Concrete filled steel tube service life prediction method and system based on machine learning

Abstract

The invention discloses a concrete filled steel tube service life prediction method and system based on machine learning. According to the method, sensor data, external environment data, unmanned aerial vehicle images and satellite remote sensing data are collected to serve as original data and real-time monitoring data, fusion, denoising, normalization and missing value filling are conducted on the original data, and a feature matrix is generated. And training the feature matrix by using a machine learning model to obtain an initial life prediction model, and continuously receiving real-time data through an incremental learning algorithm for dynamic updating to obtain a dynamically optimized life prediction model. According to the structure aging characteristics, life prediction is divided into an initial aging stage, a middle-term damage stage and an advanced-age stage, and accurate prediction is carried out by applying customized models. And finally, integrating the prediction result to an intelligent decision support platform, and displaying the prediction result, the structure health state and the maintenance suggestion through a visual tool. The precision and real-time performance of service life prediction of the concrete-filled steel tube structure are remarkably improved, and the method has a wide application prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of life prediction of concrete-filled steel tube structures, and particularly relates to a method and system for predicting the life of concrete-filled steel tubes based on machine learning. Background Art

[0002] Concrete-filled steel tube is a composite material widely used in engineering structures such as bridges, buildings, and tunnels. Its excellent compressive strength, durability, and load-bearing capacity make it an important component in modern infrastructure construction. However, with the increase in service life, the performance of concrete-filled steel tube structures will gradually degrade, and problems such as cracks and corrosion will appear, ultimately affecting their safety and stability. Therefore, the life prediction of concrete-filled steel tubes has become a key issue for ensuring engineering safety and optimizing maintenance management.

[0003] Current methods for predicting the life of concrete-filled steel tubes mainly rely on traditional theoretical models, experimental data, and rule-based life calculations. Common technical solutions include prediction methods based on stress-strain models, environmental impact factor models, etc. These methods estimate the life of the structure by establishing a mathematical model of the concrete-filled steel tube structure and combining data such as stress and load during use.

[0004] Disadvantages or problems of the existing technical solutions:

[0005] (1) Model simplification and neglect of external complex factors

[0006] Existing life prediction models are mostly based on simplified theoretical assumptions, ignoring the complexity of the influence of external environments (such as climate change conditions, traffic loads, earthquakes, etc.) on concrete-filled steel tubes in actual use. Traditional methods mostly use static models for life prediction and cannot adapt to changes in external conditions in real time, and the accuracy of prediction results is greatly limited.

[0007] (2) Inability to update in real time and optimize dynamically:

[0008] Existing methods usually rely on one-time data collection and analysis, lacking real-time monitoring and dynamic update mechanisms. As time goes by and the environment changes, the health status of the structure will constantly change. Traditional prediction methods cannot be adaptively updated with new data, resulting in lagging prediction results and being difficult to accurately reflect the actual state of the current structure.

[0009] (3) Prediction accuracy is limited by data quality and computing power

[0010] Many existing methods for predicting the life of concrete-filled steel tubes rely on a large amount of experimental data and on-site tests, but these data are difficult to obtain and have relatively high costs. At the same time, the computational amount of traditional methods is large, resulting in poor applicability and practicality in complex environments.

[0011] (4) Unable to adapt to multi-stage aging processes

[0012] The aging process of concrete-filled steel tubes is multi-stage, and the damage characteristics are different in different stages. However, existing prediction methods usually treat the entire life cycle as a whole, failing to refine the prediction for different aging stages and unable to accurately reflect the aging process in different stages and its impact on the lifespan.

[0013] (5) The prediction model lacks cross-industry applicability:

[0014] Current technologies are mostly limited to single structures or materials and are difficult to extend to other types of composite structural materials, such as prestressed concrete, fiber-reinforced concrete, etc. Therefore, their application scope is limited and cannot meet the needs of multi-field engineering. Summary of the Invention

[0015] Aiming at the problems existing in the above-mentioned prior art, the present invention provides a method and system for predicting the lifespan of concrete-filled steel tubes based on machine learning, aiming to accurately predict the remaining service life of concrete-filled steel tube structures by integrating diversified data sources, machine learning models, and real-time dynamic learning mechanisms, so as to realize the scientific and intelligent health monitoring, risk warning, and maintenance decision-making of concrete-filled steel tube structures.

[0016] To achieve the above object, the specific solution of the present invention is as follows:

[0017] A method for predicting the lifespan of concrete-filled steel tubes based on machine learning, comprising the following steps:

[0018] Step 1, data collection: Firstly, collect sensor data, external environment data, UAV images, and satellite remote sensing data of the concrete-filled steel tube structure as original data, and subsequently continuously collect sensor data, external environment data, UAV images, and satellite remote sensing data of the concrete-filled steel tube structure as real-time monitoring data. The sensor data includes key parameter data such as the stress, strain, temperature, humidity, crack width, and corrosion condition of the concrete-filled steel tube structure, and the external environment data includes data on traffic loads, climate change conditions, and seismic activities; the climate change conditions include temperature, humidity, precipitation, and wind speed.

[0019] Step 2, data preprocessing and feature construction: Integrate the original data in Step 1, and perform denoising, normalization, and missing value filling to generate a feature matrix containing structural health parameters, environmental impact factors, historical usage conditions, and maintenance records.

