Artificial intelligence-based coupled corrosion fatigue test method and system for bridge structures

Abstract

The present invention relates to the technical field of bridge engineering material testing, and discloses a bridge structure corrosion fatigue coupling test method and system based on artificial intelligence; wherein, a bridge structure corrosion fatigue coupling test method based on artificial intelligence includes: accurately expressing the interaction between physical fields by constructing a multi-physical field coupling mathematical model; applying a deep reinforcement learning algorithm to build an environmental parameter collaborative control model to achieve high-precision adaptive regulation of multiple environmental parameters; building an environmental acceleration test equivalence mapping model to balance the optimal acceleration ratio and equivalence of the acceleration test; constructing a generalized predictive control model to achieve multi-physical field collaborative regulation; constructing a real-time system identification and parameter update model to compensate for the interaction between multiple physical fields and achieve stable control; the present invention improves the test environment control accuracy, realizes the collaborative control of multiple environmental conditions, shortens the test cycle, improves the test authenticity and result stability, and provides reliable data support for bridge structure safety assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge engineering material testing, and in particular to an artificial intelligence-based bridge structure corrosion fatigue coupling test method and system. Background Art

[0002] Bridge structures are subject to the combined effects of corrosion and fatigue loading during service. The corrosion-fatigue coupling effect is a significant factor affecting bridge structural safety. Accurately understanding the evolution of corrosion-fatigue coupled damage is crucial for life prediction and safety assessment of bridge structures, and conducting corrosion-fatigue coupling tests is the primary means of obtaining relevant data.

[0003] However, the existing corrosion fatigue coupling test methods have several key technical difficulties: during the service process, bridge structures are affected by multiple physical fields such as mechanical fields, electrochemical fields, and temperature fields. These physical fields interact and influence each other, forming complex coupling effects. Traditional test methods are difficult to accurately simulate this complex interactive relationship; the development of corrosion and fatigue damage in actual service often takes years or even decades, and traditional test methods are difficult to complete effective accelerated tests within a reasonable period; traditional test methods have low control accuracy and poor stability, and it is difficult to restore the complex and changeable actual service environment conditions; traditional test methods usually control environmental conditions and load conditions separately, and cannot reflect the dynamic coupling relationship between the two, resulting in insufficient test authenticity.

[0004] With the rapid development of artificial intelligence (AI) technology, the application of advanced AI algorithms and control techniques to corrosion fatigue coupled testing is expected to overcome these technical bottlenecks and improve the accuracy, efficiency, and reliability of testing. However, there is currently a lack of systematic and effective solutions that integrate key technologies such as multi-physics modeling, coordinated control of environmental parameters, and accelerated test optimization to achieve comprehensive optimization of bridge structure corrosion fatigue coupled testing. Summary of the Invention

[0005] The present invention discloses an artificial intelligence-based coupled test method and system for bridge structure corrosion fatigue, aiming to solve problems existing in the prior art, such as the interaction of multiple physical fields, excessively long test cycles, insufficient environmental control accuracy, and the separation of environmental conditions from load conditions.

[0006] The present invention provides an artificial intelligence-based bridge structure corrosion fatigue coupling test method, comprising the following steps:

[0007] Construct a multi-physics field coupling mathematical model including mechanical field, electrochemical field and temperature field to accurately express the interaction between physical fields;

[0008] Apply deep reinforcement learning algorithms to build a collaborative control model for environmental parameters, achieving high-precision control of multiple environmental parameters and adaptive regulation based on material response;

[0009] Construct an environmental accelerated test equivalence mapping model, optimize the accelerated test parameter combination through a multi-objective optimization algorithm, and achieve a balance between the optimal acceleration ratio and equivalence;

[0010] Build a generalized predictive control model and achieve multi-physics coordinated control through a rolling optimization strategy to improve control accuracy and response speed;

[0011] Build a real-time system identification and parameter update model, realize online identification of system parameters based on the recursive least squares algorithm, and actively compensate and stabilize the interaction effects of multiple physical fields.

[0012] Furthermore, the multi-physical field coupling mathematical model is established through a group of partial differential equations, which includes the state variables of each physical field, the time derivatives of the state variables, the spatial gradients of the state variables, the interaction functions between the physical fields, and the influence functions of the control inputs on the physical fields, and is used to characterize the coupling relationship between different physical fields.

[0013] Furthermore, the deep reinforcement learning environment parameter collaborative control model adopts a deep Q network to realize strategy learning in complex environments, and realizes the learning of the optimal control strategy by iteratively updating the Q-value function, where the Q-value function represents the expected cumulative reward value of performing a specific action in a specific state.

[0014] Furthermore, the reward function of the deep reinforcement learning environment parameter collaborative control model is designed as a weighted combination of environment parameter control accuracy and material response compliance, where the environment parameter control accuracy is calculated by the deviation between the actual value and the target value, and the material response compliance is quantified by a special evaluation function.

[0015] Furthermore, the environmental acceleration test equivalence mapping model establishes a mapping relationship between the accelerated test environment parameters and the actual service environment parameters, and introduces a time compression ratio and a time-varying acceleration factor to achieve accurate setting of the accelerated test conditions.

