Deep learning-based steel structure fatigue damage prediction system and method thereof

Abstract

The invention relates to the technical field of steel structure fatigue damage prediction, in particular to a steel structure fatigue damage prediction system and method based on deep learning, and the system comprises a data collection module and a data storage module which are electrically connected with the data collection module; storing the fatigue life data of the steel structure to an external data memory; the data processing module is electrically connected with the data storage module; the data visualization module is electrically connected with the data processing module; the prediction and evaluation module is electrically connected with the data visualization module and the data storage module, and is used for receiving the prediction result sent by the data visualization module and the actual structure data sent by the data storage module; the prediction result is evaluated, and the prediction accuracy is calculated; based on prediction accuracy, parameters of a fatigue damage prediction model are dynamically adjusted, and according to the method, all modules are in close cooperation in the whole process from data acquisition, storage and processing to visualization and evaluation, so that an efficient and reliable workflow is formed.

Description

Technical Field

[0001] The present invention relates to the technical field of steel structure fatigue damage prediction, and in particular to a steel structure fatigue damage prediction system and method based on deep learning. Background Art

[0002] Steel structures are widely used in engineering fields such as buildings and bridges due to their excellent mechanical properties and ease of construction. However, long-term steel structures will inevitably face fatigue damage problems, which not only affects the service life of the structure, but is also likely to lead to catastrophic safety accidents. Therefore, accurate prediction of fatigue damage of steel structures is of great significance to ensure structural safety and optimize maintenance strategies.

[0003] The traditional steel structure fatigue damage prediction method is mainly based on Miner's linear cumulative damage theory. This method calculates the cumulative damage of the structure through the stress-life curve (SN curve) and the rain flow counting method. However, this method oversimplifies the complex process of fatigue damage and makes it difficult to consider the influence of material nonlinearity, environmental factors, load history and other aspects, resulting in the prediction results often deviating from the actual situation. In addition, the traditional method is also difficult to adapt to complex and changeable working conditions. When faced with sudden loads or extreme environments, the prediction accuracy drops significantly.

[0004] In recent years, with the development of machine learning technology, some researchers have tried to apply algorithms such as support vector machines (SVM) and artificial neural networks (ANN) to fatigue damage prediction. These methods have improved the prediction accuracy to a certain extent, but there are still many problems. For example, although the SVM method can handle nonlinear problems, its expression ability is limited and it is difficult to capture the complex characteristics of fatigue damage evolution. Although shallow neural networks have certain nonlinear fitting capabilities, their performance is often unsatisfactory when faced with high-dimensional and large-scale fatigue data.

[0005] In addition, existing intelligent prediction methods generally have the black box problem, that is, the prediction results lack interpretability. This is very likely to cause controversy in engineering practice, and engineers find it difficult to trust and adopt prediction results that lack physical meaning. At the same time, existing methods also face challenges when processing multi-source heterogeneous data, and it is difficult to effectively integrate multi-dimensional information such as structural stress, environmental parameters, and load history.

[0006] In view of the above problems, there is an urgent need for a steel structure fatigue damage prediction method that can fully utilize multi-source data, has strong nonlinear fitting capabilities, and maintains good interpretability. The present invention is an innovative solution to this technical demand. Summary of the invention

[0007] The technical problem to be solved by the present invention is how to construct a steel structure fatigue damage prediction system based on deep learning, which can achieve the organic unity of high-precision prediction, strong adaptability, and good interpretability, thereby providing reliable support for the safety assessment and maintenance decision-making of steel structures.

[0008] The present invention proposes a steel structure fatigue damage prediction system based on deep learning, comprising: Data collection module for: Collect fatigue life data of steel structures; Input the structural form, load data, and material properties into the fatigue calculation model to obtain the current real-time fatigue damage of the structure; A data storage module is electrically connected to the data collection module and is used to: Receiving the steel structure fatigue life data sent by the data collection module; storing the steel structure fatigue life data in an external data storage device; A data processing module is electrically connected to the data storage module and is used to: Using deep learning algorithms to preprocess, extract features, train and optimize the steel structure fatigue life data; Based on the deep learning algorithm, generating an optimized fatigue damage prediction model; A data visualization module is electrically connected to the data processing module and is used to: Receiving the fatigue damage prediction result sent by the data processing module; Display fatigue life prediction, fatigue status assessment and fatigue maintenance recommendations through the user interface; A prediction and evaluation module is electrically connected to the data visualization module and the data storage module, and is used to: Receive the prediction result sent by the data visualization module and the actual structure data sent by the data storage module; Evaluating the prediction results and calculating the prediction accuracy; Based on the prediction accuracy, the parameters of the fatigue damage prediction model are dynamically adjusted.

[0009] Preferably, the data collection module comprises: Fatigue calculation unit for: Based on the collected fatigue data of typical steel structures, a Miner model considering crack initiation and propagation is established; Using the Miner model, the fatigue damage evolution law of steel structure components is calculated; An input data unit is electrically connected to the fatigue calculation unit and is used for: Collect structural form data, including the number of members and connection node types; Collect load data, including temperature, wind speed, and seismic acceleration; Collect material property data and time history data; The fatigue calculation unit is further used to input the data collected by the input data unit into a fatigue calculation model to evaluate the current fatigue degree.

[0010] Preferably, the data storage module comprises: Parallel bus interface for high-speed data transmission with external data storage; A multi-level storage unit, electrically connected to the parallel bus interface, comprises: Cloud data storage, used to store large amounts of historical data; Local computer memory for storing recent data and model parameters; Mobile storage device for storing emergency backup data; The multi-level storage unit uses a hard disk, a solid-state hard disk and a flash memory as storage media to achieve hierarchical storage and fast access to data.

[0011] Preferably, the data processing module comprises: Pre-processing unit for: Clean the input data, including removing duplicates and filling missing values; Standardize the cleaned data; A feature extraction unit is electrically connected to the preprocessing unit and is used to: Use principal component analysis to reduce the dimension of preprocessed data; Use deep convolutional neural networks to extract high-level features of data; A model training unit is electrically connected to the feature extraction unit and is used to: Construct deep neural network models, including fatigue life prediction models and fatigue performance prediction models; Optimize model parameters using cross-validation techniques; An optimizer with adaptive learning rate is used to improve model training effect.