[0020] Step 3, machine learning model training: Use a machine learning model to train the feature matrix described in Step 2 to obtain an initial lifespan prediction model.

[0021] Step 4, Model Dynamic Update: After the training of the initial life prediction model described in Step 3 is completed, continuously receive the real-time monitoring data described in Step 1, and use the incremental learning algorithm to dynamically update the initial life prediction model to obtain a dynamically optimized life prediction model;

[0022] Step 5, Phased Life Prediction: According to the aging characteristics of the concrete-filled steel tubular structure, divide the life prediction into the initial aging stage, the intermediate damage stage, and the advanced age stage, and use different machine learning models for life prediction in each stage with the dynamically optimized life prediction model described in Step 3;

[0023] Step 6, Prediction Result Integration and Display: Combine the life prediction results of each stage in Step 5 with the original data and real-time monitoring data collected in Step 1, integrate them into the intelligent decision support platform, and display the prediction results, structural health status, and maintenance suggestions through visualization tools.

[0024] Furthermore, the original data described in Step 2 is fused using the weighted average method, and the fusion formula is as follows:

[0025]

[0026] In the formula: X f represents the fused feature matrix; ω i is the weight of the i-th data source, and the weight is set according to its reliability or importance; X i represents the original feature matrix of the i-th data source;

[0027] The denoising formula is as follows:

[0028]

[0029] In the formula: X denoised [i] represents the denoised data; X represents the original data; k is the size of the filtering window; j represents the index of each data point in the window;

[0030] The normalization formula is as follows:

[0031]

[0032] In the formula: X represents the original data; μ represents the mean of the data; σ represents the standard deviation; X norm represents the normalized data; the missing value filling formula is as follows:

[0033]

[0034] In the formula: n represents the number of non-missing data; represents the sum of all non-missing values; X filledDenote the data points after filling in the missing values; X i Denote the values of each data point that is not missing in the original data;

[0035] The formula for the feature matrix is as follows:

[0036] X features =[X1,X2,…,X n

[0037] In the formula: X1, X2, …, X n Are the features of different data sources extracted from sensor data, external environment data, UAV images, and satellite remote sensing data.

[0038] Furthermore, the machine learning model is a random forest model, a support vector machine model, or a long short-term memory network model.

[0039] Furthermore, the formula for dynamically updating the life prediction model using the incremental learning algorithm in step 4 is as follows:

[0040]

[0041] In the formula: θ t Denotes the parameters of the model at time point t; θ t+1 Denotes the updated model parameters, and η is the learning rate, which is used to control the amplitude of model update; Is the gradient of the loss function with respect to the model parameters, calculated based on the new data D new Calculated.

[0042] Furthermore, in step 5, for the life prediction in the initial aging stage, a support vector machine or a random forest model is selected as the machine learning model, and the dynamically optimized prediction model is used for training, where the training features are crack width, temperature and humidity changes, and initial stress distribution;

[0043] For the life prediction in the intermediate damage stage, a long short-term memory network is selected as the machine learning model, and the dynamically optimized life prediction model is used for training, where the training features are historical crack propagation speed, stress-strain data, and temperature and humidity changes;

[0044] For the life prediction in the high-age stage, a deep neural network or a convolutional neural network is selected as the machine learning model, and the dynamically optimized life prediction model is used for training, where the training features are bearing capacity, crack distribution, and material degradation.

[0045] Furthermore, the formula for the life prediction in the initial aging stage in step 5 is as follows:

[0046]

[0047] ​In the formula: represents the output value predicted by the model; K(x i , x) is the kernel function; α i is the Lagrange multiplier; b is the bias term; N is the number of support vectors; x is the input feature; x i is the feature vector of a specific sample or data point in the input features;

[0048] The life prediction formula for the intermediate damage stage is as follows:

[0049] h t = f(W h hh t-1 + W x hx t + b h )

[0050]

[0051] In the formula: h t is the hidden state at time step t; x t is the input data at time step t; W h h, W x h is the weight matrix; b h represents the bias term of the hidden state; h t-1 is the hidden state at time step t - 1, usually containing all the historical information of the previous time step; b h is the bias term of the hidden state; is the predicted data; W hy represents the weight matrix from the hidden state to the output layer; b y represents the bias term of the output layer;

[0052] The formula for life prediction in the old - age stage is as follows:

[0053] h = f(Wx + b)

[0054] In the formula: h is the output of the neural network; x is the input feature; W is the weight matrix; b is the bias term; f is the activation function.