[0016] Furthermore, the acceleration test parameter combination optimization adopts a multi-objective optimization algorithm to find the optimal acceleration test parameter combination by balancing the two objectives of time compression ratio and equivalence deviation, where the equivalence deviation is evaluated by a special measurement function to evaluate the degree of difference between the acceleration test environment and the actual service environment.

[0017] Furthermore, the generalized predictive control model constructs a weighted cost function that includes output tracking error and control input change, and performs rolling optimization in the prediction domain and the control domain to achieve coordinated regulation of multiple physical fields.

[0018] Furthermore, the real-time system identification adopts a recursive least squares algorithm to achieve online identification of the system model by continuously updating parameter estimates, wherein the parameter update process uses the Kalman gain to perform weighted correction on the error between the system output and the predicted output.

[0019] Furthermore, the active compensation method for the interaction of multiple physical fields is implemented using a feedforward-feedback composite structure. An interaction decoupling matrix is constructed based on a multi-physical field coupling model obtained based on real-time system identification, and the compensation control input is calculated according to the change in the target physical field state, thereby achieving effective compensation for the interaction of physical fields.

[0020] Furthermore, an artificial intelligence-based bridge structure corrosion fatigue coupling test system is used to perform the above-mentioned artificial intelligence-based bridge structure corrosion fatigue coupling test method, including:

[0021] Multi-physics coupling model module, used to build coupled mathematical models including mechanical fields, electrochemical fields, and temperature fields;

[0022] A deep reinforcement learning control module for coordinated control of environmental parameters and adaptive regulation based on material response;

[0023] Accelerated test mapping module, used to establish an equivalent mathematical model between environmental accelerated test parameters and actual service environment;

[0024] Generalized predictive control module, used to achieve multi-physics field coordinated regulation and rolling optimization control;

[0025] A real-time system identification module that continuously updates coupled model parameters and actively compensates for multi-physics interactions.

[0026] The beneficial effects of the present invention are:

[0027] This invention significantly improves the performance of the test system by leveraging a deep reinforcement learning environment parameter collaborative control strategy and a generalized predictive control algorithm. In terms of control accuracy, the invention's environmental parameter fluctuation control surpasses traditional methods, improving data reliability. By constructing a multi-physics field coupling mathematical model, it achieves arbitrary combinations and smooth transitions of multiple environmental conditions, closely matching the actual service environment of the bridge. Relying on an environmental accelerated test equivalence mapping model, while ensuring data reliability, it shortens the test cycle and reduces test costs. By utilizing multi-physics field collaborative control and real-time system identification, the consistency between the test and actual bridge service conditions is improved. Through real-time system identification and parameter updates, the interaction effects of multiple physical fields are actively compensated, reducing fluctuations in multiple rounds of test data and improving the comparability of results between different laboratories, providing more reliable data support for bridge structure safety assessments. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The present invention is a flow chart of an artificial intelligence-based bridge structure corrosion fatigue coupling test method. DETAILED DESCRIPTION

[0029] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0030] At least one embodiment of the present invention discloses a bridge structure corrosion fatigue coupling test method based on artificial intelligence, such as Figure 1 As shown, the following steps are included:

[0031] Step 1: Construct a multi-physics field coupling mathematical model including mechanical field, electrochemical field and temperature field to accurately express the interaction between physical fields;

[0032] Specifically include:

[0033] Step 1.1, collect response data of the bridge structure under different physical fields;

[0034] Collect response data of bridge structures under the separate and combined effects of mechanical fields (such as dynamic loads of different amplitudes and frequencies), electrochemical fields (such as corrosive media of different concentrations) and temperature fields (such as different temperature conditions), including information such as stress-strain relationship, corrosion current density, and material temperature distribution.

[0035] Step 1.2, construct the physical field coupled partial differential equations;

[0036] Based on the collected data, a group of partial differential equations describing the multi-physics field coupling relationship is established, and its calculation formula is:

[0037] ;

[0038] in, 、 Indicates the 、 The state variables of the physical field, 、 Indicates the 、 The spatial gradient of the physical field state variable, is the total number of physical fields; represents the time variable; Represents the interaction function between physical fields; Represents the influence function of control input on physical field; 、 Indicates the 、 control input variables, Enter the total number of variables for control.

[0039] Step 1.3, discretize the coupled equations using the finite element method;

[0040] The coupled partial differential equations established in step 1.2 are spatially discretized using the finite element method, transforming the continuous domain problem into a numerical calculation problem on discrete nodes, and obtaining the discretized equations:

[0041] ;

[0042] in, Represents the physical field state variable vector at the discrete node; and denote their first-order and second-order time derivatives respectively; 、 and denote the mass matrix, damping matrix and stiffness matrix respectively; The resultant force vector representing the external load and internal coupling is the control input vector and the state variable vector function.

