Flood prevention and drainage structure fracture failure prediction method and system based on data fusion

Abstract

The invention provides a flood prevention and drainage structure fracture failure prediction method and system based on data fusion, and the method comprises the steps: carrying out the multi-scale feature extraction of a standardized fusion data set, carrying out the dimension reduction and fusion of all extracted features, and obtaining a low-dimensional fusion feature vector; based on the low-dimensional fusion feature vector, a fracture failure prediction model is adopted to perform fracture failure prediction on the flood prevention and drainage structure, and a dynamic stress intensity factor prediction value and a fracture toughness time-varying prediction value of a preset key part of the flood prevention and drainage structure are obtained; and taking the ratio of the dynamic stress intensity factor prediction value to the fracture toughness time-varying prediction value as a fracture failure risk index, when the fracture failure risk index exceeds a preset threshold value, generating graded early warning information, and based on the change trend of the fracture failure risk index, positioning a potential fracture failure part and predicting the residual safety life. The intelligent level and reliability of safety management of the flood prevention and drainage structure are greatly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to a flood control structure fracture failure prediction method and system based on data fusion. BACKGROUND

[0002] Flood control structures (such as dams, dikes, spillways, etc.) are major infrastructure related to national economy and people's livelihood, and their structural safety is crucial. Fracture failure is a major failure mode of such concrete or masonry structures under the combined action of long-term hydrological load, environmental erosion and material aging. Once it occurs, the consequences are disastrous. Therefore, accurate fracture failure prediction and early warning of flood control structures is a core challenge in the field of water conservancy engineering safety operation.

[0003] Existing methods rely on a single monitoring method (such as using only strain gauges or only vibration sensors), and fail to fully utilize the complementarity and correlation between multi-source heterogeneous monitoring data (such as stress and strain, vibration, acoustic emission, environment, and hydrological data). This results in incomplete perception of the structure's state and inability to capture multi-dimensional precursor information for fracture initiation and propagation. Traditional feature extraction methods (such as simple time-domain statistics and FFT spectral analysis) are often purely data-driven, and the extracted features lack clear physical connection with key parameters of fracture mechanics (such as stress intensity factor and fracture toughness). This makes the model difficult to interpret and the prediction results lack clear physical meaning. Most artificial intelligence-based prediction models attempt to directly establish a mapping from data to "safe / failure" labels, which are end-to-end black box models. They often ignore the fact that fracture toughness is a time-varying quantity that changes with environmental erosion and fatigue load, rather than a material constant. Treating fracture toughness as a constant for failure judgment (i.e., fracture toughness) can significantly affect the accuracy of long-term prediction. Existing early warning systems usually trigger an alarm only when the monitoring value exceeds a fixed threshold, and cannot achieve trend-based graded early warning. More importantly, they are difficult to accurately locate the most dangerous potential fracture failure site, and cannot provide physically meaningful remaining safe life prediction, making it difficult to provide forward-looking and quantitative scientific basis for operation and maintenance decisions. SUMMARY

[0004] To solve the problems existing in the prior art, the application provides a flood control structure fracture failure prediction method and system based on data fusion.

[0005] To achieve the above-mentioned purpose, the application provides the following solutions:

[0006] A flood control structure fracture failure prediction method based on data fusion, comprising:

[0007] Collecting and preprocessing multi-source heterogeneous monitoring data reflecting the structural state of the flood control structure to obtain a standardized fusion data set;

[0008] performing multi-scale feature extraction on the standardized fusion data set, and performing dimension reduction and fusion on all extracted features to obtain a low-dimensional fusion feature vector;

[0009] based on the low-dimensional fusion feature vector, performing fracture failure prediction of the flood control structure by using a fracture failure prediction model to obtain a dynamic stress intensity factor prediction value and a fracture toughness time-varying prediction value of a preset key part of the flood control structure;

[0010] the ratio of the dynamic stress intensity factor prediction value to the fracture toughness time-varying prediction value is taken as a fracture failure risk index, when the fracture failure risk index exceeds a preset threshold, a hierarchical warning information is generated, and based on the change trend of the fracture failure risk index, a potential fracture failure part is located and a remaining safe life is predicted.

[0011] Preferably, the multi-source heterogeneous monitoring data includes structural stress and strain data, vibration and acoustic emission data, environmental erosion data, and hydrological load data.