[0055] The concrete - filled steel tube life prediction system based on machine learning includes:

[0056] A data acquisition module, which is used to collect sensor data, external environment data, UAV images, and satellite remote - sensing data for the first time as raw data, and continuously collect sensor data, external environment data, UAV images, and satellite remote - sensing data as real - time monitoring data in the future; among them, the sensor data includes key parameter data such as stress, strain, temperature, humidity, crack width, and corrosion condition of the concrete - filled steel tube structure, and the external environment data includes data on traffic load, climate change conditions, and seismic activities;

[0057] A data processing module, which is used to fuse the collected raw data, and perform denoising, normalization, and missing value filling to generate a feature matrix containing structural health parameters, environmental impact factors, historical usage conditions, and maintenance records;

[0058] A machine learning module, which is used to train the feature matrix using a machine learning model to obtain an initial life prediction model;

[0059] An online learning module, which is used to continuously receive real-time monitoring data after the initial life prediction model is trained, and use an incremental learning algorithm to dynamically update the life prediction model to obtain a dynamically optimized life prediction model;

[0060] A phased life prediction module, which is used to divide life prediction into an initial aging stage, a mid-term damage stage, and a high-age stage according to the aging characteristics of the concrete-filled steel tubular structure, and use different machine learning models for life prediction in each stage using the dynamically optimized life prediction model;

[0061] An intelligent decision support platform, which is used to combine the life prediction results of each stage with the collected raw data and the received real-time monitoring data, integrate them into the intelligent decision support platform, and display the prediction results and the structural health status through a visualization tool.

[0062] Advantages of the present invention

[0063] 1. Significantly improve prediction accuracy and reliability

[0064] Multi-source data fusion: By integrating multi-source information such as sensor data, environmental factors, drone images, satellite remote sensing data, etc., more comprehensive and accurate structural health information is provided compared to traditional methods, thus significantly improving the prediction accuracy of the model.

[0065] Phased life prediction: Different machine learning models and features are used for different aging stages to further refine the prediction process and effectively improve the prediction accuracy.

[0066] 2. Enhanced real-time performance and dynamic adaptability

[0067] Online learning mechanism: Introduce an online learning and dynamic update mechanism, enabling the machine learning model to receive new data in real time and dynamically optimize, significantly improving the timeliness and accuracy of prediction.

[0068] Environmental adaptability: Consider the influence of different environmental conditions, such as high temperature, humidity, cold, etc. on the structural life, and dynamically adjust the prediction model to ensure the accuracy and practicality of the prediction results in a changing environment.

[0069] 3. Wide cross-industry applicability

[0070] Multi - field expansion ability: The present invention is not only applicable to concrete - filled steel tube structures, but can also be extended and applied to other composite structural materials, such as prestressed concrete, fiber - reinforced concrete, etc., having broad cross - industry application potential.

[0071] 4. Intelligent decision - making support and risk warning

[0072] Intelligent decision - making platform: Provide an intelligent decision - making support platform, which can monitor the structural health status in real time through visualization tools, provide scientific maintenance suggestions and risk warnings, help decision - makers identify potential problems in advance, reduce maintenance costs, and improve the safety and reliability of the structure.

[0073] 5. Improve the efficiency of life - cycle management

[0074] Accurate prediction and phased management: Through accurate prediction and phased management, it helps managers plan maintenance and repair work more effectively, extend the service life of the structure, reduce the occurrence of sudden accidents, and thus greatly improve the management efficiency of the entire structural life cycle.

[0075] 6. Innovation and market competitiveness

[0076] Comprehensive innovative technology: The present invention integrates innovative technologies such as multi - source data fusion, online learning mechanism, phased life prediction, and intelligent decision - making support platform. Compared with the existing technologies, it has higher accuracy, practicability, and intelligence level.

[0077] Adaptability to variable environments: With strong environmental adaptability and cross - industry application capabilities, it can provide stable and reliable prediction results in variable environments, having broad application prospects and market value.

[0078] By introducing innovative technologies such as multi - source data fusion, online learning mechanism, phased life prediction, and intelligent decision - making support platform, the present invention significantly improves the accuracy, real - time performance, and adaptability of the life prediction of concrete - filled steel tubes. Compared with the existing technologies, the present invention can provide more accurate, more practical, and more intelligent prediction results in variable environments, having significant innovation and broad market application prospects. Brief description of the drawings

[0079] Figure 1 It is a flow chart of the method for predicting the service life of concrete - filled steel tubes based on machine learning of the present invention. Detailed implementation manners

[0080] The present invention will be further explained and illustrated below in combination with the drawings and specific implementation manners. It should be noted that this specific embodiment is not used to limit the scope of the rights of the present invention.

[0081] Such as Figure 1As shown in the figure, this specific embodiment provides a method and system for predicting the service life of concrete-filled steel tubes based on machine learning, which is mainly applied to the health monitoring, risk warning and maintenance decision-making of concrete-filled steel tube structures, and is widely applicable to the structural health management of bridges, buildings, tunnels and other infrastructure.

[0082] The method for predicting the service life of concrete-filled steel tubes based on machine learning includes the following steps:

[0083] Step 1, data collection: Firstly, collect sensor data, external environment data, UAV images and satellite remote sensing data of the concrete-filled steel tube structure as original data, and then continuously collect sensor data, external environment data, UAV images and satellite remote sensing data of the concrete-filled steel tube structure as real-time monitoring data. The sensor data includes key parameter data such as the stress, strain, temperature, humidity, crack width and corrosion condition of the concrete-filled steel tube structure, which is used to reflect the health status of the concrete-filled steel tube structure; the external environment data includes data on traffic load, climate change conditions and seismic activities; the climate change conditions include temperature, humidity, precipitation and wind speed; the external environment data has an important impact on the service life of the concrete-filled steel tube structure. The UAV images are images or videos of the concrete-filled steel tube structure taken by the UAV for the detection and analysis of structural surface defects (such as cracks and corrosion). The satellite remote sensing data is macro data related to the environment collected using satellite remote sensing technology to obtain the influence of more extensive environmental factors.