[0043] Step 1.4: Verify the accuracy of the multi-physics coupling model;

[0044] By comparing the model prediction results with the measured data, the accuracy and reliability of the model are evaluated, and the model parameters are adjusted when necessary to ensure that the model can accurately describe the interaction relationship between multiple physical fields.

[0045] Step 2: Apply deep reinforcement learning algorithms to build a collaborative control model for environmental parameters, achieving high-precision control of multiple environmental parameters and adaptive regulation based on material response;

[0046] Specifically include:

[0047] Step 2.1, construct the state space and action space of the experimental environment;

[0048] Define the test environment state space , including environmental parameters such as temperature, humidity, salinity, pH value, and state variables such as stress state and corrosion potential of the test sample; define the action space , including the adjustment parameters of each environmental control equipment, such as heating power, humidity control value, salt spray concentration, etc.

[0049] Step 2.2, build a Q-learning model based on deep reinforcement learning;

[0050] The deep Q network (DQN) is used to implement strategy learning in complex environments. Its core is the iterative update of the Q-value function, which is calculated as follows:

[0051] ;

[0052] in, Indicates that the status Next action The expected cumulative reward value of Indicates immediate reward; represents the discount factor, which is used to balance immediate rewards and long-term benefits; Indicates execution of an action The next state after Indicates the maximum Q value of the next state.

[0053] Step 2.3, construct a deep neural network to approximate the Q-value function;

[0054] In order to process high-dimensional continuous state space, a deep neural network is used to approximate the Q value function. The network structure includes an input layer (state variables), multiple hidden layers (using ReLU activation function) and an output layer (Q value of each action). By randomly sampling batch experience data , minimize the loss function for training:

[0055] ;

[0056] in, is the loss function, Indicates the current network parameters; Represents the target network parameters, which are regularly copied and updated from the current network; Indicates the desired action.

[0057] In this implementation, the deep neural network constructed has the following structure: the input layer contains 24 neurons, corresponding to the recorded values of four environmental parameters: temperature, humidity, salinity, and pH at the current moment and the past five moments. The hidden layer consists of three layers, each containing 128, 64, and 32 neurons, respectively, all using the Reluctant Unit (ReLU) activation function to enhance the network's nonlinear representation capabilities. The output layer contains 15 neurons, corresponding to different control actions for the temperature, humidity, salinity, and pH parameters. The network uses an experience replay buffer with a storage capacity of 10,000 transfer samples during training. Mini-batch training is performed by randomly sampling 512 samples at a time, and the target network parameters are updated every 200 steps during training. This neural network achieved high-precision environmental control with temperature control accuracy of ±0.5°C, relative humidity control accuracy of ±0.8%, and salt concentration control accuracy of ±0.05% in corrosion fatigue testing of reinforced concrete bridge specimens under a combined salt spray and dynamic load environment.

[0058] Step 2.4, realizing the adaptive control model based on material response;

[0059] The material response signal of the test sample (such as strain growth rate, corrosion current change rate, etc.) is used as feedback input, and the environmental parameter control strategy is dynamically adjusted through the reinforcement learning model to achieve adaptive regulation of environmental parameters based on the real-time response of the material, thereby improving the precise control capability of the test process.

[0060] In this embodiment, the adaptive control model adopts a dual-input strategy. In addition to the environmental parameter state, the material response characteristics are also introduced as the input of the model. Specifically, a deep network structure with dual-channel input is constructed: the first channel receives environmental parameter information, and the second channel receives material response signals (including strain change rate, corrosion potential change rate, corrosion current density, etc.). The two channels pass through their respective feature extraction layers and are merged in the fusion layer to finally output the control action of the environmental parameters. The reward function design of this model consists of two parts: environmental parameter control accuracy reward and material response compliance reward, and its calculation formula is:

[0061] ;

[0062] in, Indicates that the status Next action The reward value obtained; and represent the actual value and target value of the i-th environmental parameter respectively; Indicates the total number of environmental parameters; represents the material response compliance evaluation function; and are the weight coefficients of environmental control accuracy and material response compliance respectively.

[0063] In practical applications, this adaptive control model was successfully applied to coupled corrosion fatigue testing of reinforced concrete components of the Hangzhou Bay Bridge. Based on the monitored sudden changes in the steel corrosion current density, the model automatically adjusted the salt spray concentration and relative humidity to align the corrosion rate with the actual environment. Furthermore, based on changes in the material's fatigue crack growth rate, the model intelligently adjusted the load frequency and amplitude, ensuring a 31% improvement in the consistency of the test process with actual service conditions, significantly outperforming traditional fixed-parameter testing methods.

[0064] Step 3: Construct an environmental accelerated test equivalence mapping model and optimize the accelerated test parameter combination through a multi-objective optimization algorithm to achieve a balance between the optimal speedup ratio and equivalence.

[0065] Specifically include:

[0066] Step 3.1, collect actual bridge service environment data and accelerated test data;

[0067] Collect long-term monitoring data of actual bridges under different service environments, including environmental parameter variation patterns and structural response data; at the same time, collect short-term test data under different accelerated test conditions to provide a data basis for establishing a mapping relationship between accelerated tests and actual service environments.