[0012] Preferably, the method for preprocessing the multi-source heterogeneous monitoring data includes:

[0013] performing denoising, missing data filling and standardization on the multi-source heterogeneous monitoring data to obtain standardized multi-source heterogeneous monitoring data;

[0014] unifying the time stamps of the standardized multi-source heterogeneous monitoring data by using a time interpolation method, and based on the causal relationship between the physical changes of the flood control structure and the multi-source heterogeneous monitoring data, obtaining the time delay relationship between each data;

[0015] aligning the standardized multi-source heterogeneous monitoring data according to the unified time reference and the delay relationship to obtain time-aligned data;

[0016] establishing a three-dimensional space model of the flood control structure, mapping the coordinates of each sensor used to collect multi-source heterogeneous monitoring data to the three-dimensional space model to obtain the accurate positions of each sensor in the three-dimensional space model;

[0017] based on the accurate positions, constructing a sensor spatial distance matrix and a topological relationship network to obtain spatial correlation attributes;

[0018] assigning the spatial correlation data to each data point in the time-aligned data to obtain a standardized fusion data set.

[0019] Preferably, the method for obtaining a low-dimensional fusion feature vector includes:

[0020] Based on a one-dimensional convolutional neural network, time domain feature extraction is performed on the standardized fusion data set, and high-order statistics in each data window are calculated synchronously, so as to obtain local time sequence characteristics and physical statistical characteristic quantities;

[0021] The local time sequence characteristics and the physical statistical characteristic quantities are spliced in the feature dimension to obtain an enhanced time domain feature vector;

[0022] Fast Fourier transform is performed on the standardized fusion data set to obtain a power spectrum density, and based on an introduced learnable spectrum attention mechanism, weighted dominant frequency band energy features and dominant frequency amplitude features are obtained;

[0023] Based on the weighted dominant frequency band energy features and the dominant frequency amplitude features, a frequency domain feature vector is obtained;

[0024] Wavelet transform based on adaptive optimal wavelet basis selection is adopted to decompose the standardized fusion data set to obtain sub-bands;

[0025] The energy and multi-scale sample entropy of each sub-band are calculated to obtain a time-frequency feature vector;

[0026] The time domain feature vector, the frequency domain feature vector and the time-frequency feature vector are fused and dimensionally reduced to obtain a low-dimensional fusion feature vector.

[0027] Preferably, the fracture failure prediction model comprises:

[0028] A shared feature encoding layer is used to abstract and transform the low-dimensional fusion feature vector by using a fully connected neural network to obtain a shared feature representation;

[0029] A dynamic stress intensity factor prediction branch is used to obtain a dynamic stress intensity factor prediction value based on the shared feature representation;

[0030] A time-varying fracture toughness prediction branch is used to obtain a time-varying fracture toughness prediction value based on the shared feature representation.

[0031] Preferably, the loss function of the fracture failure prediction model is a total loss function composed of a supervised loss, a fracture criterion constraint, an energy conservation constraint and a trend consistency constraint.

[0032] The application also provides a data fusion-based flood control structure fracture failure prediction system for implementing the method, comprising:

[0033] A data processing module is used to collect and pre-process multi-source heterogeneous monitoring data reflecting the structure state of the flood control structure to obtain a standardized fusion data set;

[0034] a data feature extraction module, configured to perform multi-scale feature extraction on the standardized fusion data set, and to perform dimension reduction and fusion on all extracted features to obtain a low-dimensional fusion feature vector;

[0035] a prediction module, configured to perform fracture failure prediction of the flood control structure based on the low-dimensional fusion feature vector, by using a fracture failure prediction model, to obtain a dynamic stress intensity factor prediction value and a fracture toughness time-varying prediction value of a preset key position of the flood control structure;

[0036] a failure analysis module, configured to take a ratio of the dynamic stress intensity factor prediction value to the fracture toughness time-varying prediction value as a fracture failure risk index, to generate a graded early warning information when the fracture failure risk index exceeds a preset threshold, and to locate a potential fracture failure position and predict a remaining safe life based on a change trend of the fracture failure risk index.