[0084] Step 2, data preprocessing and feature construction: Integrate the original data in Step 1, and perform denoising, normalization and missing value filling. According to the actual engineering situation and expert experience, select features closely related to the service life of the concrete-filled steel tube, and the features include but are not limited to the historical crack development situation, stress-strain curve and changes in environmental factors (such as temperature fluctuations, humidity, etc.), and generate a feature matrix containing structural health parameters, environmental impact factors, historical usage conditions and maintenance records; the structural health parameters include stress and cracks. The environmental impact factors include climate change conditions and loads.

[0085] Denoising: Remove abnormal data and noise to ensure data quality.

[0086] Normalization: Standardize data with different dimensions to ensure that the influence of each data on the prediction model has the same weight.

[0087] Missing value filling: Use appropriate interpolation methods or other techniques to fill in missing values to ensure data integrity.

[0088] (1) Integrate the original data

[0089] The purpose of data fusion is to unify information from different sources such as sensor data, external environment data, and UAV image data into a usable feature matrix. The original data adopts fusion methods such as weighted average method, principal component analysis (PCA), etc. Assume X1, X2, …, X n , are each data source, and the fused feature matrix is X f . The fusion formula is as follows:

[0090]

[0091] In the formula: X f represents the fused feature matrix; ω i is the weight of the i-th data source, and the weight is set according to its reliability or importance; X i represents the original feature matrix of the i-th data source.

[0092] (2) Denoising

[0093] Denoising uses classical filtering methods such as median filtering, wavelet transform, or mean filtering, etc. Assume the original data is X, and the denoised data is X denoised . A common denoising method is mean filtering: The denoising formula is as follows:

[0094]

[0095] In the formula: X denoised [i] represents the denoised data; X represents the original data; k is the size of the filtering window; j represents the index of each data point in the window.

[0096] (3) Normalization

[0097] Data normalization is to convert data of different scales into a unified scale. Common methods include maximum normalization and standardization. The standardization formula for normalization is:

[0098]

[0099] In the formula: X represents the original data; μ represents the mean of the data; σ represents the standard deviation; X norm represents the normalized data;

[0100] (4) Missing value imputation

[0101] Common imputation methods include mean imputation, forward imputation, backward imputation, etc. Assume the missing data is X missing . The following missing value imputation formula is used for mean imputation:

[0102]

[0103] Where: n represents the number of non-missing data; represents the sum of all non-missing values; X filled represents the data point after filling in the missing values; X i represents the value of each non-missing data point in the original data.

[0104] (5) The expression of the feature matrix is as follows:

[0105] X features =[X1,X2,…,X n

[0106] Where: X1, X2, …, X n are the features of different data sources extracted from sensor data, external environment data, UAV images, and satellite remote sensing data.

[0107] Step 3, Machine learning model training: Use at least one machine learning model among random forest, support vector machine, and long short-term memory network to train the feature matrix described in Step 2 to obtain an initial life prediction model;

[0108] Random Forest (RF): Used to process complex non-linear data and make predictions through an ensemble of decision trees.

[0109] Support Vector Machine (SVM): Used for classification and regression tasks, capable of handling high-dimensional data and adapting to different environmental changes.

[0110] Long Short-Term Memory Network (LSTM): Used to process time series data, such as the changing trend of monitoring data over time. LSTM is particularly suitable for modeling the aging process of concrete-filled steel tubes and its periodic characteristics.

[0111] Adopt a model integration method to perform weighted averaging or voting selection on the prediction results of multiple machine learning models to improve the stability and accuracy of the prediction. Each machine learning model adjusts its parameters through training data and combines its respective advantages during prediction to output the initial life prediction result.

[0112] The specific steps for training the feature matrix using a machine learning model are as follows:

[0113] (1) Random Forest (Random Forest, RF) training steps

[0114] Random Forest is an ensemble learning method commonly used to process complex non-linear data and can effectively reduce the risk of overfitting. Its training process includes the following steps:

[0115] A. Data input: Use the feature matrix X features as the input data, including features such as structural health parameters and environmental impact factors.​

[0116] B. Constructing the decision tree: Random forests construct multiple decision trees through random sampling of the dataset and feature selection. The training data for each tree is a random subset of the original dataset.

[0117] C. Voting mechanism: During prediction, the results of all trees are integrated through a voting mechanism, and the final prediction result is the weighted average of all tree results (i.e., regression task) or majority voting (i.e., classification task).

[0118] The prediction formula of the random forest is expressed as:

[0119]

[0120] In the formula: is the prediction result of the i-th decision tree, N is the number of decision trees in the forest, is the final prediction result of the random forest model.

[0121] (2) Training steps of Support Vector Machine (SVM)

[0122] Support Vector Machine is a powerful supervised learning model, especially suitable for dealing with high-dimensional data and classification tasks. It classifies or regresses by finding a hyperplane with the maximum margin.

[0123] The training process includes the following steps:

[0124] A. Data input: Input the feature matrix X features into the Support Vector Machine.

[0125] B. Mapping to a high-dimensional space: Use a kernel function (such as linear kernel, RBF kernel, etc.) to map the data to a high-dimensional space so that an optimal hyperplane can be found in this space.