[0012] Preferably, the data visualization module comprises: User interface engine, used to generate interactive operation interface; A result display unit is electrically connected to the user interface engine and is used to: Present the forecast results in the form of charts, graphs, numbers, etc.; Displays the structural health status, including mild fatigue, moderate fatigue and severe fatigue levels; A maintenance suggestion unit is electrically connected to the result display unit and is used to: Generate fatigue maintenance recommendations based on prediction data and maintenance experience; Provide maintenance guidance; A historical data query unit is electrically connected to the user interface engine and is used to: Provide historical data query interface; Supports multi-dimensional data retrieval and analysis.

[0013] Preferably, the prediction and evaluation module comprises: Fatigue life prediction unit for: Establish a fatigue life prediction model based on deep neural network; The model contains 3 hidden layers, each containing 50 neurons; Adopt ReLU activation function and Dropout regularization technology; A fatigue performance prediction unit is connected in parallel with the fatigue life prediction unit and is used for: Construct a fatigue damage prediction model based on deep learning; This model also uses a 3-hidden layer structure with 50 neurons in each layer; Use mean squared error as the loss function; A model evaluation unit is electrically connected to the fatigue life prediction unit and the fatigue performance prediction unit, and is used to: Compare the predicted results with the actual structural data and calculate the model accuracy; Dynamically adjust model parameters or update training sample sets based on accuracy evaluation results.

[0014] Preferably, it further comprises a health monitoring module, which is electrically connected to the data collection module and is used for: Arrange multiple health monitoring sensors to collect health data of steel structures in real time; The health data includes: Structural stress-strain data; Ambient temperature and humidity data; Structural vibration acceleration data; The health monitoring module is also used to transmit the collected health data to the data collection module in real time.

[0015] Preferably, the health monitoring module further comprises: Data fusion unit for: Time synchronization of heterogeneous data from different sensors; The Kalman filter algorithm is used to fuse multi-source data; An abnormality detection unit is electrically connected to the data fusion unit and is used to: Based on the fused data, a baseline model of the structural health status is established; One-Class SVM algorithm is used to detect structural abnormalities; Wherein, when an abnormality is detected, the abnormality detection unit is also used to send an alarm signal to the data processing module.

[0016] Preferably, the system further comprises an edge computing module, which is electrically connected to the data collection module and the data processing module, and is used to: Deploy edge servers close to data sources; Perform preliminary processing on the collected raw data, including: Data compression; Preliminary feature extraction; Outlier detection; The processed data is transmitted to the cloud for further analysis and modeling.

[0017] A steel structure fatigue damage prediction method based on deep learning, using the system, includes the following steps: S1. Data collection: Collect steel structure fatigue life data, input the structure form, load data, and material properties into the fatigue calculation model to obtain the current real-time fatigue damage of the structure; S2. Data storage: The collected steel structure fatigue life data is stored in an external data storage device to achieve hierarchical storage and fast access to the data; S3. Data processing: Use deep learning algorithms to preprocess, extract features, train and optimize models on stored data to generate optimized fatigue damage prediction models; S4. Data visualization: Display fatigue life prediction, fatigue status assessment and fatigue maintenance suggestions through the user interface, providing intuitive display of prediction results; S5. Prediction evaluation: Evaluate the prediction results, calculate the prediction accuracy, and dynamically adjust the parameters of the fatigue damage prediction model based on the evaluation results; Wherein, step S3 comprises: Use principal component analysis to reduce the dimension of preprocessed data; Use deep convolutional neural networks to extract high-level features of data; Build a deep neural network model with 3 hidden layers, each containing 50 neurons; Use ReLU activation function and Dropout regularization technique to improve model performance; The model parameters are optimized using an optimizer with adaptive learning rate and cross-validation techniques.

[0018] The beneficial effects of the present invention are mainly reflected in the following aspects: The deep learning-based steel structure fatigue damage prediction system and method proposed in the present invention have achieved significant technological breakthroughs and performance improvements in many aspects through innovative system architecture design and deep learning algorithm application.

[0019] First, from a macroscopic perspective, the present invention constructs a complete, closed-loop fatigue damage prediction system. The system covers the entire process from data acquisition, storage, processing to visualization and evaluation, and the modules work closely together to form an efficient and reliable workflow. This systematic design not only improves the overall performance of the prediction, but also greatly enhances the practicality and scalability of the system.

[0020] Going deep into the core of the system, the deep learning model of the present invention demonstrates powerful nonlinear fitting and feature extraction capabilities. Through a multi-layer neural network structure, the model can automatically learn and capture the complex characteristics and potential laws in the evolution of fatigue damage in steel structures, which is difficult to achieve with traditional methods and shallow learning algorithms. Experimental results show that the prediction accuracy of the present invention is significantly higher than that of existing methods, providing a more reliable decision-making basis for engineering practice.

[0021] In terms of data processing, the present invention realizes the effective fusion of multi-source heterogeneous data. The system not only considers the structural stress and strain data, but also integrates multi-dimensional information such as environmental parameters and load history. This comprehensive data utilization strategy enables the model to understand and predict the fatigue damage process from multiple perspectives, greatly improving the comprehensiveness and reliability of the prediction.

[0022] The adaptability of the model is another outstanding advantage of the present invention. By introducing the edge computing module and the dynamic model adjustment mechanism, the system can quickly respond to changes in working conditions and maintain high-precision predictions under different load conditions and environments. This strong adaptability is of great significance to the complex and changeable service environment in actual engineering, and can effectively improve the safety and reliability of the structure.

[0023] This invention makes an innovative attempt to solve the black box problem. By combining traditional fatigue theory with advanced visualization technology, the system not only outputs prediction results, but also provides intuitive explanations and analyses. This interpretable design greatly enhances engineers' understanding and trust in the prediction results, which is conducive to the promotion and application of the method in actual engineering.