[0068] Step 3.2, establish the mathematical model of environmental accelerated test equivalence;

[0069] Based on the collected data, a mathematical model of equivalence between the environmental acceleration test parameters and the actual service environment is established, and its calculation formula is:

[0070] ;

[0071] in, represents the accelerated test environment parameter vector; represents the actual service environment parameter vector; Indicates the time compression ratio; Represents the environmental parameter mapping function; Represents the time-varying acceleration factor, which is used to dynamically adjust the accelerated test parameters to maintain equivalence with the actual service environment.

[0072] Step 3.3, optimize the accelerated test parameter combination;

[0073] Based on the equivalent mathematical model established in step 3.2, a multi-objective optimization algorithm is used to solve the optimal accelerated test parameter combination. The objective function includes maximizing the time compression ratio and minimizing the equivalence deviation. The calculation formula is:

[0074] ;

[0075] in, and Represent the weight coefficients of time compression ratio and equivalence deviation respectively; Indicates the time compression ratio; represents the actual service environment parameter vector; A function that measures the deviation of the equivalence between the accelerated test environment and the actual service environment; Represents the norm operation.

[0076] In this embodiment, the multi-objective optimization problem is solved by an improved non-dominated sorting genetic algorithm (NSGA-II). The specific implementation of the algorithm is as follows: First, the decision variables are defined as the accelerated test environment parameter vector , including parameters such as temperature cycle amplitude, humidity variation range, salt spray concentration, pH value range, and loading frequency in the accelerated test; secondly, a population (population size is 100) is constructed and encoded; then, a fitness function is defined, including two objectives:

[0077] Maximize time compression ratio , the value range is [5, 20];

[0078] Minimize equivalence bias , calculated by comparing key indicators such as corrosion rate and fatigue crack growth rate in accelerated tests with those in actual service environments.

[0079] After the algorithm runs for 500 generations, the non-dominated solution set is screened, and the optimal compromise solution is selected through the fuzzy set theory method.

[0080] This optimization algorithm was successfully applied in the planning of the corrosion fatigue coupling test of the steel box girder of the Wuhan Yangtze River Highway Bridge. By globally optimizing 16 environmental parameters, the corrosion fatigue test cycle, which originally required three years, was shortened to nine months, with a time compression ratio of four times, while ensuring that the equivalence deviation was less than 8%. The optimized accelerated test parameters included expanding the temperature fluctuation of ±15°C in the service environment to ±35°C, increasing the stress amplitude of the original 10MPa to 25MPa, and increasing the salt spray concentration from 3.5% to 6.8%. The test results showed that the damage morphology and failure mechanism of the accelerated test samples were highly consistent with the long-term on-site monitoring results, verifying the effectiveness and reliability of the optimization algorithm.

[0081] Step 3.4, verify the equivalence of accelerated tests;

[0082] By comparing the accelerated test results with actual service data, the equivalence and reliability of the accelerated test are verified, and the equivalent model parameters are adjusted when necessary to ensure that the accelerated test can accurately reflect the corrosion-fatigue coupling effect under actual service conditions.

[0083] Step 4: Build a generalized predictive control model and implement multi-physics coordinated control through a rolling optimization strategy to improve control accuracy and response speed.

[0084] Specifically include:

[0085] Step 4.1, establish a system prediction model;

[0086] Based on the multi-physics coupling mathematical model constructed in step 1, a control-oriented system prediction model is established to predict the response of the system in multiple time steps in the future. The calculation formula is:

[0087] ;

[0088] in, represents the length of the prediction time domain, Indicates that from the current moment The beginning of the future The system output prediction vector of time steps; 、 、 Respectively indicate at the current moment For the future 、 、 The predicted value of the system output; Represents a transpose operation.

[0089] Step 4.2, construct the generalized predictive control cost function;

[0090] The cost function of the design includes the output tracking error and the control input change, and its calculation formula is:

[0091] ;

[0092] in, represents the cost function of generalized predictive control, Indicates the length of the prediction time domain; Indicates the length of the control time domain ( ); represents the reference trajectory; Indicates the change in control input; and Represent the weighting matrices of output tracking error and control input variation respectively; Represents the time step index.

[0093] Step 4.3, solve the optimal control sequence;

[0094] Based on the cost function constructed in step 4.2, the quadratic programming method is used to solve the optimal control sequence, and its calculation formula is:

[0095] ;

[0096] in, represents the optimal control sequence; Indicates that at the current moment The calculated optimal control input value that should be applied to the system immediately; 、 Represents the current time For the future 、 The optimal control input.

[0097] Step 4.4, constructing a rolling optimization control model;

[0098] Adopting the rolling optimization strategy, only the first control input of the optimal control sequence is executed in each control cycle. ,Then the state is updated according to the actual response of the system, the optimal control sequence is re-solved, closed-loop feedback control is realized, and the robustness of the system to disturbances is improved.