[0126] C. Optimization problem: Solve the optimization problem to find the hyperplane parameters that maximize the classification boundary.

[0127] For the regression problem, the goal of SVM is to minimize the following loss function:

[0128]

[0129] In the formula, ω is the normal vector of the hyperplane, C is the regularization parameter, and ξ i is the slack variable, representing the tolerance of each data point.

[0130] For the classification problem, the goal of SVM is:

[0131] y i (ωT x i + b) ≥ 1 - ξ i , ξ i ≥ 0

[0132] Among them, y i is the sample label, x i is the input feature, ω and b are the model parameters, and ξ i is the slack variable.

[0133] (3) Training steps of Long Short-Term Memory (LSTM) network

[0134] LSTM is a specific type of Recurrent Neural Network (RNN), which is very suitable for processing time-series data and can effectively learn long-range dependencies in time-series data.

[0135] The training process includes the following steps:

[0136] A. Data input: Input the feature matrix X features into the LSTM model. The format of the data is usually a sequence of time steps, such as the structural health status data at each time point.

[0137] B. Network structure: LSTM contains several LSTM units, and each unit contains an input gate, a forget gate, an output gate, and a cell state. Information flow is controlled through these gates to retain important information and discard unimportant historical information.

[0138] C. Gradient descent optimization: Adjust the weights in LSTM through the backpropagation algorithm and use gradient descent to optimize the loss function. The core calculations of LSTM include the following steps:

[0139] f t = σ(W f · [h t-1 , x t + b f )

[0140] i t = σ(W f · [h t-1 , x t + b i )

[0141]

[0142] o t = σ(W o · [h t-1 , x t + b0)

[0143] ht = o t *tanh(C t )

[0144] In the formula, f t is the forgetting gate, i t is the input gate, is the candidate memory cell, C t is the cell state, o t is the output gate, h t is the output of the LSTM.

[0145] (4) Model Selection

[0146] Data Preparation: The feature matrix X is obtained through data preprocessing including denoising, normalization, and missing value imputation features

[0147] Model Selection: Select a suitable machine learning model according to the characteristics of the problem. The machine learning models are random forest, support vector machine, or long short-term memory network.

[0148] Model Training: Use the training data, which is the feature matrix X features , to optimize the parameters of the machine learning model.

[0149] Model Evaluation and Optimization: Optimize the hyperparameters through methods such as cross-validation and grid search, and evaluate the performance of the model.

[0150] (5) Specific Steps and Formulas for Training

[0151] A. Input data X features .

[0152] B. Use the selected machine learning model (such as random forest, support vector machine, or long short-term memory network).

[0153] C. Train the model and optimize the loss function. For the support vector machine, minimize the loss; for the random forest, make predictions based on the decision tree ensemble; for the long short-term memory network, minimize the time series prediction error.

[0154] D. Evaluate the model performance, adjust the hyperparameters, and output the initial life prediction model

[0155] Step 4, Model Dynamic Update: After the initial life prediction model described in Step 3 is trained, continuously receive the real-time monitoring data described in Step 1 to update the training data of the initial life prediction model; use the incremental learning algorithm to dynamically update the initial life prediction model, so that the initial life prediction model can continuously adapt to new environmental changes over time, improve the prediction accuracy, and make real-time adjustments according to the actual state of the structure, obtaining a dynamically optimized life prediction model to ensure that the prediction results of the dynamically optimized life prediction model are highly consistent with the current health status of the structure.

[0156] The steps of using the incremental learning algorithm to dynamically update the life prediction model in Step 4 are as follows:

[0157] A. Initial Training Model

[0158] In Step 3, use historical data and the initial feature matrix to train a basic life prediction model (such as random forest, support vector machine, or long short-term memory network, etc.). At this time, the model already has preliminary prediction capabilities.

[0159] B. Real-time Data Input

[0160] Obtain new monitoring data in real time from Step 1, including but not limited to sensor data such as stress, temperature and humidity, and crack width. The new data is gradually added to the training dataset.

[0161] C. Data Preprocessing

[0162] Preprocess the real-time data, including denoising, missing value filling, normalization, etc., to ensure that the quality of the new data meets the requirements of model update.

[0163] D. Incremental Learning to Update the Model

[0164] The formula for using the incremental learning algorithm to dynamically update the initial life prediction model is as follows:

[0165]

[0166] In the formula: θ t represents the parameters of the model at time point t; θ t+1 represents the updated model parameters; η is the learning rate; used to control the amplitude of model update; is the gradient of the loss function with respect to the model parameters; calculated based on the new data D new obtained.

[0167] Through incremental learning, the model parameters are slightly updated each time new data is received, enabling the model to continuously adapt to new data.

[0168] E. Model Validation and Adjustment

[0169] After the incremental learning update, the prediction accuracy of the model is evaluated by predicting new data and comparing it with the actual observed values. If necessary, adjust the parameters of incremental learning (such as learning rate, update frequency, etc.) to ensure the stability and accuracy of the model.

[0170] F. Update Prediction Results

[0171] Use the dynamically optimized remaining life prediction model to re-predict the remaining life of the structure. At this time, the model has adapted to the latest environmental changes and can provide accurate prediction results based on the actual state of the structure.

