Bridge structure crack spreading prediction method and system combined with incremental learning

Abstract

The invention provides a bridge structure crack spreading prediction method and system combined with incremental learning, and relates to the technical field of deep learning, and the method comprises the steps: collecting a sample test bridge crack image set and a sample test crack feature sequence set, carrying out the crack feature clustering, and determining a plurality of crack feature thresholds and a plurality of sample test data sets; using the plurality of sample test data sets to test the foundation crack propagation prediction model; and setting a plurality of adaptive incremental learning strategies according to the plurality of model prediction deviations, performing incremental learning on the basic crack propagation prediction model, and outputting a plurality of optimized crack propagation prediction models for prediction. The objective of the invention is to solve the technical problem that the accuracy and robustness of a prediction model are insufficient due to the fact that an existing prediction method generally adopts a unified incremental training strategy and lacks a differential adaptation mechanism for the actual evolution state of a crack. The prediction accuracy and robustness of the model on the complex crack evolution behavior can be improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and system for predicting crack propagation in bridge structures combined with incremental learning. Background Art

[0002] In recent years, with the advancement of artificial intelligence and deep learning technologies, a growing number of studies have incorporated neural network models into the field of bridge structural health monitoring. In particular, progress has been made in modeling and predicting the temporal evolution of cracks using recurrent neural networks (such as LSTM). Furthermore, incremental learning techniques, as an effective approach to addressing the problem of "catastrophic forgetting" during long-term learning, have also been explored for application in updating and optimizing bridge crack prediction models.

[0003] However, most existing bridge crack propagation prediction methods combined with incremental learning use a unified or fixed training strategy, which makes it difficult to fully consider the differences in shape, length, width, development speed, etc. between different cracks, and ignores the nonlinear evolution laws and diverse characteristics exhibited by the crack propagation process. In addition, the incremental learning process usually does not conduct in-depth analysis of the model prediction error and lacks a dynamic adaptation mechanism based on actual prediction performance. As a result, the model has problems such as insufficient accuracy and poor adaptability when dealing with complex crack states. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for predicting crack propagation in bridge structures using incremental learning. This method addresses the technical issues that existing methods for predicting crack propagation in bridge structures using incremental learning typically employ a unified incremental training strategy, lack a differentiated adaptation mechanism tailored to the actual evolution of cracks, and struggle to effectively address the diversity and nonlinear characteristics of crack development, resulting in insufficient accuracy and robustness of the prediction model. The system addresses the following issues: In a first aspect, the present invention provides a method for predicting crack propagation in bridge structures combined with incremental learning, comprising: taking the structural properties and environmental characteristics of the target bridge as constraints, collecting a first number of sample bridge crack image sets and sample crack feature sequence sets, training a long short-term memory network, and constructing a basic crack propagation prediction model; taking the structural properties and environmental characteristics of the target bridge as constraints, collecting a second number of sample test bridge crack image sets and sample test crack feature sequence sets, performing crack feature clustering, determining multiple crack feature thresholds, and multiple sample test data sets; using the multiple sample test data sets, iteratively testing the basic crack propagation prediction model to obtain multiple model prediction deviations; setting multiple adaptive incremental learning strategies based on the multiple model prediction deviations, performing incremental learning on the basic crack propagation prediction model, outputting multiple optimized crack propagation prediction models, and predicting crack propagation in bridge structures.

[0005] Preferably, the bridge structure crack propagation prediction method combined with incremental learning also includes: taking the structural properties of the target bridge as static constraints and the environmental characteristics as dynamic constraints, collecting a first number of sample bridge crack image sets based on historical bridge inspection records, and bridge crack images at P consecutive collection time points after the sample bridge crack image collection time point, as a second sample bridge crack image sequence, to obtain a second sample bridge crack image sequence set, wherein the first number is less than or equal to 1000 and P is greater than or equal to 10; using a convolutional neural network to construct a crack feature recognition model, extracting crack features from the sample bridge crack image set, and outputting a sample initial crack feature set, wherein the crack features include crack position, crack shape, crack length, crack width and crack depth; using the crack feature recognition model, extracting crack features from the second sample bridge crack image sequence set, and outputting a sample crack feature sequence set.

[0006] Preferably, the method for predicting crack propagation in bridge structures combined with incremental learning further includes: using the sample initial crack feature set as input, using the sample crack feature sequence set as supervision, training the long short-term memory network until the network converges, and obtaining a basic crack propagation prediction model.

[0007] Preferably, the bridge structure crack propagation prediction method combined with incremental learning also includes: taking the structural properties of the target bridge as static constraints and the environmental characteristics as dynamic constraints, collecting a second number of sample bridge crack image sets according to historical bridge inspection records as sample test bridge crack image sets, and collecting bridge crack images at P consecutive acquisition time points after the second number of acquisition time points to obtain a sample test bridge crack image sequence set, wherein the first number is greater than or equal to 5000; using the crack feature recognition model to extract crack features from the sample bridge crack image set to obtain a sample test crack feature set; using the crack feature recognition model to extract crack features from the sample test bridge crack image sequence set to obtain a sample test crack feature sequence set.

[0008] Preferably, the method for predicting crack propagation in bridge structures combined with incremental learning also includes: performing feature clustering based on the sample test crack feature set to determine multiple crack feature clustering results; performing feature interval extraction based on the multiple crack feature clustering results to determine multiple crack feature interval sets as multiple crack feature thresholds; based on the multiple crack feature thresholds, mapping and screening the sample test crack feature set respectively to determine multiple sample test input data sets; mapping and dividing the sample test crack feature sequence set based on the multiple sample test input data sets to obtain multiple sample test supervision data sets; mapping and combining the multiple sample test input data sets and multiple sample test supervision data sets to obtain multiple sample test data sets.