[0037] Preferably, the data processing module comprises:

[0038] a standardization unit, configured to perform denoising, missing data filling and standardization on the multi-source heterogeneous monitoring data to obtain standardized multi-source heterogeneous monitoring data;

[0039] a delay relationship acquisition unit, configured to unify time stamps of the standardized multi-source heterogeneous monitoring data by using a time interpolation method, and to obtain time delay relationships between data based on a causal relationship between physical changes of the flood control structure and the multi-source heterogeneous monitoring data;

[0040] a time alignment unit, configured to align the standardized multi-source heterogeneous monitoring data according to a unified time reference and delay relationship to obtain time-aligned data;

[0041] a space mapping unit, configured to establish a three-dimensional space model of the flood control structure, and to map position coordinates of each sensor used to collect the multi-source heterogeneous monitoring data to the three-dimensional space model to obtain accurate positions of each sensor in the three-dimensional space model;

[0042] a space correlation unit, configured to construct a sensor space distance matrix and a topological relationship network based on the accurate positions to obtain a space correlation attribute;

[0043] a data set construction unit, configured to assign the space correlation data to each data point in the time-aligned data to obtain a standardized fusion data set.

[0044] Compared with the prior art, the beneficial effects of the present application are: through deep fusion and high-level feature extraction of multi-source heterogeneous data in the time-space dimension, a low-dimensional fusion feature vector is constructed, which greatly improves the depth and accuracy of information utilization and state perception, and can capture the precursor signals of fracture failure earlier and more sensitively. A dual-branch decoding network with physical knowledge embedding is created, and a physical constraint loss function is used for collaborative training. This makes the prediction process deeply integrate the fracture mechanics mechanism, and the prediction results (dynamic stress intensity factor, fracture toughness) are themselves the core criteria in fracture mechanics, the model has strong explainability, and the prediction results are reliable and credible. The present application provides a complete and closed-loop technical solution from data preprocessing, feature fusion, physical mechanism model prediction to final intelligent decision-making, systematically solves multiple key problems in fracture failure prediction of flood control structures, and has good engineering application prospect and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0046] Figure 1 The flow chart of the fracture failure prediction method of the flood control structure based on data fusion according to the embodiment of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail in combination with the drawings and specific embodiments.

[0049] Embodiment one

[0050] As shown in the figure, a fracture failure prediction method of a flood control structure based on data fusion comprises: Figure 1

[0051] S1: Collect and pre-process multi-source heterogeneous monitoring data reflecting the structure state of the flood control structure to obtain a standardized fusion data set; further embodiments are that the multi-source heterogeneous monitoring data comprises structure stress and strain data, vibration and acoustic emission data, environmental erosion data and hydrological load data.​

[0052] Further embodiments are directed to a method for preprocessing multi-source heterogeneous monitoring data, comprising:

[0053] The multi-source heterogeneous monitoring data is denoised, missing data is filled, and standardized to obtain standardized multi-source heterogeneous monitoring data; the present application introduces an intelligent preprocessing paradigm based on deep fusion of physical mechanism and data driving in the data preprocessing stage. The whole process starts with deep cleaning and regularization of the original multi-source heterogeneous monitoring data, adopts an adaptive filtering algorithm, dynamically adjusts the filtering parameters according to the real-time analysis of signal spectrum characteristics and noise distribution, and adopts different denoising strategies for the burst noise in acoustic emission signals and the slow drift in strain data, so as to realize accurate noise reduction rather than indiscriminate denoising. For missing data, instead of simply using linear interpolation, a collaborative filling mechanism based on time series prediction (such as LSTM network) and spatial correlation (using adjacent sensor data) is combined to ensure the continuity of data in time and space dimensions.

[0054] The time stamps of the standardized multi-source heterogeneous monitoring data are unified by using a time interpolation method, and the time delay relationship between the data is obtained based on the causal relationship between the physical changes of the flood control structure and the multi-source heterogeneous monitoring data; specifically, on the basis of unifying the time stamps, the present application is characterized by how to scientifically determine the time delay relationship between different data sources. The method of relying on experience estimation is abandoned, and a double-checking mechanism based on mutual information and phase analysis is proposed. First, the high-frequency vibration or acoustic emission data is used to accurately capture the millisecond delay caused by mechanical wave transmission by calculating the cross-correlation function peak value between the data and the load change data; for the delay between slow-changing processes such as environmental erosion and structural response, the phase difference of the trend inflection point is analyzed and the theoretical conduction time obtained by finite element simulation is combined for comprehensive judgment, thereby establishing a set of quantitative and reliable "cause-effect-delay" dictionary, laying a solid foundation for subsequent accurate alignment.