[0172] The key technologies in the application of incremental learning are mainly the following two aspects:

[0173] A. Challenges and Countermeasures of Incremental Learning

[0174] In an environment where real-time data is constantly changing, incremental learning needs to handle the efficient processing and updating of data streams. To avoid overfitting or underfitting, the following strategies are adopted:

[0175] Data stream balance: By selectively processing certain important data (such as highly variable data, abnormal data points, boundary data, extreme value data, the latest monitoring data, data in critical time periods, and feature data showing high correlation in historical data), the model can be prevented from being over-adjusted. These important data are as follows:

[0176] a. Highly variable data: Sensor data with large changes or mutations (such as stress, temperature and humidity, crack width, etc.). These data usually indicate important changes in the structural health condition and may have a greater impact on remaining life prediction.

[0177] b. Abnormal data points: Any data points that deviate from the normal pattern, especially sudden fluctuations or anomalies, usually indicate a change in the device state and require timely model updates to reflect these changes.

[0178] c. Boundary data refers to the values close to the edge of the dataset setting range, representing some critical and extreme conditions. Extreme value data usually refers to the maximum and minimum values in the data, representing the extreme points of the data. These data usually represent extreme situations that may affect the structural health. They are crucial for the accuracy and robustness of the model, especially when predicting the remaining life.

[0179] d. The latest monitoring data: Compared with historical data, the latest sensor data can better reflect the current structural health condition. Therefore, they should be regarded as relatively important data sources and given priority for incremental updates.

[0180] e. Data during key time periods: Monitoring data for some specific time periods, such as data when major events occur to the structure or important environmental changes are encountered (such as temperature changes, impact events, etc.), need to be particularly concerned about and added to the training data in a timely manner for updating.

[0181] f. Feature data showing high correlation in historical data: Features (such as combinations of specific sensors) in which some specific monitoring data show high correlation with the structural health status. This data can be used as priority data for model updating to help better capture and predict the health changes of the structure.

[0182] Regularization method: Introduce a regularization term during model updating to prevent overfitting.

[0183] B. Frequency control of incremental update

[0184] During model updating, the incremental learning algorithm is controlled according to the change frequency of real-time data. For situations where the data changes rapidly, a higher update frequency is set; for situations where the change is slow, the update frequency is reduced to reduce the computational burden.

[0185] Step 5, phased life prediction: According to the aging characteristics of the concrete-filled steel tubular structure, the life prediction is divided into the initial aging stage, the intermediate damage stage, and the advanced age stage. The dynamically optimized life prediction model described in Step 3 is used to apply different machine learning models to each stage for life prediction;

[0186] (1) Initial aging stage

[0187] Characteristics: In the initial stage of the structure being put into use, the concrete-filled steel tubular structure is usually relatively stable, mainly monitoring the development of cracks, minor deformations, and changes in environmental conditions (such as temperature, humidity, etc.).

[0188] Prediction model selection: In this stage, models such as support vector machine (SVM) or random forest (RF) are used for prediction. These models can better handle the non-linear relationships of the initial structural health status and are suitable for monitoring minor changes in the early stage.

[0189] Prediction steps:

[0190] A. Feature selection: Mainly select features such as crack width, temperature and humidity changes, and initial stress distribution.

[0191] B. Training process: Use support vector machine (SVM) to train the initial aging data, and use classification or regression tasks based on supervised learning for model training.

[0192] C. Prediction formula:

[0193]

[0194] In the formula: represents the output value predicted by the model; K(x i , x) is the kernel function; α i is the Lagrange multiplier; b is the bias term; N is the number of support vectors; x is the input feature; x i is the feature vector of a specific sample or data point in the input features.

[0195] (2) Intermediate damage stage

[0196] Characteristics: In this stage, the damage of the concrete-filled steel tube structure begins to appear. Common damage forms include corrosion, crack propagation, material fatigue, etc. It is necessary to closely monitor factors such as crack width, stress distribution, temperature and humidity changes of the structure.

[0197] Prediction model selection: The long short-term memory network (LSTM) is adopted because this model is good at processing time series data and can effectively capture the aging characteristics changing with time.

[0198] Prediction steps:

[0199] Feature selection: Select historical crack propagation speed, stress-strain data, temperature and humidity changes, etc. as features.

[0200] Training process: Use the LSTM model for training. The model can learn the time series changes of the structure in the intermediate aging stage and predict the life through time series.

[0201] Prediction formula:

[0202] h t = f(W h hh t-1 + W x hx t + b h )

[0203]

[0204] In the formula: h t is the hidden state at time step t; x t is the input data at time step t; W h h, W x h are the weight matrices; b h represents the bias term of the hidden state; h t-1 is the hidden state at time step t - 1, usually containing all historical information of the previous time step; b h is the bias term of the hidden state; is the predicted data; W hy represents the weight matrix from the hidden state to the output layer; b yRepresents the bias term of the output layer.

[0205] (3) Advanced age stage:

[0206] Characteristics: In the advanced age stage of the concrete-filled steel tube structure, the performance degradation of the structure is obvious, and serious cracks, embrittlement of materials, and decline in bearing capacity may occur. At this time, more complex models are needed for life prediction.

[0207] Prediction model selection: Deep neural network (DNN) or convolutional neural network (CNN) is adopted. These models have strong feature learning ability and can handle complex non-linear problems and multi-dimensional data.