[0009] Preferably, the method for predicting crack propagation of a bridge structure combined with incremental learning further includes: using the multiple sample test data sets to iteratively test the basic crack propagation prediction model respectively, and outputting multiple sample test deviation sequence sets; performing mean square error and standard deviation calculations on the multiple sample test deviation sequence sets respectively, and determining multiple mean square error means and multiple standard deviation means; after normalizing the multiple mean square error means and multiple standard deviation means, performing model prediction error analysis respectively, and weightedly determining multiple model prediction deviations.

[0010] Preferably, the method for predicting crack propagation in bridge structures combined with incremental learning also includes: setting the ratio of the model prediction deviation to the preset standard model prediction deviation as an adjustment coefficient, and calculating multiple adjustment coefficients based on the multiple model prediction deviations; obtaining a fixed incremental learning strategy, wherein the fixed incremental learning strategy includes a standard loss function weight ratio and a standard new and old training data ratio; optimizing the standard loss function weight ratio and the standard new and old training data ratio according to the multiple adjustment coefficients to obtain multiple adaptive incremental learning strategies.

[0011] Preferably, the method for predicting crack propagation in bridge structures combined with incremental learning also includes: using the structural properties of the target bridge as static constraints, environmental characteristics as dynamic constraints, and crack feature thresholds as feature constraints, and collecting multiple incremental sample training data sets based on historical bridge inspection records; using the multiple incremental sample training data sets to perform incremental learning on the basic crack propagation prediction model, and outputting multiple optimized crack propagation prediction models.

[0012] Preferably, the bridge structure crack propagation prediction method combined with incremental learning also includes: establishing a mapping relationship between crack feature thresholds and optimized crack propagation prediction models, and based on the mapping relationship, constructing an optimization model matching library according to the multiple crack feature thresholds and multiple optimized crack propagation prediction models; collecting real-time bridge crack images and performing crack feature extraction to obtain real-time crack features; within the optimization model matching library, determining an adapted optimized crack propagation prediction model based on the real-time crack feature matching, performing crack propagation prediction, and outputting a predicted crack feature sequence as a bridge structure crack propagation prediction result.

[0013] In a second aspect, the present invention also provides a bridge structure crack propagation prediction system combined with incremental learning, which is used to execute a bridge structure crack propagation prediction method combined with incremental learning as described in the first aspect, including: a basic prediction model construction module, which is used to collect a first number of sample bridge crack image sets and sample crack feature sequence sets based on the structural properties and environmental characteristics of the target bridge, train a long short-term memory network, and construct a basic crack propagation prediction model; a test data set acquisition module, which is used to collect a second number of sample test bridge crack image sets and sample test crack feature sequence sets based on the structural properties and environmental characteristics of the target bridge, perform crack feature clustering, determine multiple crack feature thresholds, and multiple sample test data sets; a model prediction deviation analysis module, which is used to use the multiple sample test data sets to iteratively test the basic crack propagation prediction model respectively and obtain multiple model prediction deviations; an incremental learning optimization module, which is used to set multiple adaptive incremental learning strategies according to the multiple model prediction deviations, perform incremental learning on the basic crack propagation prediction model respectively, output multiple optimized crack propagation prediction models, and perform bridge structure crack propagation prediction.

[0014] The embodiments of the present invention include the following advantages: By taking the structural properties and environmental characteristics of the target bridge as constraints, a first number of sample bridge crack image sets and sample crack feature sequence sets are collected, a long short-term memory network is trained, and a basic crack propagation prediction model is constructed; then, taking the structural properties and environmental characteristics of the target bridge as constraints, a second number of sample test bridge crack image sets and sample test crack feature sequence sets are collected, crack feature clustering is performed, and multiple crack feature thresholds and multiple sample test data sets are determined; then, the basic crack propagation prediction model is iteratively tested using the multiple sample test data sets to obtain multiple model prediction deviations; finally, multiple adaptive incremental learning strategies are set according to the multiple model prediction deviations, and incremental learning is performed on the basic crack propagation prediction model respectively, outputting multiple optimized crack propagation prediction models to predict the propagation of cracks in bridge structures. In other words, by collecting multi-source sample data that matches the target bridge structure properties and environmental characteristics, and combining crack feature clustering with model prediction deviation analysis, it is possible to dynamically set incremental learning strategies for different crack evolution trends, thereby achieving personalized and differentiated optimization of the basic crack propagation prediction model; by constructing an adjustment coefficient based on the error normalization indicator and driving the adaptive adjustment of the loss function weight and the ratio of new and old training data, the model's prediction accuracy and robustness for complex crack evolution behavior can be improved, meeting the needs of high reliability and continuous updating capabilities in bridge structure health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of the steps of a method for predicting crack propagation in a bridge structure combined with incremental learning according to the present invention; Figure 2 This is a structural schematic diagram of a bridge structure crack propagation prediction system combined with incremental learning according to the present invention.

[0016] Description of reference numerals: Basic prediction model construction module 11, test data set acquisition module 12, model prediction deviation analysis module 13, incremental learning optimization module 14. DETAILED DESCRIPTION

[0017] The present invention provides a method and system for predicting crack propagation in bridge structures combined with incremental learning, thereby resolving the technical problem that existing methods for predicting crack propagation in bridges combined with incremental learning generally adopt a unified incremental training strategy, lack a differentiated adaptation mechanism for the actual evolution state of cracks, and are unable to effectively cope with the diversity and nonlinear characteristics in the crack development process, resulting in insufficient accuracy and robustness of the prediction model. By collecting multi-source sample data that matches the target bridge structure properties and environmental characteristics, and combining crack feature clustering with model prediction deviation analysis, it is possible to dynamically set incremental learning strategies for different crack evolution trends, thereby achieving personalized and differentiated optimization of the basic crack propagation prediction model; by constructing an adjustment coefficient based on an error normalization index and driving the adaptive adjustment of the loss function weight and the ratio of new and old training data, it is possible to improve the model's prediction accuracy and robustness for complex crack evolution behavior, meeting the requirements for high reliability and continuous updating capabilities in bridge structure health monitoring.

[0018] Below, the technical solutions of the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments described herein. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the accompanying drawings.