[0055] The standardized multi-source heterogeneous monitoring data is aligned according to the unified time reference and delay relationship to obtain time-aligned data;

[0056] A three-dimensional space model of the flood control and drainage structure is established, and the coordinates of positions of various sensors for collecting multi-source heterogeneous monitoring data are mapped onto the three-dimensional space model to obtain accurate positions of the various sensors in the three-dimensional space model; specifically, after the delay relationship calibration is completed, the time alignment process is not simply data translation, but a global time axis is constructed with a physical event as a reference. The data stream of each sensor is finely time-shifted and corrected according to the corresponding delay amount, to ensure that the same physical event (such as a water level rise and fall) can be synchronously reflected in all monitoring data. Thereafter, the processing flow is expanded from the time domain to the spatial domain. First, a three-dimensional digital twin model of the flood control and drainage structure is accurately reconstructed, and not simply the coordinates of the sensors are mapped, but through laser scanning and BIM model fusion technology, the actual installation positions of the sensors and key mechanical feature points (such as stress concentration areas and structural joints) of the structure are accurately bound, to give each sensor an accurate spatial attribute and mechanical significance on the component where the sensor is located.

[0057] Based on the accurate positions, a sensor spatial distance matrix and a topological relationship network are constructed to obtain spatial correlation attributes;

[0058] The spatial correlation data is given to each data point in the time-aligned data to obtain a standardized fusion data set. The spatial topological relationship network of the sensors not only includes a geometric distance matrix between the sensors, but also integrates a “functional distance” based on a structure mechanical transmission path, for example, sensors on the same stress path are defined as “functionally adjacent”, even if the geometric distance is far. This spatial correlation attribute is creatively given to each time-aligned data point as metadata. Finally, each data point becomes a high-dimensional data unit containing a timestamp, a monitoring value, a sensor spatial coordinate, a component ID to which the data point belongs, and information of all adjacent sensor sets related to the function of the data point, thereby generating a standardized fusion data set that is highly fused and correlated in time, space and physical dimensions, to provide an unprecedented consistency and richness basis for subsequent deep feature mining.

[0059] S2: multi-scale feature extraction is performed on the standardized fusion data set, and all extracted features are dimensionally reduced and fused to obtain a low-dimensional fusion feature vector; a further implementation manner is that the method for obtaining the low-dimensional fusion feature vector comprises:

[0060] The one-dimensional convolutional neural network is used for extracting time domain features of the standardized fusion data set, and high-order statistics in each data window are calculated synchronously, so as to obtain local time sequence features and physical statistical characteristic quantities. The local time sequence features and the physical statistical characteristic quantities are spliced in the feature dimension to obtain an enhanced time domain feature vector. Specifically, the extraction of the time domain features is not dependent on a single method, but is realized through a parallel hybrid architecture. The one-dimensional convolutional neural network takes original time sequence data as input, and a small convolution kernel is used in a shallow structure to focus on capturing local mutations, transient impacts and microscopic trend patterns contained in the signal, which are often early signs of structural damage. At the same time, a series of high-order statistics are calculated synchronously in the same sliding data window, including not only the mean and variance, but also the arrangement entropy and HURST index, which can quantify the complexity and randomness of the sequence, so as to form a set of statistical characteristics with clear physical meaning. Finally, the abstract local features learned by the CNN and the physical statistical features designed manually are spliced in the dimension to form an enhanced time domain feature vector with complementary information, which contains details and summarizes the global information.

[0061] The standardized fusion data set is subjected to fast Fourier transform to obtain a power spectral density, and based on the introduced learnable spectrum attention mechanism, weighted dominant band energy features and dominant frequency amplitude features are obtained. The frequency domain feature vector is obtained based on the weighted dominant band energy features and the dominant frequency amplitude features. Specifically, in the frequency domain processing, the present application goes beyond simple spectrum calculation. After obtaining the initial power spectral density through fast Fourier transform, a trainable spectrum attention mechanism is introduced. The mechanism is essentially a lightweight neural network that can autonomously learn and analyze the entire spectrum, automatically identify the frequency bands highly related to key physical phenomena such as structural inherent characteristics and crack propagation acoustic emission characteristic frequencies, and assign higher weights to these frequency bands while suppressing the frequency band contributions caused by environmental noise or irrelevant vibrations. This data-driven "spectrum focusing" process makes the extracted weighted dominant band energy features and dominant frequency amplitude features more discriminative, and directly constitutes the frequency domain feature vector with clear physical meaning.