[0099] In this embodiment, the rolling optimization control model is implemented with the following specific steps:

[0100] At the current moment , based on the current state of the system and multi-physics coupled mathematical models to predict the future System output at a moment ;

[0101] Solve the optimization problem to get the future The optimal control sequence at each moment ;

[0102] Only the first control input is executed ;

[0103] Measure the actual state of the system at the next moment ;

[0104] Move the time pointer forward one step to , repeat the above process.

[0105] The key parameters of this model are set as: prediction time domain , control time domain , the control cycle is 200ms. The quadratic programming solver qpOASES in the MATLAB environment is used to solve the optimization problem. An optimization calculation can be completed within 50ms, meeting the real-time control requirements.

[0106] This rolling optimization control model was successfully applied to the corrosion fatigue coupled test system in the cable anchorage area of the Sutong Bridge, achieving coordinated control of multiple parameters such as temperature, humidity, salt concentration, and mechanical load. When faced with sudden disturbances (such as a sudden temperature change of ±5°C), the control model was able to suppress the effects of the disturbance within three control cycles (600ms), allowing the system to quickly return to the target state. Traditional PID control requires 12 control cycles to achieve the same effect. In practical applications, this control model improved the stability of the corrosion fatigue test environment by 78% and the smoothness of environmental condition changes during critical transition periods by 65%. This effectively avoided the overshoot and oscillation problems common in traditional control methods, providing a reliable guarantee for accurately simulating the corrosion fatigue coupled behavior of bridge structures under real-world service conditions.

[0107] Step 5: Build a real-time system identification and parameter update model, implement online identification of system parameters based on the recursive least squares algorithm, and actively compensate and stabilize the interaction of multiple physical fields.

[0108] Specifically include:

[0109] Step 5.1, construct a parameterized system model;

[0110] In order to realize the online identification of system parameters, a parameterized system model is constructed, and its calculation formula is:

[0111] ;

[0112] in, Indicates system output; Represents the regression vector, which contains past input and output data; represents the parameter vector to be identified; Represents the model error.

[0113] Step 5.2, design a recursive least squares algorithm for parameter identification;

[0114] The recursive least squares (RLS) algorithm is used to realize the online identification and parameter update of the system parameters. The calculation formula is:

[0115] ;

[0116] in, 、 Respectively indicate time 、 Parameter estimates of ;

[0117] represents the Kalman gain, which is calculated as follows:

[0118] ;

[0119] in, Indicates time The parameter estimate covariance matrix of .

[0120] Covariance matrix The update formula is:

[0121] ;

[0122] In this embodiment, the specific implementation of the recursive least squares algorithm includes the following steps: first, initialize the parameter estimation value Initialize the covariance matrix to a zero vector ,in is a large positive number (such as 1000), is the unit matrix; then, in each control cycle (typical value is 100ms), based on the newly acquired system input and output data, a regression vector is constructed , contains the input and output data of the past four control cycles; then the parameter update amount is calculated according to equations (11)-(13). In order to improve the robustness of the algorithm, this embodiment also introduces the forgetting factor (Value range 0.95-0.99), modify the covariance matrix update formula to:

[0123] ;

[0124] This algorithm has been successfully used in practical applications to identify multi-physics interaction parameters in a large suspension bridge main cable corrosion fatigue coupled test system. The algorithm converges model parameters to within ±3% of their true values within 30 iterations. It can also rapidly track changes in system dynamics under sudden changes in environmental conditions (such as temperature and humidity), completing adaptive updates of model parameters within 200 milliseconds. This real-time identification capability enables the system to accurately capture the nonlinear interactions between different physical fields, providing a precise model foundation for subsequent active compensation control.

[0125] Step 5.3, construct an adaptive control method based on model parameters;

[0126] The system parameters identified in step 5.2 Substitute it into the generalized predictive control model in step 4, dynamically update the control model, implement adaptive control based on model parameters, and improve the system's adaptability to parameter changes.

[0127] Step 5.4, construct an active compensation method for multi-physics field interaction;

[0128] Based on the identified system parameters, the interaction between multiple physical fields is analyzed, and an active compensation method is constructed to offset the system instability factors caused by the interaction, thereby improving the stability of the test conditions and the control accuracy.

[0129] In this embodiment, the active compensation method for the interaction of multiple physical fields is implemented using a feedforward feedback composite structure. First, based on the multi-physical field coupling model obtained by real-time system identification, an interaction decoupling matrix is constructed. ,in To control the input dimension, is the physical field state dimension. The decoupling matrix is generated using the following algorithm:

[0130] Calculate the interaction sensitivity matrix between various physical fields based on the identified system model ,element Indicates the The physical field The degree of influence of a physical field;

[0131] Based on the sensitivity matrix, construct the decoupling matrix ,in, Represents the sensitivity matrix The pseudo-inverse matrix of

[0132] The singular value decomposition (SVD) method is used to deal with the condition number problem of the decoupling matrix to ensure the stability of the numerical solution.

[0133] The implementation process of compensation control is as follows:

[0134] Through the decoupling matrix , the target physical field state change Mapped as compensation control input :

[0135] ;

[0136] Superimpose the compensation control input and the basic control input to form the final control command;

[0137] Based on the actual response, the decoupling matrix is dynamically adjusted to adapt to changes in system characteristics.