[0019] For example, see the attached Figure 1 The present invention provides a method for predicting crack propagation in a bridge structure combined with incremental learning, which is applied to a system for predicting crack propagation in a bridge structure combined with incremental learning, and specifically includes the following steps: S10: Based on the structural properties and environmental characteristics of the target bridge, a first number of sample bridge crack image sets and sample crack feature sequence sets are collected, a long short-term memory network is trained, and a basic crack propagation prediction model is constructed.

[0020] Furthermore, step S10 of the present invention further includes: S11: Taking the structural properties of the target bridge as static constraints and the environmental characteristics as dynamic constraints, a first number of sample bridge crack image sets and bridge crack images at P consecutive acquisition time points after the sample bridge crack image acquisition time point are collected according to historical bridge detection records as a second sample bridge crack image sequence to obtain a second sample bridge crack image sequence set, wherein the first number is less than or equal to 1000 and P is greater than or equal to 10; S12: Using a convolutional neural network to construct a crack feature recognition model, perform crack feature extraction on the sample bridge crack image set, and output a sample initial crack feature set, wherein the crack features include crack position, crack shape, crack length, crack width and crack depth; S13: Using the crack feature recognition model, perform crack feature extraction on the second sample bridge crack image sequence set, and output a sample crack feature sequence set.

[0021] Specifically, the structural attributes and environmental characteristics of the target bridge are first acquired. Structural attributes refer to inherent, short-term, characteristics of the bridge, such as bridge type (e.g., beam bridge, arch bridge), material (e.g., concrete, steel structure), span, age, and construction techniques. These characteristics determine the fundamental context for crack formation and serve as the primary constraints for selecting similar bridge samples. Environmental characteristics refer to variable external factors that influence crack development, such as temperature, humidity, precipitation, load (traffic volume), earthquakes, and corrosive environments. By matching similar environmental backgrounds, the true evolution of cracks over time can be more accurately simulated. Next, using the structural attributes of the target bridge as static constraints and the environmental characteristics as dynamic constraints, a first set of sample bridge crack image sets is collected based on historical bridge inspection records. Historical bridge inspection records refer to the long-term accumulation of bridge crack images and their acquisition time records through methods such as drone inspections, manual inspections, and sensor monitoring, constituting the sample data source. Specifically, while satisfying the aforementioned static and dynamic constraints, no more than 1,000 sets of bridge crack images are selected from the historical inspection database as observation data at the crack initiation moment (the first time point). Then, for each initial crack image, its acquisition time point is located, and image data of the crack at P consecutive time points thereafter (P is greater than or equal to 10) is further extracted to form a crack time-series image sequence. Each sequence records the changing state of the same crack at multiple consecutive moments, including changes in length and width, and serves as the second sample bridge crack image sequence, thus obtaining the second sample bridge crack image sequence set.

[0022] A crack feature recognition model is constructed using a convolutional neural network for crack feature extraction, wherein crack features include crack location, crack shape, crack length, crack width, and crack depth. Crack location refers to the spatial location of the crack on an image or bridge deck (e.g., coordinates, region labels); crack shape refers to the geometric outline of the crack, which can be obtained through edge detection or segmentation; crack length refers to the physical length along the main axis of the crack; crack width refers to the average or maximum spacing between the edges on both sides of the crack; crack depth refers to the extent of internal crack expansion, which can be estimated indirectly (e.g., based on brightness / texture changes). The crack feature recognition model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer is used to receive a bridge crack image, typically a size-normalized grayscale image or RGB image; the convolutional layer is used to extract local features, such as the texture, edge, and direction of the crack; the pooling layer is used to reduce feature dimensionality and computational complexity; the fully connected layer is used to perform feature fusion and comprehensive judgment, and output feature representations for classification / regression; and the output layer is used to output the structured features of the crack, including location, shape, length, width, and depth.

[0023] Manually annotated crack image data is further selected as training data, which must include the annotation information corresponding to each image (position, shape, length, width, depth), which can be obtained through image segmentation and annotation software (such as LabelMe); the crack feature recognition model is supervised and trained using the training data. During the training process, a batch gradient descent method is adopted, and an optimizer such as Adam or SGD is used. Each batch inputs images with corresponding annotations, and the multi-task prediction results of the crack features are calculated through forward propagation. The error is then calculated based on the multi-task loss function, and the model parameters are subsequently updated through back propagation. During training, the validation set is regularly used to evaluate the model performance, and the learning rate and hyperparameters are dynamically adjusted according to the performance until the loss function converges or the preset number of training rounds is reached, thereby obtaining a trained crack feature recognition model.

[0024] Then, using the crack feature recognition model, crack features are extracted from multiple sample bridge crack images in the sample bridge crack image set, and a sample initial crack feature set is output. The feature set includes the crack position, shape and quantized size data corresponding to each sample image; and using the crack feature recognition model, crack features are extracted from the second sample bridge crack image sequence set, and a sample crack feature sequence set is output.

[0025] Furthermore, step S10 of the present invention further includes: S14: Using the sample initial crack feature set as input and the sample crack feature sequence set as supervision, the long short-term memory network is trained until the network converges to obtain a basic crack propagation prediction model.

[0026] Specifically, feature vectors consisting of multidimensional crack features (location, shape, length, width, and depth) extracted from a sample set of bridge crack images serve as the model's input data. A chronological sequence of continuous crack feature data, reflecting the development and changes of bridge cracks at multiple time points, serves as supervisory labels to guide the model in learning the temporal dynamics of crack propagation. The Long Short-Term Memory (LSTM) network is used as the basic model architecture. Because LSTM can capture long-term dependencies in time series, it is well-suited for handling the temporal evolution of crack features. The network input is a sequence of crack feature vectors, and the network output predicts the future state of the crack features. The long short-term memory network is further trained using input data and supervision data. During the training process, the initial crack characteristics of the sample are first fed into the LSTM network as an input sequence. The network predicts the changing trend of future crack characteristics based on the input sequence through multiple iterations of time steps. Next, the predicted results are compared with the actual subsequent features in the sample crack feature sequence set, and the loss (such as mean square error) is calculated. The backpropagation algorithm and optimizer (such as Adam) are then used to adjust the network weights to minimize the prediction error. The training process continues until the loss function converges and the model performance reaches stability. After training is completed, a basic crack propagation prediction model is obtained, which can accurately predict the future development trend of cracks based on the current crack characteristics.