[0062] The standard fusion data set is decomposed by using wavelet transform based on adaptive optimal wavelet basis selection to obtain sub-bands; the energy and multi-scale sample entropy of each sub-band are calculated to obtain a time-frequency feature vector; for the time-frequency feature, adaptive wavelet packet transform is used to cope with the non-stationary characteristics of the signal. The selection of the wavelet basis is not determined empirically, but the optimal basis function that can best concentrate the energy of the signal is automatically selected by calculating the energy entropy of the signal after decomposition of different candidate wavelet bases, so that the best matching between the analysis tool and the signal characteristics is realized. On this basis, not only the energy of each sub-band coefficient obtained by decomposition is calculated, but also the multi-scale sample entropy of each sub-band is further innovatively calculated. The sample entropy can finely depict the complexity and irregularity of the signal at a certain scale, and is extremely sensitive to the subtle changes in the signal caused by structural micro-damage. By fusing the energy and entropy values of each sub-band, a time-frequency feature vector that can represent the distribution of signal energy and complexity in the time-frequency plane is obtained.

[0063] The time-domain feature vector, the frequency-domain feature vector, and the time-frequency feature vector are fused and dimensionally reduced to obtain a low-dimensional fusion feature vector. Specifically, the high-dimensional feature vectors extracted from the time, frequency, and time-frequency domains are spliced to form a super-dimensional mixed feature set. To this end, the embodiment adopts a supervised manifold learning dimension reduction algorithm. When constructing the sample neighborhood relationship graph, it not only considers the Euclidean distance in the feature space, but also more importantly, integrates the sample fracture failure risk label information, so that samples belonging to the same risk level are closer in the low-dimensional space, and samples of different risk levels are pushed away. This process ensures that the low-dimensional fusion feature vector after dimension reduction retains the most relevant discriminant information for the final prediction target to the greatest extent, eliminates redundancy, forms a high-value feature representation that is both concise and informative, and lays a solid foundation for subsequent accurate prediction.

[0064] S3: Based on the low-dimensional fusion feature vector, a fracture failure prediction model is used to predict the fracture failure of the flood control structure, to obtain the dynamic stress intensity factor prediction value and the time-varying fracture toughness prediction value of the predetermined key part of the flood control structure; a further implementation manner is that the fracture failure prediction model comprises:

[0065] A shared feature encoding layer is used to abstract and transform the low-dimensional fusion feature vector using a fully connected neural network to obtain a shared feature representation; a dynamic stress intensity factor prediction branch is used to obtain a dynamic stress intensity factor prediction value based on the shared feature representation; a time-varying fracture toughness prediction branch is used to obtain a time-varying fracture toughness prediction value based on the shared feature representation.

[0066] The core creativity of the application in constructing the fracture failure prediction model is to completely surpass the traditional "black box" regression model. Through the design of a physical mechanism deeply embedded neural network architecture and a multi-constraint joint optimization strategy, the prediction of key parameters of fracture mechanics is accurate, interpretable and consistent with physical laws. Specifically, the model first abstracts the low-dimensional fusion feature vector input through a shared feature coding layer. These feature vectors already contain rich time, frequency, and time-frequency domain information. The role of the shared layer is to further refine the deep patterns related to the fracture process, forming a highly refined shared feature representation, providing a common knowledge base for the subsequent two specialized branches.

[0067] The dynamic stress intensity factor prediction branch does not directly output a numerical value, but adopts a parameterized physical guidance strategy. This branch takes shared features as input, but its final output layer is coupled with a parameterized physical formula. For example, the network learns the "equivalent calibration coefficient" and "equivalent crack length change" related to the current load condition and structural response. The final prediction value is calculated by these network output parameters through the built-in physical formula consistent with the I-type crack theory (such as the relationship between stress and crack length square root). This ensures that the learning process of the network is strictly constrained within the framework of physical laws, ensuring the physical reasonableness of the prediction value.

[0068] Similarly, the innovation of the time-varying fracture toughness prediction branch lies in explicitly simulating the degradation process of material performance. This branch not only receives shared features, but also injects additional inputs such as environmental erosion (such as humidity, chemical ion concentration) and load history (such as cycle number, average stress). Its internal uses a recurrent neural network (such as LSTM) to capture the time-varying degradation effect of material performance caused by the above factors, and finally outputs a slowly changing fracture toughness prediction value over time, breaking the traditional assumption that time-varying fracture toughness is constant.