[0024] In addition, the present invention has also made considerable progress in computing efficiency and anomaly detection capabilities. By optimizing algorithms and hardware acceleration, the system can complete the processing and prediction of large-scale data in a reasonable time. At the same time, the deep learning model shows a keen ability to identify structural abnormalities, which is of great significance for early detection of potential risks and prevention of major accidents.

[0025] In general, the present invention achieves the organic unity of multiple goals such as prediction accuracy, model adaptability, and interpretability through innovative system design and advanced algorithm application. This comprehensive performance improvement not only solves many problems in the prior art, but also opens up a new research direction in the field of fatigue damage prediction of steel structures. The results of the present invention are expected to play an important role in many steel structure engineering fields such as bridges, high-rise buildings, and offshore platforms, and provide more reliable technical support for structural safety assessment, maintenance decision-making, and life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a logic block diagram of the entire system of the present invention.

[0027] Figure 2 It is a logic block diagram of the data collection module of the present invention.

[0028] Figure 3 It is a logic block diagram of the data storage module of the present invention.

[0029] Figure 4 It is a logic block diagram of the data processing module of the present invention.

[0030] Figure 5 It is a logic block diagram of the data visualization module of the present invention.

[0031] Figure 6 It is a logic block diagram of the prediction and evaluation module of the present invention.

[0032] Figure 7 It is a logic block diagram of the health monitoring module of the present invention.

[0033] Figure 8 This is a logical block diagram of the edge computing module of the present invention. DETAILED DESCRIPTION

[0034] See also Figure 1-8 The present invention provides a steel structure fatigue damage prediction system and method based on deep learning. The system includes a data collection module 1, a data storage module 2, a data processing module 3, a data visualization module 4 and a prediction and evaluation module 5. These modules work together to achieve accurate prediction and evaluation of steel structure fatigue damage.

[0035] First, the data collection module 1 is used to collect the fatigue life data of the steel structure, and input the structural form, load data, and material properties into the fatigue calculation model to obtain the current real-time fatigue damage of the structure. The data collection module 1 includes a fatigue calculation unit 11, which is used to: establish a Miner model that considers crack initiation and propagation based on the collected typical steel structure fatigue data; Preferably, the present invention adopts an improved Miner model as a fatigue calculation model, which takes into account the effects of crack initiation and propagation and can more accurately describe the fatigue damage evolution process of steel structures. Specifically, the improved Miner model can be expressed as: , in, is the cumulative fatigue damage degree, is the stress amplitude The actual number of cycles under is the stress amplitude Fatigue life under is the crack influence factor (usually ranging from 0.1 to 0.3), is the current crack length, is the critical crack length.

[0036] The data collection module 1 also includes an input data unit 12 for collecting structural form data, load data, material property data and time history data. For example, the structural form data may include the number of rods (such as 100) and the type of connection nodes (such as rigid connection or hinged connection); the load data may include temperature (such as -20°C to 60°C), wind speed (such as 0-50m / s) and earthquake acceleration (such as 0-0.5g); the material property data may include elastic modulus (such as 210GPa), yield strength (such as 345MPa) and fracture toughness (such as 100MPa·m^0.5).

[0037] Next, the data storage module 2 is electrically connected to the data collection module 1, and is used to receive the steel structure fatigue life data sent by the data collection module 1, and store these data in an external data storage device. In one embodiment of the present invention, the data storage module 2 includes a parallel bus interface 21 and a multi-level storage unit 22. The parallel bus interface 21 uses a high-speed PCI-Express interface, which can achieve a data transmission rate of up to 32GB / s, greatly improving data access efficiency.

[0038] The multi-level storage unit 22 includes a cloud data storage 221, a local computer storage 222, and a mobile storage 223. This multi-level storage structure can realize hierarchical management and fast access to data. For example, the cloud data storage 221 can use a distributed file system (such as Hadoop HDFS), with a storage capacity of up to PB level, mainly used to store historical data; the local computer storage 222 can use a solid-state hard disk with a capacity of 1-2TB, used to store recent data and model parameters; the mobile storage 223 can use a high-speed USB flash drive with a capacity of 128-256GB, used to store emergency backup data.

[0039] In addition, the data storage module 2 also adopts data compression technology, which can compress the original data to 20% - 30% of the original size, greatly saving storage space. Preferably, the present invention uses a lossless compression algorithm (such as the LZ77 algorithm) to ensure the integrity and accuracy of the data.

[0040] The data processing module 3 is electrically connected to the data storage module 2 and is used to preprocess, extract features, train and optimize models for the stored steel structure fatigue life data by using deep learning algorithms, and finally generate an optimized fatigue damage prediction model. Specifically, the data processing module 3 includes a preprocessing unit 31, a feature extraction unit 32, and a model training unit 33.

[0041] The preprocessing unit 31 first cleans the input data, including removing duplicates and filling missing values. For example, for missing values in time series data, linear interpolation can be used for filling; for missing values in categorical data, the mode can be used for filling. Then, the preprocessing unit 31 normalizes the cleaned data. Common methods include Z-score normalization and Min-Max normalization. The present invention preferably uses Z-score normalization, and its calculation formula is: , where, is the normalized data, is the original data, is the data mean, is the data standard deviation.

[0042] The feature extraction unit 32 first uses the principal component analysis (PCA) method to reduce the dimension of the preprocessed data. The basic idea of PCA is to map n-dimensional features to k dimensions (k < n). These k dimensions are new orthogonal features, also known as principal components, which are re-constructed k-dimensional features, rather than simply removing n - k dimensions from the n-dimensional features. Through PCA, the original high-dimensional data can be reduced to a suitable dimension (for example, reducing 100-dimensional data to 20 dimensions), which not only retains the main information of the data but also reduces the computational complexity.