[0138] This active compensation method was successfully applied in the corrosion fatigue coupling test of the steel structure of the Nanjing Yangtze River Bridge. For example, during the test, when the mechanical load change caused the specimen temperature to rise by 3.2°C, the traditional method required 20 minutes for the temperature to return to stability. However, using the active compensation control of this method, the system was able to predict and compensate for this interaction in advance, controlling the temperature fluctuation within the range of ±0.6°C and shortening the stabilization time to 3 minutes. Similarly, for the interaction between the electrochemical field and the mechanical field, this method reduced the corrosion potential fluctuation from the traditional control of ±25mV to ±5mV, significantly improving the accuracy and reliability of the test data. The application of this method reduced the field interaction interference in the multi-physics field coupling test by 87%, providing technical support for accurately reproducing the complex physical field coupling effects in the actual service environment.

[0139] This implementation method proposes an artificial intelligence-based bridge structure corrosion fatigue coupling test method. With the help of key technologies such as multi-physics field coupling modeling and deep reinforcement learning environment control, it solves the difficulties of existing test methods and achieves significant technical results. In terms of test environment control accuracy, deep reinforcement learning environmental parameter collaborative control strategy and generalized predictive control algorithm are used to control environmental parameter fluctuations within ±0.8%, which is much better than the ±5% of traditional methods, thereby improving the reliability of test data. In terms of collaborative control of multiple environmental conditions, by constructing a multi-physics field coupling mathematical model and a generalized predictive control algorithm, any combination and smooth transition of more than 10 environmental conditions such as temperature and humidity can be achieved, which is close to the actual service environment of the bridge. With the help of the environmental accelerated test equivalence mapping model, the test cycle is shortened by 65% while ensuring data reliability, reducing test costs. Through multi-physics field collaborative control and real-time system identification, intelligent matching of environmental and load conditions is achieved, and the consistency between the test and the actual bridge service conditions is improved by 90%. Through real-time system identification and parameter updating, the interaction of multiple physical fields is actively compensated, the fluctuation of multiple rounds of test data is reduced by 75%, and the comparability of results between different laboratories is improved by 85%, providing reliable data for bridge structure safety assessment, and overcoming the shortcomings of traditional test methods in terms of accuracy, authenticity, efficiency and stability. 4. Feedforward feedback Real application examples of this implementation method

[0140] In one embodiment of the present invention, an example of the aforementioned bridge structure corrosion fatigue coupling test method based on artificial intelligence is provided:

[0141] Application Scenario Description: The methods of this embodiment are being applied to a durability research project for a heavy-duty railway steel bridge in a coastal region. Located in my country's southeastern coastal region, the bridge operates in an environment characterized by high humidity, high salt spray concentration, and cyclical temperature fluctuations. It also undergoes high-frequency fatigue loading from heavy trains. The research objective is to evaluate the evolution of coupled corrosion-fatigue damage in key steel bridge structures and its impact on the bridge's remaining service life, providing a scientific basis for bridge maintenance decisions. Traditional research methods require eight to ten years of field testing or use simplified accelerated tests, which have poor equivalence and cannot accurately predict the evolution of structural performance under actual service conditions. To address these issues, this project utilizes the artificial intelligence-based coupled corrosion-fatigue testing method proposed in this embodiment to establish an equivalent accelerated testing method, thereby obtaining valuable research data within a reasonable timeframe. The test subjects were 25 standard specimens taken from the steel bridge, including key structural features such as welded joints, bolted connections, and main beam sections. The test apparatus included an intelligent environmental control system (for temperature, humidity, salt spray concentration, pH, etc.) and a multi-channel electro-hydraulic servo loading system, designed and controlled entirely based on the methods of this embodiment.

[0142] In this application example, a three-field coupled mathematical model was established based on one year of field monitoring data, encompassing mechanical, electrochemical, corrosion, and temperature fields. Targeting the unique environment of coastal heavy-load railway steel bridges, the key variables and their interactions are shown in Table 1:

[0143] Table 1: Key multi-physics variables and their interactions in the feedforward feedback steel bridge corrosion fatigue coupling test:

[0144]

[0145] Based on actual monitoring data and material test results, the established mechanical field and electrochemical field coupling function is as follows:

[0146] ;

[0147] in, and Respectively represent the corrosion current density with and without strain; and represent the current strain and initial strain respectively; and Respectively represent the current temperature and reference temperature; and Represents the strain sensitivity coefficient and temperature sensitivity coefficient respectively, which are determined by experiments , .

[0148] In view of the particularity of this application scenario, a stress corrosion crack growth rate correction function is introduced based on the traditional model:

[0149] ;

[0150] in, represents the crack growth rate; represents the effective stress intensity factor amplitude; and is the material constant; is the corrosion temperature correction function, and its expression is:

[0151] ;

[0152] in, and are the reference corrosion current density and reference temperature, respectively; 、 and is the fitting parameter, determined by comparative test , .