[0027] S20: Taking the structural properties and environmental characteristics of the target bridge as constraints, collect a second number of sample test bridge crack image sets and sample test crack feature sequence sets, perform crack feature clustering, and determine multiple crack feature thresholds and multiple sample test data sets.

[0028] Furthermore, step S20 of the present invention further includes: S21: Taking the structural properties of the target bridge as static constraints and the environmental characteristics as dynamic constraints, a second number of sample bridge crack image sets are collected according to historical bridge inspection records as sample test bridge crack image sets, and bridge crack images at P consecutive acquisition time points after the second number of acquisition time points are collected to obtain a sample test bridge crack image sequence set, wherein the first number is greater than or equal to 5000; S22: Using the crack feature recognition model, crack features are extracted from the sample bridge crack image set to obtain a sample test crack feature set; S23: Using the crack feature recognition model, crack features are extracted from the sample test bridge crack image sequence set to obtain a sample test crack feature sequence set.

[0029] Specifically, using the structural attributes of the target bridge as static constraints (such as fixed attributes like bridge type, material, and span) and environmental characteristics as dynamic constraints (such as changing conditions like temperature, humidity, traffic load, and wind speed), a second set of sample bridge crack images is collected based on historical bridge inspection records to serve as the crack image sample library for testing. The second set of images must be no less than 5,000 to ensure data adequacy and diversity. Subsequently, after each of the second set of crack image acquisition points, bridge crack images are continuously collected at P time points to form time series data. These time series data reflect the evolution of cracks at multiple consecutive time points, providing a basis for dynamic analysis, thereby obtaining a sample test bridge crack image sequence set.

[0030] Then, the crack feature recognition model is used to extract crack features from the sample bridge crack image set to obtain a sample test crack feature set; and the crack feature recognition model is used to extract crack features from the sample test bridge crack image sequence set to obtain a sample test crack feature sequence set.

[0031] Furthermore, step S20 of the present invention further includes: S24: Perform feature clustering based on the sample test crack feature set to determine multiple crack feature clustering results; S25: Perform feature interval extraction based on the multiple crack feature clustering results to determine multiple crack feature interval sets as multiple crack feature thresholds; S26: Based on the multiple crack feature thresholds, perform mapping and screening on the sample test crack feature set respectively to determine multiple sample test input data sets; S27: Map and divide the sample test crack feature sequence set based on the multiple sample test input data sets to obtain multiple sample test supervision data sets; S28: Perform mapping and combination based on the multiple sample test input data sets and the multiple sample test supervision data sets to obtain multiple sample test data sets.

[0032] Specifically, a clustering algorithm (such as the K-means algorithm) is first used to analyze the multidimensional feature data in the sample test crack feature set. The purpose of clustering is to group the crack features according to their similarity, obtaining multiple crack feature clustering results. Each cluster represents a group of crack features with similar attributes, and multiple crack feature clustering results are determined. Next, feature interval extraction is performed based on the multiple crack feature clustering results. That is, the range and distribution of each dimensional crack feature in the data of each crack feature clustering result are statistically analyzed. The feature interval of each cluster is extracted, including the location coverage area (the spatial distribution range of the crack on the bridge structure), shape feature interval (the shape feature range of the crack, such as the ratio of length to width, tortuosity, etc.), length interval (the minimum and maximum range of crack length), width interval (the distribution range of crack width), and depth interval (the variation range of crack depth). Multiple crack feature interval sets are determined, and each crack feature interval set is used as a crack feature threshold to obtain multiple crack feature thresholds. The feature interval threshold can accurately characterize the key attribute range of different crack types, providing a scientific basis for the differentiated adjustment of model strategies. By clustering and analyzing sample tests of crack characteristics, various crack feature intervals are extracted, and multiple multidimensional crack feature thresholds are determined. This provides support for the subsequent design of incremental learning strategies based on feature differences, lays a data foundation for the optimization of personalized and targeted crack propagation prediction models, and thus improves the accuracy and robustness of the prediction results.

[0033] Next, based on the multiple crack feature thresholds, the sample test crack feature sets are mapped and screened. Specifically, each crack feature threshold (including position, shape, length, width, depth, etc.) obtained in the previous step is used as a screening criterion. Samples in the sample test crack feature set are matched against each threshold one by one (i.e., determined to determine whether they fall within the interval corresponding to a threshold). Successfully matched sample features are assigned to corresponding categories, forming multiple sample test input data sets. Then, based on the multiple sample test input data sets, the sample test crack feature sequence set is mapped and partitioned. Specifically, for each crack image (or sample number) involved in the sample test input data set, its subsequent evolution sequence is found in the corresponding sample test crack feature sequence set. Time series crack feature evolution data corresponding to each input sample is obtained as a supervision label, ultimately resulting in multiple sample test supervision data sets, corresponding one-to-one to the input set. Finally, each set of "sample test input data sets" is combined with its corresponding "sample test supervision data set" to form a complete set of test data pairs, resulting in multiple sample test data sets for subsequent model evaluation.

[0034] S30: Using the multiple sample test data sets, iteratively test the basic crack propagation prediction model to obtain multiple model prediction deviations.

[0035] Furthermore, step S30 of the present invention further includes: S31: Using the multiple sample test data sets, the foundation crack propagation prediction model is iteratively tested respectively, and multiple sample test deviation sequence sets are output; S32: The mean square error and standard deviation of the multiple sample test deviation sequence sets are calculated respectively to determine multiple mean square error means and multiple standard deviation means; S33: After normalizing the multiple mean square error means and multiple standard deviation means, the model prediction error analysis is performed respectively to determine multiple model prediction deviations by weight.