[0043] Then, the feature extraction unit 32 uses a deep convolutional neural network (DCNN) to extract high-level features of the data. The advantage of DCNN is that it can automatically learn hierarchical features of the data and is particularly suitable for processing data with a grid topology structure (such as time series data). The DCNN structure adopted in the present invention includes 3 convolutional layers and 2 fully connected layers, where the convolutional kernel size is 3x3, the stride is 1, and the number of convolutional kernels in each layer is 32, 64, and 128 respectively.

[0044] The model training unit 33 is responsible for constructing a deep neural network model, including a fatigue life prediction model and a fatigue performance prediction model. Both models use a deep neural network structure with three hidden layers, each containing 50 neurons. The activation function uses the ReLU function, which is expressed as: , The advantage of the ReLU function is that it can effectively alleviate the gradient vanishing problem and speed up the model convergence.

[0045] The advantage of the ReLU function is that it can effectively alleviate the gradient vanishing problem and accelerate the model convergence speed. In addition, the present invention also adopts the Dropout regularization technology, and the dropout rate is set to 0.5, which helps to prevent the model from overfitting and improve the generalization ability.

[0046] During the model training process, the model training unit 33 uses an adaptive learning rate optimizer (such as the Adam optimizer) and cross-validation technology to optimize the model parameters. The initial value of the learning rate of the Adam optimizer is set to 0.001, and the learning rate will gradually decay as the number of training rounds increases. The cross-validation adopts a 5-fold cross-validation method, that is, the data set is divided into 5 parts, 4 parts are used as training sets each time, and 1 part is used as a validation set, and the cycle is repeated 5 times, and finally the average performance is taken as the evaluation indicator of the model.

[0047] Through the above detailed description, it can be seen that the system of the present invention adopts advanced technologies and algorithms in data acquisition, storage and processing, laying a solid foundation for subsequent fatigue damage prediction. Compared with traditional methods, this method based on deep learning has stronger nonlinear fitting ability and adaptive ability, and can better capture the complex evolution law of fatigue damage of steel structures, thereby improving the accuracy and reliability of prediction. Next, the system of the present invention also includes a data visualization module 4, which is electrically connected to the data processing module 3, and is used to receive the fatigue damage prediction results sent by the data processing module 3, and display fatigue life prediction, fatigue state assessment and fatigue maintenance recommendations through the user interface. The design of the data visualization module 4 is intended to provide users with an intuitive and easy-to-understand display of prediction results, thereby assisting decision making.

[0048] Specifically, the data visualization module 4 includes a user interface engine 41, a result display unit 42, a maintenance suggestion unit 43, and a historical data query unit 44. The user interface engine 41 is responsible for generating an interactive operation interface, which adopts a responsive design and can be adapted to display devices of different sizes, such as desktops, tablet computers, and smart phones. In a preferred embodiment of the present invention, the user interface adopts the Material Design design language, which has a clear sense of hierarchy and a good user experience.

[0049] The result display unit 42 is electrically connected to the user interface engine 41 and is responsible for displaying the prediction results in the form of charts, graphs, numbers, etc. For example, a line chart can be used to display the fatigue damage evolution trend of the steel structure in the future, a heat map can be used to display the fatigue damage degree of different parts of the structure, or a dashboard can be used to intuitively display the overall health status of the current structure. Preferably, the present invention uses the D3.js library to achieve these complex visualization effects, which has powerful data-driven capabilities and rich chart types.

[0050] In addition, the result display unit 42 is also responsible for displaying the structural health status, which is divided into three levels: mild fatigue, moderate fatigue and severe fatigue. The system of the present invention determines the fatigue level according to the value of the cumulative fatigue damage degree D: when 0 ≤ D< 0.3, it is mild fatigue, 0.3 ≤ D < 0.7 is moderate fatigue, and D ≥ 0.7 is severe fatigue. This classification method is based on engineering experience and a large amount of experimental data, and can provide an intuitive reference for structural maintenance.

[0051] The maintenance suggestion unit 43 is electrically connected to the result display unit 42, and generates fatigue maintenance suggestions and provides maintenance guidance based on the prediction data and maintenance experience. In one embodiment of the present invention, the maintenance suggestion unit 43 uses a rule-based expert system to generate maintenance suggestions. The system includes a knowledge base that stores a large number of if-then rules, such as: ``` IF (Fatigue level = Mild) AND (Critical part stress < Critical value) THEN Recommendation = Continue monitoring, no maintenance required for the time being IF (Fatigue level = Moderate) AND (Crack length > Warning value) THEN Recommendation = Non-destructive testing is recommended and local repair is recommended if necessary IF (Fatigue level = severe) OR (Accumulated fatigue damage > 0.9) THEN Recommendation = Immediate overhaul and replacement of damaged components if necessary ``` These rules are based on a wealth of engineering practice experience and expert knowledge, and can provide reasonable suggestions for structural maintenance in different situations.

[0052] The historical data query unit 44 is electrically connected to the user interface engine 41, provides a historical data query interface, and supports multi-dimensional data retrieval and analysis. Users can query historical data based on multiple dimensions such as time, location, fatigue level, etc. Preferably, the present invention uses ElasticSearch as a full-text search engine, combined with Kibana to achieve complex data visualization and analysis functions. This not only facilitates users to trace back historical data, but also provides a basis for subsequent data mining and model optimization.

[0053] The system of the present invention also includes a prediction and evaluation module 5, which is electrically connected to the data visualization module 4 and the data storage module 2, and is used to receive the prediction results sent by the data visualization module 4 and the actual structural data sent by the data storage module 2, evaluate the prediction results, calculate the prediction accuracy, and dynamically adjust the parameters of the fatigue damage prediction model based on the prediction accuracy.

[0054] Specifically, the prediction and evaluation module 5 includes a fatigue life prediction unit 51, a fatigue performance prediction unit 52 and a model evaluation unit 53. The fatigue life prediction unit 51 is responsible for establishing a fatigue life prediction model based on a deep neural network. The model includes 3 hidden layers, each layer includes 50 neurons. Preferably, the present invention uses ReLU activation function and Dropout regularization technology to improve model performance. The expression of ReLU activation function is: , The principle of Dropout regularization is to randomly discard a portion of neurons during the training process to prevent the model from overfitting. In the present invention, the Dropout rate is set to 0.5, that is, in each training iteration, each neuron has a 50% probability of being temporarily removed from the network.