[0153] The prediction accuracy of the model in practical applications is shown in Table 2:

[0154] Table 2: Prediction accuracy evaluation of the feedforward-feedback multi-physics coupling model:

[0155]

[0156] In this application example, the deep reinforcement learning environmental control model constructed in Step 2 enabled high-precision coordinated control of environmental parameters during a steel bridge corrosion fatigue coupled test. To address the unique service environment of coastal steel bridges, the state space of environmental parameters was expanded, and targeted optimizations were performed on the basic network architecture.

[0157] Aiming at the corrosion fatigue characteristics of steel bridges, a dual reward function is designed, including environmental control accuracy reward and material response matching reward:

[0158] ;

[0159] in, and represent the actual value and target value of the i-th environmental parameter respectively; Function that represents the matching degree between material corrosion current and crack growth rate and field data; and are the weight coefficients for control accuracy and material response matching, respectively.

[0160] After 200 hours of actual test system training, this deep reinforcement learning model achieved high-precision coordinated control of environmental parameters. In response to rapidly changing environmental scenarios, the environmental control effects of the traditional PID control method and this method are compared as shown in Table 3:

[0161] Table 3: Comparison of the effects of traditional PID with feedforward feedback and deep reinforcement learning control methods:

[0162]

[0163] In actual applications, the system can accurately simulate the diurnal and seasonal changes in coastal environments. Material response evaluation shows that the obtained corrosion current density changes are 93.2% consistent with the data monitored in the actual service environment, and the crack growth rate consistency reaches 91.5%, which is much higher than traditional control methods.

[0164] In this application example, the environmental accelerated test equivalence mapping model constructed in step 3 achieved time compression of the corrosion fatigue coupled damage process of steel bridges, significantly shortening the test cycle while ensuring equivalence.

[0165] First, based on three years of historical monitoring data collected on-site, typical variation patterns of parameters in actual service environments and the development patterns of corrosion fatigue damage were established. Then, a modified non-dominated sorting genetic algorithm (NSGA-II) was used for multi-objective optimization to determine the optimal combination of accelerated test parameters. For the steel bridge application scenario, the optimization process used the following key parameters, as shown in Table 4:

[0166] Table 4: Key parameters for optimization of feedforward feedback accelerated test equivalence:

[0167]

[0168] Based on a multi-physics coupling model and extensive test data, an equivalence evaluation standard was established between the accelerated test and the actual service environment, including indicators such as damage mechanism consistency, corrosion morphology similarity, and fatigue crack growth behavior matching. The genetic algorithm used a population size of 150 and ran 800 generations. The optimal acceleration parameter combination was finally determined as shown in Table 5:

[0169] Table 5: Acceleration test parameter combinations and time compression effects after feedforward feedback optimization:

[0170]

[0171] Comprehensive optimization results demonstrate a 5.3-fold time reduction while maintaining 90.5% equivalence. This means the corrosion-fatigue coupled damage process, which would normally take 10 years to complete, was completed in just 23 months in this application example, significantly improving research efficiency. Comparative analysis of damage morphology revealed a 92.7% consistency between the corrosion fatigue crack morphology of the accelerated test and that of field-tested steel structures, achieving a highly reliable fatigue life prediction accuracy of ±8.5%.

[0172] In this application example, the generalized predictive control model and real-time system identification model constructed based on steps 4 and 5 achieved highly coordinated regulation of environmental conditions and load conditions, effectively solving the problem of insufficient test authenticity caused by the separate control of environment and load in traditional test systems.

[0173] The hardware configuration of the collaborative control system is as follows: Intel Core i7 processor, 16GB memory, NVIDIA GeForce RTX 3080 GPU acceleration, and a control cycle set to 50ms. The following key parameters were configured for the coastal steel bridge test scenario, as shown in Table 6:

[0174] Table 6: Key parameter configuration of feedforward feedback multi-physics field coordinated control system:

[0175]

[0176] The application effect of this system in the steel bridge corrosion fatigue coupling test is manifested in the following aspects:

[0177] It achieves highly coordinated changes in temperature field, electrochemical field and mechanical field, and accurately reproduces the complex coupling relationship of the three physical fields in the actual service environment. For example, when the temperature rises, the salt spray concentration is automatically increased to maintain the consistency of the corrosion current density, and when the load increases, the temperature and humidity parameters are actively adjusted to compensate for the environmental disturbance caused by the load.

[0178] The model parameters are continuously updated through system identification, effectively coping with nonlinear changes and uncertainties in the test process.