[0208] Prediction steps:

[0209] A. Feature selection: Select the bearing capacity, crack distribution, material degradation, etc. of the concrete-filled steel tube as input features.

[0210] B. Training process: Use the deep neural network (DNN) to train the data. The model can automatically learn the complex relationships in the advanced age stage and conduct life prediction.

[0211] C. Prediction formula:

[0212] h = f(Wx + b)

[0213] In the formula: h is the output of the neural network, x is the input feature, W is the weight matrix, b is the bias term, and f is the activation function (such as ReLU, Sigmoid).

[0214] According to the aging characteristics of the concrete-filled steel tube structure, in this embodiment, different machine learning models (such as SVM, RF, LSTM, DNN, etc.) are used for accurate prediction in different aging stages. These models are dynamically optimized and updated in each stage to ensure the accuracy and reliability of life prediction and provide a scientific basis for structural health management and maintenance decisions.

[0215] Step 6, Prediction result integration and display: Combine the life prediction results of each stage in Step 5 with the original data and real-time monitoring data collected in Step 1, integrate them into the intelligent decision support platform, and display the prediction results, structural health status, and maintenance suggestions through visualization tools.

[0216] The functions of the intelligent decision support platform include:

[0217] (1) Real-time monitoring: Display the health data and life prediction results of the concrete-filled steel tube.

[0218] (2) Risk warning: According to the prediction results, when the structure is close to the end of its life cycle or abnormal conditions occur, the platform will send out warning signals.

[0219] (3) Decision support: The intelligent decision support platform provides multi-dimensional decision support functions to help management personnel formulate maintenance plans, optimize resource allocation, and extend the service life of the structure.

[0220] The concrete-filled steel tube life prediction system based on machine learning includes:

[0221] A data acquisition module, which is used to initially collect sensor data, external environment data, UAV images, and satellite remote sensing data as raw data, and subsequently continuously collect sensor data, external environment data, UAV images, and satellite remote sensing data as real-time monitoring data; among them, the sensor data includes key parameter data such as the stress, strain, temperature, humidity, crack width, and corrosion condition of the concrete-filled steel tube structure, and the external environment data includes data on traffic load, climate change conditions, and seismic activities;

[0222] A data processing module, which is used to fuse the collected raw data, and perform denoising, normalization, and missing value filling to generate a feature matrix including structural health parameters, environmental impact factors, historical usage conditions, and maintenance records;

[0223] A machine learning module, which is used to train the feature matrix using a machine learning model to obtain an initial life prediction model;

[0224] An online learning module, which is used to continuously receive real-time monitoring data after the initial life prediction model is trained, and use an incremental learning algorithm to dynamically update the life prediction model to obtain a dynamically optimized life prediction model;

[0225] A phased life prediction module, which is used to divide life prediction into an initial aging stage, a mid-term damage stage, and a high-age stage according to the aging characteristics of the concrete-filled steel tube structure, and use the dynamically optimized life prediction model to apply different machine learning models to each stage for life prediction;

[0226] An intelligent decision support platform, which is used to combine the life prediction results of each stage with the collected raw data and the received real-time monitoring data, integrate them into the intelligent decision support platform, and display the prediction results and the structural health status through a visualization tool.

[0227] To enhance the adaptability of the machine learning model, the present invention adjusts the machine learning model according to different environments, such as high temperature, humidity, cold regions, etc., so that the prediction results can accurately reflect the life of the concrete-filled steel tube structure under different climate conditions. In addition, this specific embodiment also has cross-industry adaptability and can be extended to other composite structural materials, such as prestressed concrete, fiber-reinforced concrete, etc., to provide life prediction services in multiple fields.

[0228] In this specific embodiment, through the integration of diversified data sources, the integration of machine learning models, the introduction of an online learning module, the application of a phased life prediction method, and the implementation of an intelligent decision support platform, the accuracy and real-time performance of the life prediction of concrete-filled steel tubes are comprehensively improved, providing a scientific basis for the health management of concrete-filled steel tube structures.

Claims

1. A method for predicting the service life of concrete-filled steel tubes based on machine learning, characterized in that, It includes the following steps: Step 1, data collection: For the first time, collect sensor data, external environment data, UAV images and satellite remote sensing data of the concrete-filled steel tubular structure as original data, and subsequently continuously collect sensor data, external environment data, UAV images and satellite remote sensing data of the concrete-filled steel tubular structure as real-time monitoring data. The sensor data includes key parameter data such as the stress, strain, temperature, humidity, crack width and corrosion condition of the concrete-filled steel tubular structure, and the external environment data includes data on traffic load, climate change conditions and seismic activities; the climate change conditions include temperature, humidity, precipitation and wind speed; Step 2, data preprocessing and feature construction: Integrate the original data in Step 1, and perform denoising, normalization and missing value filling to generate a feature matrix containing structural health parameters, environmental impact factors, historical usage conditions and maintenance records; Step 3, machine learning model training: Use a machine learning model to train the feature matrix described in Step 2 to obtain an initial life prediction model; Step 4, model dynamic update: After the initial life prediction model in Step 3 is trained, continuously receive the real-time monitoring data described in Step 1, and use the incremental learning algorithm to dynamically update the initial life prediction model to obtain a dynamically optimized life prediction model; Step 5, phased life prediction: According to the aging characteristics of the concrete-filled steel tubular structure, divide the life prediction into the initial aging stage, the intermediate damage stage and the high-age stage, and use different machine learning models for life prediction for each stage by using the dynamically optimized life prediction model in Step 4; Step 6, prediction result integration and display: Combine the life prediction results of each stage in Step 5 with the original data and real-time monitoring data in Step 1, integrate them into the intelligent decision support platform, and display the prediction results, structural health status and maintenance suggestions through visualization tools.