[0069] The comprehensive loss function consists of four parts: first, a supervision loss based on a small amount of labeled data (such as the mean square error between predicted and simulated or experimental values); second, a fracture criterion constraint loss, which acts as a powerful physical regularization term, severely punishing absurd situations where the predicted dynamic stress intensity factor is greater than the time-varying fracture toughness prediction at "safe" moments, and conversely rewarding such situations at "dangerous" moments, forcing the model to strictly adhere to the Irwin fracture criterion; third, an energy conservation constraint, ensuring that the predicted crack propagation behavior is consistent with the energy release rate trend reflected in the monitoring data; and fourth, a trend consistency constraint, ensuring that the prediction results from adjacent sensors are spatially smooth and continuous, avoiding abrupt changes that violate mechanical principles. By combining these loss terms in a weighted sum, the model not only learns to fit the data during training but is also forced to learn to obey basic physical laws, thus maintaining the physical credibility and consistency of its predictions even without a large amount of labeled data. Ultimately, the model can simultaneously output the predicted values ​​of the dynamically changing dynamic stress intensity factor and the time-varying fracture toughness, providing a solid and scientific basis for subsequent dynamic assessment of failure risk based on the ratio relationship.

[0070] A further implementation method is that the loss function of the fracture failure prediction model is a total loss function composed of supervision loss, fracture criterion constraint, energy conservation constraint, and trend consistency constraint.

[0071] S4: The ratio of the predicted value of dynamic stress intensity factor to the predicted value of fracture toughness time-varying value is used as the fracture failure risk index. When the fracture failure risk index exceeds the preset threshold, graded early warning information is generated, and potential fracture failure sites are located and the remaining safe life is predicted based on the changing trend of the fracture failure risk index.

[0072] In the final decision-making stage of early warning, location and life prediction based on prediction results, this invention constructs a complete system from global risk field construction to intelligent inverse problem solving, realizing a leap from passive alarm to active prediction and accurate diagnosis.

[0073] Specifically, the system does not view the risk indices of individual sensor points in isolation, but first places them within a three-dimensional spatial model. Using a physically constrained Kriging space interpolation algorithm, the risk indices of discrete points are fused to generate a continuously visualized three-dimensional dynamic fracture risk field covering the entire structure surface. The innovation of this interpolation process lies in the fact that its covariates not only include geometric distance but also incorporate principal stress transfer path information derived from finite element analysis. This ensures that the distribution of high-risk areas conforms to structural mechanics principles, avoiding unreasonable high-risk values ​​in rigid supports or non-load-bearing areas, and making the risk field more accurately reflect the potential failure modes of the structure.

[0074] Based on this dynamic risk field, the early warning mechanism transcends simple threshold judgment. The system establishes multi-level adaptive thresholds, such as Level 1 (observation), Level 2 (alert), and Level 3 (danger). Its innovation lies in the fact that these thresholds are not fixed but dynamically fine-tuned based on the confidence interval of the predicted value and the current hydrological load conditions. For example, during high water levels in the flood season, the system automatically lowers the Level 2 warning threshold to capture risks with higher sensitivity; conversely, if the confidence interval of a predicted value is wide (high uncertainty), the system will also adopt a conservative strategy to lower the threshold, thereby achieving intelligent, situation-adaptive, tiered early warning.

[0075] For locating potential fracture failure sites, this invention employs risk field gradient analysis. The initiation and propagation of potential cracks are not only characterized by a high absolute value of the risk index, but more fundamentally by a dramatic spatial change in the risk field, i.e., a high gradient value. The system performs differentiation on the three-dimensional risk field to calculate its spatial gradient field, VRiskField. The true potential crack tip and propagation front will be precisely located on spatial units that simultaneously satisfy a high RiskField value (e.g., >0.8) and a maximum gradient modulus |VRiskField|. This method can clearly delineate the potential crack path, achieving a leap from "point warning" to "field diagnosis" and "path prediction."

[0076] Finally, in predicting remaining safe life, this invention abandons the simple time series extrapolation method and introduces a physical information neural network for scientific prediction. A PINN model is trained using a dynamic risk field sequence constructed over a past period as training data, taking spatial coordinates and future time as input and the risk index at that point as output. The network's loss function incorporates constraints from physical laws controlling crack propagation. This ensures that the network adheres to the fundamental laws of fracture mechanics when learning and extrapolating the risk field evolution. After training, the entire three-dimensional risk field is iteratively predicted over multiple future time steps until the predicted risk index value at any point consistently exceeds 1.0. The time span from the current moment to this critical moment represents the remaining safe life at that point. The system ultimately outputs a remaining life distribution cloud map of the entire structure, rather than a single value, providing a highly intuitive and scientific basis for maintenance decisions.

