A Real-Time Hybrid Test Method for Fire in Large-Scale Substructures of Bridge Cables

Abstract

This invention discloses a real-time hybrid test method for large-scale substructure fire testing of bridge cables, belonging to the field of bridge fire resistance performance testing technology. The entire bridge is divided into a physical test substructure and an LSTM numerical substructure. The numerical substructure is constructed based on a multi-scale gated residual LSTM, achieving high-precision prediction of boundary forces or displacements of the cable system under fire conditions within seconds through parallel multi-branch networks, dynamic feature selection gates, and residual cross-layer connections. Combined with an LSTM-PID intelligent control system, the prediction commands are converted into displacement / stress and temperature fields into loading signals, synchronously integrating a time-delay compensation mechanism to drive the through-hole jacks, burners, and fans in real time, reproducing non-uniform wind-fire coupling scenarios and forming a three-dimensional visualized fire response for the entire bridge. This invention breaks through the traditional finite element calculation speed bottleneck, achieving second-level thermo-mechanical coupling interaction, providing a high-precision and high-reliability test method and platform for the fire resistance design of long-span bridges.

Description

Technical Field

[0001] This invention relates to the field of bridge fire resistance performance testing technology, and in particular to a real-time hybrid test method for fire in large-scale substructures of bridge cables. Background Technology

[0002] In modern transportation networks, long-span bridges serve as the core arteries connecting regional economies, and their structural safety under fire conditions directly impacts public safety and the sustainable operation of these vital transportation networks. Fires not only soften steel but also trigger internal force redistribution, local buckling, and even the continuous collapse of the entire bridge through non-uniform temperature fields. This thermo-coupling disaster mechanism poses a severe challenge to traditional fire-resistant design methods.

[0003] Current fire resistance research methods have significant shortcomings: traditional experiments rely on homogenized heating in combustion furnaces or scaled-down models, making it difficult to reproduce the coupling effects of non-uniform fire loads, wind fields, and force loads. While hybrid simulation technology is mature in seismic engineering, it is still in the exploratory stage in fire scenarios. Existing numerical tools, such as OpenSees and ABAQUS, can model and perform thermo-mechanical coupling analysis, but their computational efficiency is low and cannot meet the needs of real-time hybrid experiments. Some foreign studies have attempted to combine experimental substructures with numerical models, but due to computational delays and scaling effects, they have been unable to achieve high-fidelity analysis of full-scale structures. Domestically, progress has been slow due to the lack of efficient numerical substructure calculation models and experimental platforms.

[0004] The unique characteristics of fire resistance assessment for long-span bridges further exacerbate technical bottlenecks: First, the collaborative working mechanism of the cable system, bridge towers, and stiffening girders exhibits strong nonlinear characteristics during fires, making it impossible to reveal the failure path and evolution of the overall structure under fire conditions through local component tests. Second, the structural stiffness degradation and cable force redistribution caused by high fire temperatures exhibit significant spatiotemporal asynchrony, requiring real-time coupling of thermo-mechanical boundary conditions. However, existing hybrid test systems lack sufficient accuracy to meet the demands of long-term loading. Third, existing systems rely on traditional finite element calculations, making it difficult to overcome the "speed-accuracy" contradiction, resulting in significant error accumulation under long-term loading. Furthermore, although some studies have attempted to improve efficiency through GPU acceleration and model order reduction, the computational speed still cannot meet the requirements of real-time hybrid experiments.