[0055] The fatigue performance prediction unit 52 is connected in parallel with the fatigue life prediction unit 51 to construct a fatigue damage prediction model based on deep learning. The structure of this model is similar to that of the fatigue life prediction model, and it also uses 3 hidden layers with 50 neurons in each layer. The difference is that this model uses the mean square error (MSE) as the loss function, and its expression is: , in, is the sample size, is the true value, is the predicted value. MSE can effectively measure the difference between the predicted value and the true value and is a commonly used loss function in regression problems.

[0056] The model evaluation unit 53 is electrically connected to the fatigue life prediction unit 51 and the fatigue performance prediction unit 52, and is responsible for comparing the prediction results with the actual structural data and calculating the model accuracy. The present invention uses the mean absolute percentage error (MAPE) as an evaluation index, and its calculation formula is: , The advantage of MAPE is that it can eliminate the impact of dimension and facilitate the comparison of data of different scales. Generally speaking, a MAPE less than 10% indicates a good prediction effect, a MAPE between 10% and 20% indicates a general prediction effect, and a MAPE greater than 20% indicates a poor prediction effect.

[0057] Based on the calculation results of MAPE, the model evaluation unit 53 will dynamically adjust the model parameters or update the training sample set. For example, when MAPE is greater than 15%, the system will automatically increase the number of training rounds or randomly extract new samples from historical data to add to the training set to improve the prediction accuracy of the model.

[0058] In addition, the system of the present invention also includes a health monitoring module 6, which is electrically connected to the data collection module 1 and is used to arrange multiple health monitoring sensors to collect health data of the steel structure in real time. These health data include structural stress and strain data, environmental temperature and humidity data, and structural vibration acceleration data.

[0059] Preferably, the present invention uses a fiber grating sensor (FBG) to monitor structural stress and strain. The sensor has the advantages of high precision (typical accuracy of 1με), anti-electromagnetic interference, and can be buried in the structure for a long time. The environmental temperature and humidity data are collected by a digital temperature and humidity sensor (such as SHT30), whose temperature measurement range is -40°C to 125°C and the accuracy is ±0.2°C; the relative humidity measurement range is 0% to 100%RH, and the accuracy is ±2%RH. The structural vibration acceleration data is collected by a MEMS acceleration sensor (such as ADXL355), which has a measurement range of ±2g / ±4g / ±8g optional, and a resolution of up to 1.0 mg / LSB.

[0060] The health monitoring module 6 is also responsible for transmitting the collected health data to the data collection module 1 in real time. The data transmission adopts wireless sensor network technology, specifically using the ZigBee protocol, which has the characteristics of low power consumption, high reliability, and support for large-scale networks, and is very suitable for the application scenario of structural health monitoring.

[0061] The health monitoring module 6 is electrically connected to the data collection module 1 and is used to arrange a plurality of health monitoring sensors to collect health data of the steel structure in real time. Preferably, the health monitoring module 6 of the present invention further includes: The data fusion unit 61 is used for: Time synchronization of heterogeneous data from different sensors; The Kalman filter algorithm is used to fuse multi-source data.

[0062] Specifically, the data fusion unit 61 first uses the timestamp information to align the data collected by different sensors to ensure the time consistency of the data. Then, the multi-source data is fused through the Kalman filter algorithm. The core idea of ​​the Kalman filter is to continuously optimize the state estimation through two steps of prediction and update.

[0063] The abnormality detection unit 62 is electrically connected to the data fusion unit 61 and is used to: Based on the fused data, a baseline model of the structural health status is established; One-Class SVM algorithm is used to detect structural abnormalities.

[0064] The anomaly detection unit 62 first uses the data under normal working conditions to establish a baseline model of the structural health status. Then, the One-Class SVM algorithm is used for anomaly detection. The goal of the One-Class SVM is to find a hyperplane in the feature space so that most normal samples are located on one side of the hyperplane, while the abnormal samples are located on the other side. Its decision function can be expressed as: , in, is the input sample, is the support vector, is the Lagrange multiplier, is the kernel function, is the bias term. When it is judged as abnormal state, is the symbolic function, represents the number of support vectors.

[0065] Through the above detailed description, it can be seen that the system of the present invention adopts advanced technologies and methods in data visualization, predictive evaluation and health monitoring. This comprehensive and systematic design not only improves the accuracy of fatigue damage prediction, but also greatly enhances the practicality and reliability of the system, providing strong support for the safety assessment and maintenance decision-making of steel structures. The system of the present invention also includes an edge computing module 7, which is electrically connected to the data collection module 1 and the data processing module 3, and is used to deploy an edge server at a location close to the data source, perform preliminary processing on the collected raw data, and transmit the processed data to the cloud for further analysis and modeling.

[0066] The introduction of the edge computing module 7 is an important innovation of the present invention, which can effectively solve the problems of high data transmission delay, high bandwidth consumption, and difficulty in privacy protection in the traditional cloud computing architecture. In a preferred embodiment of the present invention, the edge computing module 7 is developed using Intel's OpenVINO™ toolkit, which is optimized for edge computing and can significantly improve the inference speed of deep learning models.

[0067] The edge computing module 7 is electrically connected to the data collection module 1 and the data processing module 3, and is used to deploy an edge server near the data source, perform preliminary processing on the collected raw data, and transmit the processed data to the cloud for further analysis and modeling. Specifically, the edge computing module 7 includes: The data compression unit 71 is used to compress the collected raw data. The present invention preferably adopts the LZ4 lossless compression algorithm, which has extremely fast compression and decompression speeds and a compression ratio of 2:1 to 5:1. The core idea of ​​the LZ4 algorithm is to use the repetitive patterns in the data to replace the repetitive long character strings with short codes. The compression process can be described as follows: 1. Find a match: Find the longest string in the processed data that matches the current data block.