[0036] Specifically, first, the basic crack propagation prediction model is iteratively tested using the multiple sample test data sets, that is, each sample test data set constructed previously (that is, a subset of test samples under different crack characteristic intervals) is input into the basic crack propagation prediction model (such as LSTM) for prediction, and the predicted value sequence of the group of test samples at each time point is output; then, the predicted value sequence is compared with the actual observed crack feature sequence, and the prediction deviation is recorded to obtain multiple sample test deviation sequence sets, wherein each test set corresponds to a sample test deviation sequence set, and the prediction error at each time point is recorded.

[0037] Then, the mean square error and standard deviation are calculated for the multiple sample test deviation sequence sets respectively. The mean square error is used to measure the overall deviation between the predicted value and the true value, reflecting the prediction accuracy; the standard deviation is used to measure the volatility and stability of the error, reflecting the stability of the model output. For each sample test deviation sequence set, the mean square error and the mean standard deviation are calculated respectively to obtain multiple mean square error means and multiple standard deviation means. Next, because the error index scales in different subsets may be different, it is necessary to unify the dimensions for comparison, normalize the multiple mean square error means and multiple standard deviation means, and weight the normalized error indicators to form a comprehensive prediction deviation indicator for each test set. The weights of the mean square error and standard deviation can be set according to the prediction requirements. For example, the weight of the two is set to 0.5, which indicates that both the prediction accuracy and the prediction stability are concerned; the weight of the mean square error is set to 0.7 and the weight of the standard deviation is set to 0.3, which indicates that the overall deviation between the model prediction value and the true value is minimized, and the accuracy of the prediction is given priority; the attention to error fluctuation is relatively weakened, and the stability of the model prediction is appropriately considered to avoid drastic fluctuations in the prediction results.

[0038] S40: Setting a plurality of adaptive incremental learning strategies according to the plurality of model prediction deviations, performing incremental learning on the basic crack propagation prediction model respectively, outputting a plurality of optimized crack propagation prediction models, and performing crack propagation prediction on the bridge structure.

[0039] Furthermore, step S40 of the present invention further includes: S41: Set the ratio of the model prediction deviation to the preset standard model prediction deviation as the adjustment coefficient, and calculate multiple adjustment coefficients based on the multiple model prediction deviations; S42: Obtain a fixed incremental learning strategy, wherein the fixed incremental learning strategy includes a standard loss function weight ratio and a standard new and old training data ratio; S43: Optimize the standard loss function weight ratio and the standard new and old training data ratio according to the multiple adjustment coefficients to obtain multiple adaptive incremental learning strategies.

[0040] Specifically, first, the ratio of the model prediction deviation to the preset standard model prediction deviation is set as the adjustment coefficient, wherein the preset standard model prediction deviation can be set by the ideal performance of the basic model on the validation set, such as the reference benchmark in the initial stage of system operation, and multiple adjustment coefficients are calculated based on the multiple model prediction deviations. Among them, if the adjustment coefficient is greater than 1, it indicates that the current model prediction effect is poor, and the model adjustment should be strengthened; if the adjustment coefficient is less than 1, it indicates that the current model is good, and the adjustment range can be appropriately reduced to maintain stability.

[0041] Next, a fixed incremental learning strategy is obtained. This strategy includes a standard loss function weight ratio and a standard old-new training data ratio. The loss function weight ratio refers to the weighted weights of the error between time steps and the error of different crack characteristics (location, width, etc.). The old-new training data ratio refers to the ratio of the number of new data to the number of old data during the incremental learning process, which is used to control whether the model retains old knowledge or adapts to new changes. The standard loss function weight ratio and the standard old-new training data ratio are then optimized based on multiple adjustment coefficients. Specifically, the training strategy is adjusted using the adjustment coefficient corresponding to each test dataset to achieve personalized incremental training. For example, the adjustment coefficient is multiplied by the new data ratio in the standard old-new training data ratio, and the optimized new data ratio is calculated as 1 minus the optimized new data ratio. If the adjustment coefficient is greater than 1, the weight of time series correlation or crack shape error is increased. By setting an adaptive incremental learning strategy, the model can be personalized based on different crack characteristics, thereby better addressing the diversity and nonlinearity of crack changes.

[0042] Furthermore, step S40 of the present invention further includes: S44: Using the structural properties of the target bridge as static constraints, the environmental characteristics as dynamic constraints, and the crack feature threshold as feature constraints, multiple incremental sample training data sets are collected based on historical bridge inspection records; S45: Using the multiple incremental sample training data sets, incremental learning is performed on the basic crack propagation prediction model respectively, and multiple optimized crack propagation prediction models are output.

[0043] Specifically, first, the structural properties of the target bridge are used as static constraints, the environmental characteristics are used as dynamic constraints, and the crack feature threshold is used as the feature constraint. That is, data that is "structurally similar" to the target bridge is screened to ensure that the training data is representative; historical data with "similar service environment" to the target bridge is screened to enhance the model's ability to adapt to actual working conditions; it is ensured that the training data reflects crack evolution trends similar to the target prediction task; multiple incremental sample training data sets are collected based on historical bridge inspection records, where the sample training data includes sample crack features and sample crack feature sequences used for incremental training, which can be obtained by feature extraction through the crack feature recognition model constructed above.