2. The method according to claim 1, wherein The original data in Step 2 is integrated by the weighted average method, and the integration formula is as follows: Where: X f represents the fused feature matrix; ω i is the weight of the i-th data source, and the weight is set according to its reliability or importance; X i represents the original feature matrix of the i-th data source; The formula for denoising is as follows: Where: X denoised [i] represents the denoised data; X represents the original data; k is the size of the filtering window; j represents the index of each data point in the window; The formula for normalization is as follows: Where: X represents the original data; μ represents the mean of the data; σ represents the standard deviation; X norm represents the data after normalization; The formula for missing value filling is as follows: Where: n represents the number of non-missing data; represents the sum of all non-missing values; X filled represents the data point after filling in the missing value; X i represents the value of each data point that is not missing in the original data; The formula for the feature matrix is as follows: X features = [X1, X2, …, X n ​ Where: X1, X2, …, X n are features of different data sources extracted from sensor data, external environment data, UAV images, and satellite remote sensing data.

3. The method according to claim 1, wherein The machine learning model is a random forest model, a support vector machine model or a long short-term memory network model.

4. The life prediction method according to claim 1, characterized in that The formula for dynamically updating the life prediction model using the incremental learning algorithm in Step 4 is as follows: where: θ t represents the parameters of the model at time point t; θ t+1 represents the updated model parameters, and η is the learning rate used to control the magnitude of model update; is the gradient of the loss function with respect to the model parameters, calculated based on the new data D new obtained.

5. The life prediction method according to claim 1, wherein For the life prediction in the initial aging stage in Step 5, select a support vector machine or a random forest model as the machine learning model, and use the dynamically optimized prediction model for training, where the training features are crack width, temperature and humidity changes, and initial stress distribution; For the life prediction in the intermediate damage stage, select a long short-term memory network as the machine learning model, and use the dynamically optimized life prediction model for training, where the training features are historical crack propagation speed, stress and strain data, and temperature and humidity changes; For the life prediction in the high-age stage, select a deep neural network or a convolutional neural network as the machine learning model, and use the dynamically optimized life prediction model for training, where the training features are bearing capacity, crack distribution and material degradation.

6. The life prediction method according to claim 1, wherein The life prediction formula for the initial aging stage in Step 5 is as follows: In the formula: represents the output value predicted by the model; K(x i , x) is the kernel function; α i is the Lagrange multiplier; b is the bias term; N is the number of support vectors; x is the input feature; x i is the feature vector of a specific sample or data point in the input features; The life prediction formula for the intermediate damage stage is as follows: h t = f(W h hh t-1 + W x hx t + b h ) where: h t is the hidden state at time step t; x t is the input data at time step t; W h h, W x h is the weight matrix; b h represents the bias term of the hidden state; h t-1 is the hidden state at time step t-1, usually containing all historical information of the previous time step; b h is the bias term of the hidden state; is the predicted data; W hy represents the weight matrix from the hidden state to the output layer; b y represents the bias term of the output layer; The life prediction formula for the advanced age stage is as follows: h = f(Wx + b) Where: h is the output of the neural network; x is the input feature; W is the weight matrix; b is the bias term; f is the activation function.

7. A system for predicting the service life of concrete-filled steel tubes based on machine learning, characterized in that It includes: A data acquisition module, which is used to collect sensor data, external environment data, UAV images and satellite remote sensing data for the first time as raw data, and continuously collect sensor data, external environment data, UAV images and satellite remote sensing data as real-time monitoring data in the follow-up; among them, the sensor data includes key parameter data such as the stress, strain, temperature, humidity, crack width and corrosion condition of the concrete-filled steel tube structure, and the external environment data includes data on traffic load, climate change conditions and seismic activities; A data processing module, which is used to fuse the collected raw data, and perform denoising, normalization and missing value filling to generate a feature matrix containing structural health parameters, environmental impact factors, historical usage conditions and maintenance records; A machine learning module, which is used to train the feature matrix using a machine learning model to obtain an initial life prediction model; An online learning module, which is used to continuously receive real-time monitoring data after the initial life prediction model is trained, and use the incremental learning algorithm to dynamically update the life prediction model to obtain a dynamically optimized life prediction model; A stage-based life prediction module, which is used to divide the life prediction into the initial aging stage, the intermediate damage stage and the advanced age stage according to the aging characteristics of the concrete-filled steel tube structure, and apply different machine learning models to each stage respectively for life prediction using the dynamically optimized life prediction model; An intelligent decision support platform, which is used to combine the life prediction results of each stage with the collected raw data and the received real-time monitoring data, integrate them into the intelligent decision support platform, and display the prediction results and the structural health status through visualization tools.

Images