[0179] Technical Effect Verification: This application example fully verifies the technical effects of Implementation Method 1, and has achieved remarkable results in the following two aspects:

[0180] Improved consistency between feedforward feedback test and actual service conditions: Compared with traditional corrosion fatigue test methods, this implementation significantly improves the consistency between test and actual service conditions, providing more reliable data support for the remaining life assessment of steel bridge structures. Detailed comparison results are shown in Table 7:

[0181] Table 7: Comparison of consistency between feedforward feedback test and actual service:

[0182]

[0183] Feedforward feedback test efficiency and cost optimization effect: Implementation method 1 significantly improves test efficiency and reduces research costs while ensuring test quality. The efficiency and cost comparison of three different scale corrosion fatigue coupling tests is shown in Table 8:

[0184] Table 8: Comparison of efficiency and cost of feedforward feedback at different test scales:

[0185]

[0186] The above technical validation data demonstrates that, in a coupled corrosion fatigue test of a coastal heavy-load railway steel bridge, this method achieved a 39.6% improvement in consistency between test and actual service conditions, while reducing testing costs by 64.7%, compared to conventional methods. This fully validates the technical effectiveness of this invention. This technological achievement has been successfully applied in three steel bridge life assessment projects of varying scales.

[0187] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A bridge structure corrosion fatigue coupling test method based on artificial intelligence, characterized in that: The following steps are involved: Construct a multi-physics field coupling mathematical model including mechanical field, electrochemical field and temperature field to accurately express the interaction between physical fields; A deep reinforcement learning algorithm is used to construct an environmental parameter collaborative control model to achieve high-precision control of multiple environmental parameters and adaptive regulation based on material response. The deep reinforcement learning environmental parameter collaborative control model uses a deep Q-network to achieve policy learning in complex environments, and learns the optimal control strategy by iteratively updating the Q-value function, where the Q-value function represents the expected cumulative reward value for performing a specific action under a specific state. The reward function of the deep reinforcement learning environmental parameter collaborative control model is designed to be a weighted combination of environmental parameter control accuracy and material response compliance. The environmental parameter control accuracy is calculated by the deviation between the actual value and the target value, and the material response compliance is quantified using a dedicated evaluation function. Construct an environmental accelerated test equivalence mapping model, optimize the accelerated test parameter combination through a multi-objective optimization algorithm, and achieve a balance between the optimal acceleration ratio and equivalence; Build a generalized predictive control model and achieve multi-physics coordinated control through a rolling optimization strategy to improve control accuracy and response speed; Build a real-time system identification and parameter update model, realize online identification of system parameters based on the recursive least squares algorithm, and actively compensate and stabilize the interaction effects of multiple physical fields.

2. The bridge structure corrosion fatigue coupling test method based on artificial intelligence according to claim 1 is characterized in that: The multi-physical field coupling mathematical model is established through a group of partial differential equations, which includes the state variables of each physical field, the time derivatives of the state variables, the spatial gradients of the state variables, the interaction functions between the physical fields, and the influence functions of the control inputs on the physical fields, and is used to characterize the coupling relationship between different physical fields.

3. The bridge structure corrosion fatigue coupling test method based on artificial intelligence according to claim 1 is characterized in that: The environmental acceleration test equivalence mapping model establishes a mapping relationship between the accelerated test environment parameters and the actual service environment parameters, and introduces a time compression ratio and a time-varying acceleration factor to achieve accurate setting of the accelerated test conditions.

4. The bridge structure corrosion fatigue coupling test method based on artificial intelligence according to claim 3 is characterized in that: The acceleration test parameter combination optimization adopts a multi-objective optimization algorithm to find the optimal acceleration test parameter combination by balancing the two objectives of time compression ratio and equivalence deviation. The equivalence deviation is evaluated by a special measurement function to evaluate the degree of difference between the acceleration test environment and the actual service environment.

5. The bridge structure corrosion fatigue coupling test method based on artificial intelligence according to claim 1 is characterized in that: The generalized predictive control model constructs a weighted cost function that includes output tracking error and control input variation, and performs rolling optimization in the prediction time domain and the control time domain to achieve coordinated regulation of multiple physical fields.

6. The bridge structure corrosion fatigue coupling test method based on artificial intelligence according to claim 1 is characterized in that: The real-time system identification adopts a recursive least squares algorithm to achieve online identification of the system model by continuously updating parameter estimates, wherein the parameter update process uses the Kalman gain to perform weighted correction on the error between the system output and the predicted output.

7. The bridge structure corrosion fatigue coupling test method based on artificial intelligence according to claim 6 is characterized in that: The active compensation method for the interaction of multiple physical fields is implemented using a feedforward-feedback composite structure. The interaction decoupling matrix is constructed by the multi-physical field coupling model obtained based on real-time system identification. The compensation control input is calculated according to the change in the target physical field state to achieve effective compensation for the interaction of physical fields.

8. An artificial intelligence-based bridge structure corrosion fatigue coupling test system, used to execute the artificial intelligence-based bridge structure corrosion fatigue coupling test method according to any one of claims 1 to 7, characterized in that: include: Multi-physics coupling model module, used to build coupled mathematical models including mechanical fields, electrochemical fields, and temperature fields; A deep reinforcement learning control module for coordinated control of environmental parameters and adaptive regulation based on material response; Accelerated test mapping module, used to establish an equivalent mathematical model between environmental accelerated test parameters and actual service environment; Generalized predictive control module, used to achieve multi-physics field coordinated regulation and rolling optimization control; A real-time system identification module that continuously updates coupled model parameters and actively compensates for multi-physics interactions.

Images