[0068] 2. Encoding: If a match is found, it is encoded as a (distance, length) pair; otherwise, the original character is output directly.

[0069] Repeat steps 1 and 2 until all data has been processed.

[0070] The preliminary feature extraction unit 72 is electrically connected to the data compression unit 71 and is used to perform preliminary feature extraction on the compressed data. The present invention uses a lightweight convolutional neural network MobileNetV2 for feature extraction. The core of MobileNetV2 is deep separable convolution, which includes two steps: deep convolution and point-by-point convolution: 1. Depth convolution: Perform convolution operation on each channel of the input separately.

[0071] 2. Point-wise convolution: Linearly combine the outputs of depthwise convolution using a 1x1 convolution kernel.

[0072] This design greatly reduces the amount of computation and the number of parameters, allowing the model to run efficiently on edge devices.

[0073] The outlier detection unit 73 is electrically connected to the preliminary feature extraction unit 72 and is used to perform outlier detection on the extracted features. The present invention uses an isolation forest algorithm for outlier detection. The core idea of ​​the algorithm is that outliers are more easily isolated. The specific steps are as follows: 1. Randomly select a feature.

[0074] 2. Randomly select a split point between the maximum and minimum values ​​of the feature.

[0075] 3. Divide the data into two parts.

[0076] Repeat steps 1-3 until the specimen is individually isolated or the specified tree height is reached.

[0077] The outlier detection unit 73 is also responsible for transmitting the processed data to the data processing module 3 for further analysis and modeling.

[0078] In general, the edge computing module 7 mainly performs the following three tasks: data compression, preliminary feature extraction, and outlier detection. Data compression uses the lossless compression algorithm LZ4, which has extremely fast compression and decompression speeds, and a compression ratio of 2:1 to 5:1, which is very suitable for real-time data stream processing. Preliminary feature extraction uses the lightweight convolutional neural network MobileNetV2, which has a small number of network model parameters (about 3.4M) but excellent performance, and can achieve real-time feature extraction on edge devices.

[0079] For outlier detection, the present invention adopts a method based on the Isolation Forest algorithm. The core idea of ​​the algorithm is that outliers are easier to isolate. Specifically, the algorithm randomly selects a feature and randomly selects a split point between its maximum and minimum values ​​to divide the data into two parts. This process is repeated until the sample is isolated individually or the specified tree height is reached. Outliers are often easier to isolate, so fewer segmentation steps are required. The mathematical expression of outlier detection is as follows: , in, It is a sample The anomaly score, It is a sample The average path length, is given The average path length when samples are Near 1 o'clock, is likely to be an outlier; when When it is much smaller than 0.5, Most likely normal.

[0080] Through the processing of the edge computing module 7, the amount of original data can be reduced by 70% to 80%, greatly reducing the pressure of data transmission and storage, while also improving the efficiency of subsequent analysis and modeling.

[0081] Finally, the present invention also proposes a steel structure fatigue damage prediction method based on deep learning, which is closely integrated with the above system and includes the following steps: S1. Data collection: Collect steel structure fatigue life data, input the structure form, load data, and material properties into the fatigue calculation model to obtain the current real-time fatigue damage of the structure. In this step, the fatigue calculation model uses the improved Miner model, which takes into account the influence of crack initiation and propagation. The mathematical expression of the model is as follows: , in, is the cumulative fatigue damage degree, is the stress amplitude The actual number of cycles under is the stress amplitude Fatigue life under is the crack influence factor (the value range is 0.1~0.3), is the current crack length, is the critical crack length.

[0082] S2. Data storage: The collected steel structure fatigue life data is stored in an external data storage device to achieve hierarchical storage and fast access to data. The method of the present invention adopts a multi-level storage structure, including a cloud data storage device, a local computer storage device, and a mobile storage device. Preferably, the cloud storage uses the Apache Hadoop Distributed File System (HDFS), which can store and process PB-level data; the local storage uses a solid-state drive (SSD), with a read and write speed of more than 500MB / s; the mobile storage uses a high-speed USB flash drive with a USB 3.1 interface, with a transmission rate of up to 10Gbps.

[0083] S3. Data processing: Use deep learning algorithms to preprocess, extract features, train and tune models for stored data to generate optimized fatigue damage prediction models. This step is the core of the entire method and includes the following key links: First, the principal component analysis (PCA) is used to reduce the dimension of the preprocessed data. The mathematical principle of PCA can be expressed as: , in, is the covariance matrix, is the number of samples, is the sample vector, is the sample mean vector, is the eigenvector matrix, is a diagonal matrix of eigenvalues.

[0084] Then, a deep convolutional neural network (DCNN) is used to extract high-level features of the data. The convolution operation of DCNN can be expressed as: , in, is the input signal, is the convolution kernel, Represents a convolution operation.

[0085] Finally, a deep neural network model with 3 hidden layers is constructed, with 50 neurons in each layer. The model uses the ReLU activation function and Dropout regularization technology, and the mean square error (MSE) is used as the loss function. The forward propagation process of the model can be expressed as: , in, and Respectively The weight matrix and bias vector of the layer, is the ReLU activation function.

[0086] S4. Data visualization: The fatigue life prediction, fatigue state assessment and fatigue maintenance suggestions are displayed through the user interface, providing an intuitive display of the prediction results. The method of the present invention uses the D3.js library to achieve complex data visualization effects, such as using a force-directed graph to display the fatigue state of each part of the structure and its mutual influence. The core algorithm can be expressed as: , in, For Node and The force between and are the elastic coefficient and repulsion coefficient, and and The distance between For the ideal distance, For slave nodes Point to Node The unit vector of .

[0087] S5. Prediction evaluation: Evaluate the prediction results, calculate the prediction accuracy, and dynamically adjust the parameters of the fatigue damage prediction model based on the evaluation results. The method of the present invention uses the mean absolute percentage error (MAPE) as an evaluation indicator, and its calculation formula is: , in, is the sample size, is the true value, is the predicted value.