[0005] Therefore, in response to the urgent need for fire safety assessment of long-span bridges, it is imperative to develop a hybrid testing method that integrates large-scale substructure physical experiments, rapid computational numerical models based on neural networks, and intelligent time delay compensation. This method aims to overcome the spatial scale limitations and real-time bottlenecks of traditional approaches, providing technical support for bridge fire-resistant design that combines scientific rigor with the operability of physical experiments. Summary of the Invention

[0006] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the first objective of this invention is to propose a real-time hybrid testing method for fire in large-scale substructures of bridge cables, breaking through the bottleneck of traditional finite element calculation speed and achieving second-level thermo-mechanical coupling interaction, thus providing a high-precision and high-reliability testing method for the fire-resistant design of long-span bridges.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a real-time hybrid test method for fire in large-scale substructures of bridge cables, comprising:

[0008] S1, divide the cable-bearing bridge to be analyzed into experimental substructure and numerical substructure;

[0009] S2, deploy the trained fire temperature field rapid generation model and the LSTM-based numerical substructure rapid calculation model to the computing server;

[0010] S3, fabricate the test specimens corresponding to the experimental substructure, and build a vehicle-fire-wind coupling test platform;

[0011] S4, Deploy a fire scene environmental monitoring system;

[0012] S5. The specimen is preloaded on the vehicle-fire-wind coupling test platform, and the initial time delay of each system is obtained and input into the LSTM-PID intelligent control system for feedforward compensation.

[0013] S6, the computing server receives data from the fire environment monitoring system of the monitoring system; the numerical substructure fast calculation model based on LSTM predicts the boundary displacement of the next time step in real time, and converts the predicted boundary displacement into force or displacement loading command of physical test according to the displacement equivalence principle.

[0014] S7. Apply force load and open flame non-uniform heat load to the specimen. Send the force or displacement loading command of the test substructure boundary extracted in step S6 to the self-balancing loading reaction frame through the LSTM-PID intelligent control system and synchronously collect data and transmit it back to the calculation server.

[0015] S8 activates the LSTM-based time delay compensator for dynamic compensation, which predicts the time delay and issues control commands in advance to gradually reduce the time delay error.

[0016] S9. Repeat steps S6 to S8 until the deformation or stress of the specimen exceeds the limit in a certain analysis step, then end the loading.

[0017] S10 uses 3D animation software to dynamically display the stress, deformation, and temperature field distribution under a full-bridge fire.

[0018] Preferably, in step S1, when the prototype of the specimen is the main cable, its two ends pass through the free rotation anchoring end and the free rotation loading end at the same height on the self-balancing loading reaction frame, and are then anchored using anchors. As needed, a large-scale equivalent model with a specimen diameter not less than 1:4 compared to the actual bridge main cable diameter is selected.

[0019] When it is necessary to test a full-scale boom or sling, its upper end is fixed to the test model of the main cable by a cable clamp, and its lower end is connected to the hydraulic loading device embedded in the self-balancing loading reaction frame.

[0020] Preferably, in step S2, constructing the rapid generation model of the fire temperature field includes:

[0021] S21. Construct a GAN network, which includes a generator and a discriminator; among which...

[0022] The generator takes fire source power, normalized location coordinates, and wind speed as input conditions. It extracts local fire dynamic features through a two-layer spatiotemporal convolutional LSTM branch and embeds residual terms of the heat diffusion equation to construct a physical constraint branch. The two features are fused through a fully connected layer to output a normalized temperature field matrix, which is then denormalized to the actual temperature-time curve. The two branches of the two-layer spatiotemporal convolutional LSTM have 64 and 128 channels, respectively, and are activated using LeakyReLU. The output normalized temperature field matrix is ​​activated using Sigmoid and has a value range of [0,1].

[0023] The discriminator takes the generated or real temperature field matrix and fire source parameters as joint inputs, extracts spatiotemporal joint features through a three-layer convolutional network, and embeds a gradient penalty term to constrain the smoothness of the generated data; the output layer is compressed into a single node by global average pooling, outputting the probability of the temperature field's authenticity, and simultaneously calculates the residual of the heat diffusion equation through residual connections as a criterion for physical rationality. The three-layer convolutional network has 128, 64, and 32 channels respectively, and uses ReLU activation; the output layer uses Sigmoid activation.

[0024] The residual of the thermal diffusion equation is the difference between the generated temperature field and the predicted value of the theoretical equation. The calculation formula is as follows:

[0025]

[0026] In the formula, To generate a temperature field, For ambient temperature, To generate the spatial second derivative of the temperature field, This is the balance term between the heat source and the ambient temperature; The L2 norm is used to measure the magnitude of the residuals.