[0077] Example 2

[0078] This invention also provides a data fusion-based system for predicting the fracture failure of flood control and drainage structures, and a method for implementing this system, comprising:

[0079] The data processing module is used to collect and preprocess multi-source heterogeneous monitoring data reflecting the structural status of flood control and drainage structures to obtain a standardized fusion dataset;

[0080] The data feature extraction module is used to perform multi-scale feature extraction on the standardized fusion dataset, and to reduce and fuse all extracted features to obtain a low-dimensional fusion feature vector.

[0081] The prediction module is used to predict the fracture failure of flood control and drainage structures based on low-dimensional fused feature vectors and a fracture failure prediction model, and to obtain the predicted values ​​of dynamic stress intensity factors and time-varying fracture toughness of preset key parts of flood control and drainage structures.

[0082] The failure analysis module uses the ratio of the predicted value of dynamic stress intensity factor to the predicted value of fracture toughness as the fracture failure risk index. When the fracture failure risk index exceeds a preset threshold, it generates graded early warning information and locates potential fracture failure sites and predicts the remaining safe life based on the changing trend of the fracture failure risk index.

[0083] A further embodiment of the implementation includes a data processing module comprising:

[0084] Standardized units are used to denoise, fill in missing data, and standardize multi-source heterogeneous monitoring data to obtain standardized multi-source heterogeneous monitoring data.

[0085] The delay relationship acquisition unit is used to standardize the timestamps of multi-source heterogeneous monitoring data using time interpolation methods, and to obtain the time delay relationship between each data based on the causal relationship between the physical changes of flood control and drainage structures and multi-source heterogeneous monitoring data.

[0086] The time alignment unit is used to align standardized multi-source heterogeneous monitoring data according to a unified time reference and delay relationship to obtain time-aligned data;

[0087] The spatial mapping unit is used to establish a three-dimensional spatial model of flood control and drainage structures, and to map the position coordinates of each sensor used to collect multi-source heterogeneous monitoring data onto the three-dimensional spatial model to obtain the accurate position of each sensor in the three-dimensional spatial model.

[0088] The spatial association unit is used to construct a sensor spatial distance matrix and topological relationship network based on accurate location to obtain spatial association attributes;

[0089] The dataset building unit is used to assign spatially correlated data to each data point in time-aligned data to obtain a standardized fused dataset.

[0090] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the fracture failure of flood control and drainage structures based on data fusion, characterized in that, include: Collect and preprocess multi-source heterogeneous monitoring data reflecting the structural status of flood control and drainage structures to obtain a standardized fusion dataset; Multi-scale feature extraction is performed on the standardized fusion dataset, and all extracted features are reduced in dimensionality and fused to obtain a low-dimensional fusion feature vector. Based on the low-dimensional fusion feature vector, a fracture failure prediction model is used to predict the fracture failure of flood control and drainage structures, and to obtain the predicted values ​​of dynamic stress intensity factors and time-varying fracture toughness of preset key parts of flood control and drainage structures. The ratio of the predicted value of the dynamic stress intensity factor to the predicted value of the time-varying fracture toughness is used as the fracture failure risk index. When the fracture failure risk index exceeds a preset threshold, graded early warning information is generated, and potential fracture failure sites are located and the remaining safe life is predicted based on the changing trend of the fracture failure risk index.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous monitoring data includes structural stress and strain data, vibration and acoustic emission data, environmental erosion data, and hydrological load data.

3. The method according to claim 1, characterized in that, The method for preprocessing the multi-source heterogeneous monitoring data includes: The multi-source heterogeneous monitoring data is denoised, missing data is filled in, and standardized to obtain standardized multi-source heterogeneous monitoring data. The timestamps of the standardized multi-source heterogeneous monitoring data are unified by time interpolation method, and the time delay relationship between each data is obtained based on the causal relationship between the physical changes of flood control and drainage structures and the multi-source heterogeneous monitoring data. The standardized multi-source heterogeneous monitoring data are aligned according to a unified time base and delay relationship to obtain time-aligned data; A three-dimensional spatial model of the flood control and drainage structure is established, and the position coordinates of each sensor used to collect multi-source heterogeneous monitoring data are mapped onto the three-dimensional spatial model to obtain the accurate position of each sensor in the three-dimensional spatial model. Based on the accurate location, a sensor spatial distance matrix and topological relationship network are constructed to obtain spatial correlation attributes; The spatial correlation data is assigned to each data point in the time-aligned data to obtain a standardized fused dataset.