[0088] Through the above steps, the method of the present invention achieves accurate prediction and evaluation of fatigue damage of steel structures. This method combines the powerful nonlinear fitting ability of deep learning with the physical basis of traditional fatigue theory, which not only improves the accuracy of prediction, but also enhances the interpretability of the model. In addition, the data visualization and prediction and evaluation steps in the method provide intuitive and reliable decision support for engineering practice, which has important application value.

[0089] In order to verify the superiority of the steel structure fatigue damage prediction system and method based on deep learning of the present invention, a set of experiments were designed, including an embodiment and two comparative examples. These experiments are intended to simulate the steel structure fatigue damage prediction scenario in actual engineering and evaluate the performance of the present invention from multiple perspectives.

[0090] Example 1 The system and method of the present invention are used to predict fatigue damage of a large steel bridge for one year. The steel bridge has a span of 200 meters, the main beam adopts a box-type section, and the average daily traffic volume is about 50,000 vehicles. The system collects data such as structural stress and strain, ambient temperature and humidity, and vibration acceleration in real time through the arranged health monitoring sensors, and combines historical data for deep learning model training and prediction.

[0091] Comparative Example 1 uses the traditional Miner linear cumulative damage theory to predict fatigue life. This method is based on the stress-life curve (SN curve) and the rain flow counting method to calculate the cumulative damage of the structure.

[0092] Comparative Example 2 uses the support vector regression (SVR) method for fatigue damage prediction. SVR is a commonly used machine learning algorithm that can handle nonlinear problems, but its expressiveness and flexibility are not as good as deep learning models.

[0093] The following five indicators were selected to evaluate the performance of each method: prediction accuracy, computational efficiency, model adaptability, anomaly detection capability, and interpretability. The detection standards and methods of these indicators are as follows: 1. Prediction accuracy: The mean absolute percentage error (MAPE) is used to evaluate the prediction accuracy, which is calculated by comparing the predicted value with the measured value. The detection method is to randomly select 100 time points for comparison during the prediction period.

[0094] 2. Computational efficiency: record the average time required to complete a full bridge fatigue damage prediction. The detection method is to repeat the prediction process 10 times and take the average value.

[0095] 3. Model adaptability: Evaluate the model's adaptability to different load conditions. The detection method is to make predictions under normal load, 50% overload and extreme weather (strong wind) conditions, and calculate the average MAPE.

[0096] 4. Anomaly detection capability: evaluates the model’s ability to identify abnormal structural conditions. The detection method is to artificially introduce five tiny cracks into the structure and calculate the proportion of successful detections.

[0097] 5. Explainability: Evaluate the model's ability to explain the prediction results. Qualitative scoring is used, and 5 domain experts score (1-10 points) based on the explanatory information output by the model, and the average is taken.

[0098] The experimental results are shown in Table 1: Table 1. Comparison of experimental results of Example 1, Comparative Example 1 and Comparative Example 2 Performance Indicators Example 1 Comparative Example 1 Comparative Example 2 Prediction accuracy 93.2% 78.5% 85.7% Computational efficiency 2.3 minutes 0.5 minutes 1.8 minutes Model adaptability 91.5% 72.3% 83.1% Anomaly detection capabilities 4 / 5 1 / 5 3 / 5 Explainability 8.2 points 9.5 points 6.8 points , It can be seen from the experimental results in Table 1 that the method of the present invention has shown obvious advantages in multiple key indicators: In terms of prediction accuracy, the method of the present invention reached 93.2%, which is much higher than the traditional Miner theory (78.5%) and SVR method (85.7%). This shows that the deep learning model can better capture the complex nonlinear characteristics of fatigue damage of steel structures, thereby providing more accurate prediction results.

[0099] In terms of computational efficiency, although the method of the present invention (2.3 minutes) is not as good as the simple Miner theory (0.5 minutes), this computational time is acceptable considering the significant improvement in prediction accuracy. Moreover, compared with the SVR method (1.8 minutes), the method of the present invention shows better scalability when processing large-scale data.

[0100] In terms of model adaptability, the method of the present invention performs well (91.5%), far exceeding the other two methods. This proves the powerful generalization ability of deep learning models, which can better adapt to different load conditions and environmental changes, which is extremely important for practical engineering applications.

[0101] In terms of anomaly detection capability, the method of the present invention successfully detected 4 out of 5 artificially introduced micro cracks, which is better than the other two methods. This is due to the multi-source data fusion and deep feature extraction technology used in the present invention, which can capture subtle structural anomalies.

[0102] In terms of interpretability, although the traditional Miner theory scored the highest (9.5 points), the method of the present invention also reached a very high level (8.2 points). This shows that through reasonable model design and visualization technology, deep learning methods can also provide good interpretability, which helps engineers understand and trust the prediction results.

[0103] In summary, the steel structure fatigue damage prediction system and method based on deep learning of the present invention have shown significant advantages in key indicators such as prediction accuracy, model adaptability and anomaly detection capability. Although it is slightly insufficient in computational efficiency, considering the performance improvement it brings, this trade-off is worthwhile. In addition, the method of the present invention also maintains good interpretability, which is crucial for decision support in engineering practice.

[0104] These experimental results fully demonstrate the innovation and practicality of the present invention in the field of fatigue damage prediction of steel structures. It not only improves the accuracy and reliability of prediction, but also enhances the adaptability of the model to complex working conditions, providing more reliable technical support for safety assessment and maintenance decision-making of steel structures. In the future, with the further development and optimization of deep learning technology, the method of the present invention is expected to play an important role in a wider range of structural health monitoring fields.

