In the formula: For time points The rate of evolution and growth of structural damage. The structural damage evolution index at the current time point. This represents the structural damage evolution index at the previous time point. The sampling interval; The risk identification threshold for the structural damage evolution index is set as follows: When the structural damage evolution index at any spatial level When the above risk identification threshold is met, the system automatically determines that the spatial level is a potential over-limit risk area and writes the unique structural level identifier and its structural damage evolution time series corresponding to the level into the early warning index queue.

6. The method for early warning of stress exceeding limits in support structures according to claim 1, characterized in that: Step S500 includes the following sub-steps: Step S500.1: Configure edge computing nodes and cloud intelligent center, establish a multi-level collaborative computing architecture, and configure real-time early warning triggering mechanism and event package generation mechanism; Specifically, a distributed computing architecture consisting of edge computing nodes and a cloud-based intelligent center is configured. The edge computing nodes are used to perform real-time calculation tasks of the multimodal stress raw data matrix, while the cloud-based intelligent center is used to perform global trend analysis and model iteration training tasks. The edge computing nodes integrate embedded edge processing chips and 5G communication modules and are deployed in the near-field position of the stress sensing array of the support structure. The cloud-based intelligent center centrally manages and integrates the stress evolution rate, acoustic emission energy flux, structural damage evolution index, and dynamic stress threshold uploaded by multiple edge computing nodes through a unified data platform. The cache period of the edge computing nodes is set, and a sliding cache mechanism is used to segment and encapsulate the continuously acquired multimodal stress raw data matrix. Set the stress evolution rate of edge computing nodes Acoustic emission energy flux With dynamic stress threshold Real-time monitoring of the time series is performed, and the measured stress value satisfies: If the event is detected, it is determined to be an over-limit warning state, and the event package generation and reporting process is executed within 1 second. The event package includes three-dimensional layout coordinates, detection timestamp, stress evolution rate sequence, acoustic emission energy flux sequence, structural damage evolution index sequence, and dynamic stress threshold estimation sequence. The event packets are encapsulated using a standardized data structure and transmitted encrypted via a 5G private network to the cloud-based intelligent center for subsequent fusion analysis and instruction distribution. Step S500.2: Configure the edge consistency judgment mechanism and the cloud risk escalation response mechanism; Specifically, an edge consistency judgment mechanism is configured to determine the consistency of multiple consecutive early warning results from edge computing nodes. The judgment criterion is that the stress judgment results within three consecutive sampling periods all meet the over-limit early warning conditions. When the consistency conditions are met, the edge computing node outputs a level-one risk confirmation signal. The cloud intelligent center is configured with an event fusion analysis module to perform fusion analysis on the event packets uploaded by multiple edge computing nodes and the level-one risk confirmation signal. Based on the fusion results, the cloud intelligent center performs structural damage evolution trend fitting and stress accumulation effect assessment, generates enhanced response instructions, and triggers a level-one risk alarm.

7. The method for early warning of stress exceeding limits in support structures according to claim 1, characterized in that: Step S600 includes the following sub-steps: Step S600.1: Configure the digital twin model of the support structure and perform real-time state mapping; configure the cognitive graph structure and knowledge reasoning engine. Specifically, a digital twin model of the support structure is configured. The digital twin model of the support structure takes the original multimodal stress data matrix collected by fiber optic grating sensing units, micro piezoelectric accelerometer sensing units and acoustic emission array sensing units as input to establish a digital equivalent structural model of the real-time stress state and deformation state of the support structure monitoring points. The digital twin model of the support structure uses a structural topology mapping algorithm to topologically bind the location of each monitoring point to the type of support structure component, generating a set of digitally mapped entities with topological consistency and response synchronization. The digital twin model of the support structure updates the set of state parameters with a fixed refresh cycle. The refresh cycle is set, and the stress state and deformation state of the monitoring points are continuously updated through a real-time refresh mechanism. The state change trajectory obtained from each refresh is written into the dynamic state buffer and used as the input feature set of the cognitive graph knowledge reasoning engine. Configure a cognitive graph structure. The cognitive graph structure takes the historical operation data of the support structure and the structural design parameters as input to establish a semantic association network for risk identification and trend prediction. The cognitive graph structure contains a set of semantic nodes: component type semantic nodes, material property semantic nodes, stress threshold semantic nodes, damage evolution path semantic nodes, and risk level semantic nodes. Each semantic node is connected through structural association semantic edges, response evolution semantic edges, and risk propagation semantic edges to form a directed graph structure with multi-layered semantic relationships, which is used to represent the stress evolution logic and risk evolution path of the support structure during long-term operation. Configure a knowledge reasoning engine module. The knowledge reasoning engine module calculates the similarity between the set of state parameters from the digital twin model of the support structure and the historical semantic path in the cognitive graph structure. It uses a vectorized graph matching algorithm to obtain the similarity "Sim" between the current state path and the historical semantic path, and performs a judgment: "Sim>0.90". When the similarity value meets the judgment condition, the knowledge reasoning engine module determines the risk level semantic node and the response strategy semantic node corresponding to the high similarity path, and generates a risk path backtracking result and response measure recommendation set. Step S600.2: Configure the risk triplet output mechanism and generate a structured early warning report; Specifically, a risk triplet output mechanism is configured. The risk triplet output mechanism takes the risk path backtracking results and the set of recommended countermeasures output by the knowledge reasoning engine module as input, and generates a risk triplet structure consisting of three elements: risk source semantic node, propagation path semantic node, and countermeasure semantic node. The risk source semantic node corresponds to the unique identifier of the support structure component or monitoring node; the propagation path semantic node corresponds to the set of intermediate states in the risk propagation path; and the response measure semantic node corresponds to the valid response operation stored in the historical knowledge graph. The risk triplet structure is encapsulated in a standardized data format to form a structured early warning report. The structured early warning report includes a unique event number, trigger timestamp, similarity score, and a set of recommended operation instructions. The structured early warning report is transmitted through the early warning notification module and written to the log database for risk event tracing and decision optimization. The structured early warning report generated by the cognitive graph decision engine serves as the core data basis for classifying the early warning level of support structure stress over-limit, scheduling response instructions, and human-computer interaction feedback, constituting the final intelligent decision output unit of the support structure stress over-limit early warning method.

Images