[0044] Then, the basic crack propagation prediction model is incrementally learned respectively according to the multiple adaptive incremental learning strategies using the multiple incremental sample training data sets. First, for each group of incremental sample training data sets, a copy of the trained basic crack propagation prediction model is made as the initial model of the current incremental training; each incremental sample training data set corresponds to a configured incremental learning strategy, which includes the loss function weight ratio, that is, the contribution ratio of old samples and new samples in the training process (such as 0.4:0.6); and the ratio of old and new training data, that is, the ratio of old samples and new samples mixed in each training batch (such as 1:2); then, the current incremental sample training data set is mixed with the sample data representing the old knowledge according to the strategy to construct a training batch to ensure that the model can take into account both historical knowledge and new knowledge. Crack behavior; each batch of training data is then input into the current model copy, and forward propagation calculations are performed in sequence to output a sequence of predicted values for crack propagation; then, according to the weight ratio of the loss function set in the incremental learning strategy, the loss of old samples and the loss of new samples are weightedly calculated, and based on the above weighted loss function, the gradient is calculated by back propagation, and the model parameters are updated using an optimizer (such as Adam) to achieve the model's continuous learning ability; after each round or several rounds of training, the validation set is used to evaluate the prediction effect of the current model in different feature intervals, and the model's performance in generalization ability and balance between old and new features is monitored; when the loss function converges or reaches the preset number of training rounds, the optimized crack propagation prediction model is output as one of the personalized prediction models corresponding to the current feature threshold, and multiple optimized crack propagation prediction models are obtained.

[0045] Furthermore, step S40 of the present invention further includes: S46: Establish a mapping relationship between crack feature thresholds and optimized crack propagation prediction models. Based on the mapping relationship, construct an optimized model matching library according to the multiple crack feature thresholds and multiple optimized crack propagation prediction models. S47: Collect real-time bridge crack images and extract crack features to obtain real-time crack features. S48: Within the optimized model matching library, determine an adapted optimized crack propagation prediction model based on the real-time crack feature matching, perform crack propagation prediction, and output a predicted crack feature sequence as a bridge structure crack propagation prediction result.

[0046] Specifically, a mapping relationship between crack feature thresholds and optimized crack propagation prediction models is first established, i.e., a one-to-one mapping relationship exists between different crack feature thresholds and their corresponding trained and optimized crack propagation prediction models. Next, based on this mapping relationship, an optimized model matching library (i.e., a lookup table of feature-model correspondences) is constructed based on the multiple crack feature thresholds and multiple optimized crack propagation prediction models. Then, during bridge inspection or online monitoring, an image of the bridge crack at the current moment is captured and input into a trained crack feature recognition model to obtain the current real-time crack feature vector. Furthermore, the optimized model matching library is used to search for the crack feature threshold interval closest to the real-time extracted crack feature, and the optimal crack propagation prediction model matching this feature interval is determined. Finally, the real-time crack feature is used as input to the optimized crack propagation prediction model for crack propagation prediction, and a predicted crack feature sequence is output as the crack propagation prediction result for the bridge structure. This feature-based dynamic model matching mechanism achieves precise adaptation of the crack propagation prediction model to the current structural state, effectively improving the timeliness, personalization, and robustness of the prediction, meeting the dual requirements of high dynamic response and high accuracy in structural health monitoring.

[0047] In summary, the method for predicting crack propagation in bridge structures combined with incremental learning provided by the present invention has the following technical effects: By taking the structural properties and environmental characteristics of the target bridge as constraints, a first number of sample bridge crack image sets and sample crack feature sequence sets are collected, a long short-term memory network is trained, and a basic crack propagation prediction model is constructed; then, taking the structural properties and environmental characteristics of the target bridge as constraints, a second number of sample test bridge crack image sets and sample test crack feature sequence sets are collected, crack feature clustering is performed, and multiple crack feature thresholds and multiple sample test data sets are determined; then, the basic crack propagation prediction model is iteratively tested using the multiple sample test data sets to obtain multiple model prediction deviations; finally, multiple adaptive incremental learning strategies are set according to the multiple model prediction deviations, and incremental learning is performed on the basic crack propagation prediction model respectively, outputting multiple optimized crack propagation prediction models to predict the propagation of cracks in bridge structures. In other words, by collecting multi-source sample data that matches the target bridge structure properties and environmental characteristics, and combining crack feature clustering with model prediction deviation analysis, it is possible to dynamically set incremental learning strategies for different crack evolution trends, thereby achieving personalized and differentiated optimization of the basic crack propagation prediction model; by constructing an adjustment coefficient based on the error normalization indicator and driving the adaptive adjustment of the loss function weight and the ratio of new and old training data, the model's prediction accuracy and robustness for complex crack evolution behavior can be improved, meeting the needs of high reliability and continuous updating capabilities in bridge structure health monitoring.

[0048] In the second embodiment, based on the same inventive concept as the method for predicting crack propagation of a bridge structure combined with incremental learning in the above embodiment, the present invention also provides a system for predicting crack propagation of a bridge structure combined with incremental learning, as shown in the attached figure. Figure 2 , including: a basic prediction model construction module 11, which is used to collect a first number of sample bridge crack image sets and sample crack feature sequence sets based on the structural properties and environmental characteristics of the target bridge, train a long short-term memory network, and construct a basic crack propagation prediction model; a test data set acquisition module 12, which is used to collect a second number of sample test bridge crack image sets and sample test crack feature sequence sets based on the structural properties and environmental characteristics of the target bridge, perform crack feature clustering, determine multiple crack feature thresholds, and multiple sample test data sets; a model prediction deviation analysis module 13, which is used to use the multiple sample test data sets to iteratively test the basic crack propagation prediction model respectively, and obtain multiple model prediction deviations; an incremental learning optimization module 14, which is used to set multiple adaptive incremental learning strategies according to the multiple model prediction deviations, perform incremental learning on the basic crack propagation prediction model respectively, output multiple optimized crack propagation prediction models, and perform bridge structure crack propagation prediction.

[0049] Furthermore, the bridge structure crack propagation prediction system combined with incremental learning is also used to: use the structural properties of the target bridge as static constraints and the environmental characteristics as dynamic constraints, collect a first number of sample bridge crack image sets based on historical bridge inspection records, and bridge crack images at P consecutive collection time points after the sample bridge crack image collection time point, as a second sample bridge crack image sequence, to obtain a second sample bridge crack image sequence set, wherein the first number is less than or equal to 1000 and P is greater than or equal to 10; use a convolutional neural network to construct a crack feature recognition model, extract crack features from the sample bridge crack image set, and output a sample initial crack feature set, wherein the crack features include crack position, crack shape, crack length, crack width and crack depth; use the crack feature recognition model to extract crack features from the second sample bridge crack image sequence set, and output a sample crack feature sequence set.