[0105] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A steel structure fatigue damage prediction system based on deep learning, characterized in that: include: Data collection module for: Collect fatigue life data of steel structures; Input the structural form, load data, and material properties into the fatigue calculation model to obtain the current real-time fatigue damage of the structure; A data storage module is electrically connected to the data collection module and is used to: Receiving the steel structure fatigue life data sent by the data collection module; storing the steel structure fatigue life data in an external data storage device; A data processing module is electrically connected to the data storage module and is used to: Using deep learning algorithms to preprocess, extract features, train and optimize the steel structure fatigue life data; Based on the deep learning algorithm, generating an optimized fatigue damage prediction model; A data visualization module is electrically connected to the data processing module and is used to: Receiving the fatigue damage prediction result sent by the data processing module; Display fatigue life prediction, fatigue status assessment and fatigue maintenance recommendations through the user interface; A prediction and evaluation module is electrically connected to the data visualization module and the data storage module, and is used to: Receive the prediction result sent by the data visualization module and the actual structure data sent by the data storage module; Evaluating the prediction results and calculating the prediction accuracy; Based on the prediction accuracy, the parameters of the fatigue damage prediction model are dynamically adjusted.

2. The system according to claim 1, characterized in that The data collection module includes: Fatigue calculation unit for: Based on the collected fatigue data of typical steel structures, a Miner model considering crack initiation and propagation is established; Using the Miner model, the fatigue damage evolution law of steel structure components is calculated; An input data unit is electrically connected to the fatigue calculation unit and is used for: Collect structural form data, including the number of members and connection node types; Collect load data, including temperature, wind speed, and seismic acceleration; Collect material property data and time history data; The fatigue calculation unit is further used to input the data collected by the input data unit into a fatigue calculation model to evaluate the current fatigue degree.

3. The system according to claim 1, characterized in that The data storage module comprises: Parallel bus interface for high-speed data transmission with external data storage; A multi-level storage unit, electrically connected to the parallel bus interface, comprises: Cloud data storage, used to store large amounts of historical data; Local computer memory for storing recent data and model parameters; Mobile storage device for storing emergency backup data; The multi-level storage unit uses a hard disk, a solid-state hard disk and a flash memory as storage media to achieve hierarchical storage and fast access to data.

4. The system according to claim 1, characterized in that The data processing module comprises: Pre-processing unit for: Clean the input data, including removing duplicates and filling missing values; Standardize the cleaned data; A feature extraction unit is electrically connected to the preprocessing unit and is used to: Use principal component analysis to reduce the dimension of preprocessed data; Use deep convolutional neural networks to extract high-level features of data; A model training unit is electrically connected to the feature extraction unit and is used to: Construct deep neural network models, including fatigue life prediction models and fatigue performance prediction models; Optimize model parameters using cross-validation techniques; An optimizer with adaptive learning rate is used to improve model training effect.

5. The system according to claim 1, characterized in that The data visualization module includes: User interface engine, used to generate interactive operation interface; A result display unit is electrically connected to the user interface engine and is used to: Present the forecast results in charts, graphs, and numbers; Displays the structural health status, including mild fatigue, moderate fatigue and severe fatigue levels; A maintenance suggestion unit is electrically connected to the result display unit and is used to: Generate fatigue maintenance recommendations based on prediction data and maintenance experience; Provide maintenance guidance; A historical data query unit is electrically connected to the user interface engine and is used to: Provide historical data query interface; Supports multi-dimensional data retrieval and analysis.

6. The system according to claim 1, characterized in that The prediction and evaluation module comprises: Fatigue life prediction unit for: Establish a fatigue life prediction model based on deep neural network; The model contains 3 hidden layers, each containing 50 neurons; Adopt ReLU activation function and Dropout regularization technology; A fatigue performance prediction unit is connected in parallel with the fatigue life prediction unit and is used for: Construct a fatigue damage prediction model based on deep learning; This model also uses a 3-hidden layer structure with 50 neurons in each layer; Use mean squared error as the loss function; A model evaluation unit is electrically connected to the fatigue life prediction unit and the fatigue performance prediction unit, and is used to: Compare the predicted results with the actual structural data and calculate the model accuracy; Dynamically adjust model parameters or update training sample sets based on accuracy evaluation results.

7. The system according to claim 1, characterized in that It also includes a health monitoring module, which is electrically connected to the data collection module and is used to: Arrange multiple health monitoring sensors to collect health data of steel structures in real time; The health data includes: Structural stress-strain data; Ambient temperature and humidity data; Structural vibration acceleration data; The health monitoring module is also used to transmit the collected health data to the data collection module in real time.

8. The system according to claim 7, characterized in that The health monitoring module also includes: Data fusion unit for: Time synchronization of heterogeneous data from different sensors; The Kalman filter algorithm is used to fuse multi-source data; An abnormality detection unit is electrically connected to the data fusion unit and is used to: Based on the fused data, a baseline model of the structural health status is established; One-Class SVM algorithm is used to detect structural abnormalities; Wherein, when an abnormality is detected, the abnormality detection unit is also used to send an alarm signal to the data processing module.

9. The system according to claim 1, characterized in that The system further includes an edge computing module, which is electrically connected to the data collection module and the data processing module and is used to: Deploy edge servers close to data sources; Perform preliminary processing on the collected raw data, including: Data compression; Preliminary feature extraction; Outlier detection; The processed data is transmitted to the cloud for further analysis and modeling.

10. A steel structure fatigue damage prediction method based on deep learning, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Data collection: Collect steel structure fatigue life data, input the structure form, load data, and material properties into the fatigue calculation model to obtain the current real-time fatigue damage of the structure; S2. Data storage: The collected steel structure fatigue life data is stored in an external data storage device to achieve hierarchical storage and fast access to the data; S3. Data processing: Use deep learning algorithms to preprocess, extract features, train and optimize models on stored data to generate optimized fatigue damage prediction models; S4. Data visualization: Display fatigue life prediction, fatigue status assessment and fatigue maintenance suggestions through the user interface, providing intuitive display of prediction results; S5. Prediction evaluation: Evaluate the prediction results, calculate the prediction accuracy, and dynamically adjust the parameters of the fatigue damage prediction model based on the evaluation results; Wherein, step S3 comprises: Use principal component analysis to reduce the dimension of preprocessed data; Use deep convolutional neural networks to extract high-level features of data; Build a deep neural network model with 3 hidden layers, each containing 50 neurons; Use ReLU activation function and Dropout regularization technique to improve model performance; The model parameters are optimized using an optimizer with adaptive learning rate and cross-validation techniques.

Images