4. The method according to claim 3, characterized in that, Methods for obtaining low-dimensional fused feature vectors include: Temporal features are extracted from the standardized fusion dataset based on a one-dimensional convolutional neural network, and higher-order statistics within each data window are calculated simultaneously to obtain local temporal features and physical statistical features. The local temporal features and the physical statistical features are concatenated along the feature dimension to obtain an enhanced temporal feature vector. The standardized fusion dataset is subjected to Fast Fourier Transform to obtain the power spectral density, and the weighted dominant frequency band energy features and dominant frequency amplitude features are obtained based on the introduced learnable spectral attention mechanism. Based on the weighted dominant frequency band energy characteristics and dominant frequency amplitude characteristics, a frequency domain feature vector is obtained; The standard fusion dataset is decomposed using wavelet transform based on adaptive optimal wavelet basis selection to obtain sub-bands; Calculate the energy of each sub-band and the multi-scale sample entropy to obtain the time-frequency feature vector; The time-domain feature vector, the frequency-domain feature vector, and the time-frequency feature vector are fused and their dimensions reduced to obtain a low-dimensional fused feature vector.

5. The method according to claim 1, characterized in that, The fracture failure prediction model includes: A shared feature encoding layer is used to perform feature abstraction and transformation on the low-dimensional fused feature vector using a fully connected neural network to obtain a shared feature representation. The dynamic stress intensity factor prediction branch is used to obtain the predicted value of the dynamic stress intensity factor based on the shared feature representation. The time-varying fracture toughness prediction branch is used to obtain the time-varying predicted value of fracture toughness based on the shared feature representation.

6. The method according to claim 1, characterized in that, The loss function of the fracture failure prediction model is a total loss function composed of supervised loss, fracture criterion constraint, energy conservation constraint, and trend consistency constraint.

7. A data fusion-based system for predicting the fracture failure of flood control and drainage structures, used to implement the method described in any one of claims 1-6, characterized in that, include: The data processing module is used to collect and preprocess multi-source heterogeneous monitoring data reflecting the structural status of flood control and drainage structures to obtain a standardized fusion dataset; The data feature extraction module is used to perform multi-scale feature extraction on the standardized fusion dataset, and to reduce and fuse all extracted features to obtain a low-dimensional fusion feature vector. The prediction module is used to predict the fracture failure of flood control and drainage structures based on the low-dimensional fused feature vector and the fracture failure prediction model, and to obtain the predicted values ​​of dynamic stress intensity factors and time-varying fracture toughness of preset key parts of the flood control and drainage structures. The failure analysis module is used to take the ratio of the predicted value of the dynamic stress intensity factor to the predicted value of the fracture toughness as the fracture failure risk index. When the fracture failure risk index exceeds a preset threshold, it generates graded early warning information and locates potential fracture failure sites and predicts the remaining safe life based on the changing trend of the fracture failure risk index.

8. The system according to claim 7, characterized in that, The data processing module includes: The standardization unit is used to denoise, fill in missing data, and standardize the multi-source heterogeneous monitoring data to obtain standardized multi-source heterogeneous monitoring data. The delay relationship acquisition unit is used to unify the timestamps of the standardized multi-source heterogeneous monitoring data using a time interpolation method, and to obtain the time delay relationship between each data based on the causal relationship between the physical changes of flood control and drainage structures and the multi-source heterogeneous monitoring data. The time alignment unit is used to align the standardized multi-source heterogeneous monitoring data according to a unified time reference and delay relationship to obtain time-aligned data; The spatial mapping unit is used to establish a three-dimensional spatial model of the flood control and drainage structure, and to map the position coordinates of each sensor used to collect multi-source heterogeneous monitoring data onto the three-dimensional spatial model to obtain the accurate position of each sensor in the three-dimensional spatial model. The spatial association unit is used to construct a sensor spatial distance matrix and topological relationship network based on the accurate location to obtain spatial association attributes; The dataset construction unit is used to assign the spatially correlated data to each data point in the time-aligned data to obtain a standardized fused dataset.

Images