[0050] Furthermore, the bridge structure crack propagation prediction system combined with incremental learning is also used to: use the sample initial crack feature set as input, use the sample crack feature sequence set as supervision, train the long short-term memory network until the network converges, and obtain a basic crack propagation prediction model.

[0051] Furthermore, the bridge structure crack propagation prediction system combined with incremental learning is also used to: use the structural properties of the target bridge as static constraints and the environmental characteristics as dynamic constraints, collect a second number of sample bridge crack image sets based on historical bridge inspection records as sample test bridge crack image sets, and collect bridge crack images at P consecutive acquisition time points after the second number of acquisition time points to obtain a sample test bridge crack image sequence set, wherein the first number is greater than or equal to 5000; use the crack feature recognition model to extract crack features from the sample bridge crack image set to obtain a sample test crack feature set; use the crack feature recognition model to extract crack features from the sample test bridge crack image sequence set to obtain a sample test crack feature sequence set.

[0052] Furthermore, the bridge structure crack propagation prediction system combined with incremental learning is also used to: perform feature clustering based on the sample test crack feature set to determine multiple crack feature clustering results; perform feature interval extraction based on the multiple crack feature clustering results to determine multiple crack feature interval sets as multiple crack feature thresholds; based on the multiple crack feature thresholds, respectively map and screen the sample test crack feature set to determine multiple sample test input data sets; map and divide the sample test crack feature sequence set based on the multiple sample test input data sets to obtain multiple sample test supervision data sets; map and combine the multiple sample test input data sets and the multiple sample test supervision data sets to obtain multiple sample test data sets.

[0053] Furthermore, the bridge structure crack propagation prediction system combined with incremental learning is also used to: use the multiple sample test data sets to iteratively test the basic crack propagation prediction model, and output multiple sample test deviation sequence sets; calculate the mean square error and standard deviation of the multiple sample test deviation sequence sets, and determine multiple mean square error means and multiple standard deviation means; after normalizing the multiple mean square error means and multiple standard deviation means, perform model prediction error analysis, and weightedly determine multiple model prediction deviations.

[0054] Furthermore, the bridge structure crack propagation prediction system combined with incremental learning is also used to: set the ratio of the model prediction deviation to the preset standard model prediction deviation as an adjustment coefficient, and calculate multiple adjustment coefficients based on the multiple model prediction deviations; obtain a fixed incremental learning strategy, wherein the fixed incremental learning strategy includes a standard loss function weight ratio and a standard new and old training data ratio; optimize the standard loss function weight ratio and the standard new and old training data ratio according to the multiple adjustment coefficients to obtain multiple adaptive incremental learning strategies.

[0055] Furthermore, the bridge structure crack propagation prediction system combined with incremental learning is also used to: use the structural properties of the target bridge as static constraints, the environmental characteristics as dynamic constraints, and the crack feature threshold as feature constraints, and collect multiple incremental sample training data sets based on historical bridge inspection records; use the multiple incremental sample training data sets to perform incremental learning on the basic crack propagation prediction model, and output multiple optimized crack propagation prediction models.

[0056] Furthermore, the bridge structure crack propagation prediction system combined with incremental learning is also used to: establish a mapping relationship between crack feature thresholds and optimized crack propagation prediction models, and based on the mapping relationship, construct an optimization model matching library according to the multiple crack feature thresholds and multiple optimized crack propagation prediction models; collect real-time bridge crack images and extract crack features to obtain real-time crack features; within the optimization model matching library, determine an adapted optimized crack propagation prediction model based on the real-time crack feature matching, perform crack propagation prediction, and output a predicted crack feature sequence as a bridge structure crack propagation prediction result.

[0057] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The method and specific examples of the bridge structure crack propagation prediction method combined with incremental learning in the aforementioned embodiment 1 are also applicable to the bridge structure crack propagation prediction system combined with incremental learning in this embodiment. Through the aforementioned detailed description of the bridge structure crack propagation prediction method combined with incremental learning, those skilled in the art can clearly understand the bridge structure crack propagation prediction system combined with incremental learning in this embodiment. Therefore, for the sake of brevity of the specification, it will not be described in detail here. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0058] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

[0059] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention is intended to include these modifications and variations.

Claims

1. A method for predicting crack propagation in bridge structures combined with incremental learning, characterized in that: Methods include: Taking the structural properties and environmental characteristics of the target bridge as constraints, a first set of sample bridge crack image sets and sample crack feature sequence sets are collected, a long short-term memory network is trained, and a basic crack propagation prediction model is constructed; Taking the structural attributes and environmental characteristics of the target bridge as constraints, collecting a second number of sample test bridge crack image sets and sample test crack feature sequence sets, performing crack feature clustering, and determining multiple crack feature thresholds and multiple sample test data sets; Using the multiple sample test data sets, iteratively testing the basic crack propagation prediction model to obtain multiple model prediction deviations; A plurality of adaptive incremental learning strategies are set according to the plurality of model prediction deviations, and incremental learning is performed on the basic crack propagation prediction model respectively to output a plurality of optimized crack propagation prediction models to predict crack propagation of the bridge structure.

2. The method for predicting crack propagation in bridge structures combined with incremental learning according to claim 1, characterized in that: Taking the structural attributes and environmental characteristics of the target bridge as constraints, a first number of sample bridge crack image sets and sample crack feature sequence sets are collected, including: Using the structural properties of the target bridge as static constraints and the environmental characteristics as dynamic constraints, a first number of sample bridge crack image sets are collected based on historical bridge inspection records, as well as bridge crack images collected at P consecutive time points after the time point at which the sample bridge crack image was collected, as a second sample bridge crack image sequence, to obtain a second sample bridge crack image sequence set, wherein the first number is less than or equal to 1000 and P is greater than or equal to 10; A crack feature recognition model is constructed using a convolutional neural network to extract crack features from the sample bridge crack image set, and an initial crack feature set of the sample is output, wherein the crack features include crack location, crack shape, crack length, crack width, and crack depth; The crack feature recognition model is used to extract crack features from the second sample bridge crack image sequence set, and a sample crack feature sequence set is output.

3. The method for predicting crack propagation in bridge structures combined with incremental learning according to claim 2, characterized in that: Train the long short-term memory network and build a basic crack propagation prediction model, including: The sample initial crack feature set is used as input, the sample crack feature sequence set is used as supervision, and the long short-term memory network is trained until the network converges to obtain a basic crack propagation prediction model.

4. The method for predicting crack propagation in bridge structures combined with incremental learning according to claim 2, characterized in that: Taking the structural attributes and environmental characteristics of the target bridge as constraints, a second number of sample test bridge crack image sets and sample test crack feature sequence sets are collected, including: Using the structural properties of the target bridge as static constraints and the environmental characteristics as dynamic constraints, a second number of sample bridge crack image sets are collected based on historical bridge inspection records as the sample test bridge crack image set, and bridge crack images are collected at P consecutive acquisition time points after the second number of acquisition time points to obtain a sample test bridge crack image sequence set, where the first number is greater than or equal to 5000. Using the crack feature recognition model, extract crack features from the sample bridge crack image set to obtain a sample test crack feature set; The crack feature recognition model is used to extract crack features from the sample test bridge crack image sequence set to obtain a sample test crack feature sequence set.

5. The method for predicting crack propagation in bridge structures combined with incremental learning according to claim 4, characterized in that: Perform crack feature clustering, determine multiple crack feature thresholds, and multiple sample test data sets, including: Performing feature clustering based on the sample test crack feature set to determine multiple crack feature clustering results; Extracting characteristic intervals based on the plurality of fracture feature clustering results, and determining a plurality of fracture feature interval sets as a plurality of fracture feature thresholds; Based on the multiple crack feature thresholds, mapping and screening the sample test crack feature sets are performed respectively to determine multiple sample test input data sets; Mapping and dividing the sample test crack feature sequence set according to the multiple sample test input data sets to obtain multiple sample test supervision data sets; Mapping and combining the multiple sample test input data sets and the multiple sample test supervision data sets are performed to obtain multiple sample test data sets.

6. The method for predicting crack propagation in bridge structures combined with incremental learning according to claim 1, characterized in that: The basic crack propagation prediction model is iteratively tested using the multiple sample test data sets to obtain multiple model prediction deviations, including: Using the multiple sample test data sets, iteratively test the basic crack propagation prediction model respectively, and output multiple sample test deviation sequence sets; Calculating the mean square error and the standard deviation of each of the plurality of sample test deviation sequence sets to determine a plurality of mean square error means and a plurality of standard deviation means; After normalizing the multiple mean square error means and the multiple standard deviation means, model prediction error analysis is performed respectively, and multiple model prediction deviations are determined by weighting.

7. The method for predicting crack propagation in bridge structures combined with incremental learning according to claim 1, characterized in that: Setting multiple adaptive incremental learning strategies according to the multiple model prediction deviations includes: The ratio of the model prediction deviation to the preset standard model prediction deviation is set as an adjustment coefficient, and multiple adjustment coefficients are calculated based on the multiple model prediction deviations; Obtain a fixed incremental learning strategy, where the fixed incremental learning strategy includes a standard loss function weight ratio and a standard ratio of new and old training data; The standard loss function weight ratio and the standard new and old training data ratio are optimized according to the multiple adjustment coefficients to obtain multiple adaptive incremental learning strategies.

8. The method for predicting crack propagation in bridge structures combined with incremental learning according to claim 5, characterized in that: Incremental learning is performed on the basic crack propagation prediction model to output multiple optimized crack propagation prediction models, including: Taking the structural properties of the target bridge as static constraints, environmental characteristics as dynamic constraints, and crack feature thresholds as feature constraints, multiple incremental sample training datasets are collected based on historical bridge inspection records; The basic crack propagation prediction model is incrementally learned using the multiple incremental sample training data sets to output multiple optimized crack propagation prediction models.

9. The method for predicting crack propagation in bridge structures combined with incremental learning according to claim 8, characterized in that: Predict crack propagation in bridge structures, including: Establishing a mapping relationship between crack feature thresholds and optimized crack propagation prediction models, and constructing an optimized model matching library based on the mapping relationship and the multiple crack feature thresholds and the multiple optimized crack propagation prediction models; Collect real-time bridge crack images and extract crack features to obtain real-time crack features; In the optimization model matching library, an adaptive optimization crack propagation prediction model is determined according to the real-time crack feature matching, crack propagation prediction is performed, and a predicted crack feature sequence is output as a bridge structure crack propagation prediction result.

10. A bridge structure crack propagation prediction system combined with incremental learning, characterized in that: The steps for implementing the method for predicting crack propagation in a bridge structure combined with incremental learning as described in any one of claims 1 to 9 include: A basic prediction model construction module is used to collect a first number of sample bridge crack image sets and sample crack feature sequence sets based on the structural properties and environmental characteristics of the target bridge, train a long short-term memory network, and construct a basic crack propagation prediction model; a test data set acquisition module, configured to collect a second number of sample test bridge crack image sets and sample test crack feature sequence sets based on the structural properties and environmental characteristics of the target bridge, perform crack feature clustering, determine multiple crack feature thresholds, and multiple sample test data sets; a model prediction deviation analysis module, configured to perform iterative testing on the basic crack propagation prediction model using the multiple sample test data sets to obtain multiple model prediction deviations; The incremental learning optimization module is used to set multiple adaptive incremental learning strategies according to the multiple model prediction deviations, perform incremental learning on the basic crack propagation prediction model respectively, output multiple optimized crack propagation prediction models, and perform bridge structure crack propagation prediction.

Images