[0027] S22 generates full-bridge three-dimensional non-uniform temperature field data based on fire dynamics software simulation, covering combinations of different fire source power, location, and wind speed; it extracts spatiotemporal temperature distribution data to construct a training dataset, with fire source parameters as input and gridded temperature field matrix as output labels; it verifies the reliability of the data through experimental substructure tests, reproduces the target temperature field of the fire dynamics software, and compares the measured and simulated temperature gradient distributions to ensure that the error range is ≤5%;

[0028] S23, Offline training of a rapid fire temperature field generation model based on GAN network, and verification data based on fire dynamics software simulation and physical experiment.

[0029] S24 deploys the trained fire temperature field generation model to the computing server and binds the fire source parameter input interface; the input targets are fire source power, location, and wind speed, and the target temperature-time curve is generated.

[0030] Preferably, in step S2, the LSTM-based numerical substructure fast calculation model is a multi-scale gated residual LSTM neural network prediction model. This model is constructed by introducing a parallel multi-branch structure, dynamic feature selection gates, and residual cross-layer connections on top of LSTM. It is used to predict the evolution process under a full-bridge fire based on measured values ​​from multi-dimensional monitoring sensors during the experimental substructure test and to output loading boundary forces or displacement commands in real time. The specific steps for training the LSTM-based numerical substructure fast calculation model are as follows:

[0031] S31, Construct a multi-scale gated residual LSTM neural network, wherein the multi-scale gated residual LSTM neural network includes a dual-branch LSTM unit, namely a first branch LSTM unit and a second branch LSTM unit.

[0032] S32, Constructing a dataset for a fast numerical substructure computation model based on LSTM, specifically including: using fire dynamics software to simulate the three-dimensional non-uniform temperature field distribution within the full-bridge space under different fire scenarios, and extracting temperature-time curves; establishing a thermo-mechanical coupling finite element model of the full-bridge and the experimental substructure based on finite element analysis software; fabricating specimens corresponding to the experimental substructure, subjecting the specimens to non-uniform temperature field thermo-mechanical coupling loading, and verifying the accuracy of the specimen thermo-mechanical coupling model; using the verified simulation data to train a multi-scale gated residual LSTM neural network model;

[0033] S33, offline pre-training, uses a sliding time window to split the dataset, the loss function is Huber Loss, the optimizer is Nadam, the initial learning rate is 0.001, and it decays by 50% every 10 epochs; the first branch LSTM unit is input with the original step size data, the second branch LSTM unit is input with downsampled data, and the dynamic feature selection gate weights are optimized through backpropagation of historical errors.

[0034] S34, online incremental learning, dynamically adapts to changes in fire conditions by capturing sensor data in real time, fine-tuning the parameters of the model's downsampling branch and constraining the weight change amplitude to ≤0.01;

[0035] S35 performs real-time prediction and control during real-time mixed testing;

[0036] S36 iteratively updates the thermal-displacement / force boundary conditions until the specimen deformation or stress exceeds the limit, and outputs the failure mode and critical threshold.

[0037] Preferably, step S31 includes:

[0038] S311, the input layer receives historical time-series data, and the input features include location spatial coordinates, temperature field, boundary displacement, cable force and wind speed;

[0039] S312 is a parallel dual-branch LSTM unit. The first branch LSTM unit processes the original time-series data with a step size of 1, 128 hidden units, and a tanh activation function, and is used to extract short-term local mutation features. The second branch LSTM unit processes downsampled time-series data with a step size of 5, 64 hidden units, and a tanh activation function, and is used to extract long-term global evolution features. Both the first branch LSTM unit and the second branch ISTM unit contain a first-layer LSTM and a second-layer LSTM.

[0040] S313, the dynamic feature selection gate calculates branch weights using the sigmoid function, and the fusion formula is:

[0041]

[0042] Among them, h final h is the hidden state vector after fusing the two branch LSTM units. branch1 h is the output hidden state vector of the first branch LSTM unit. branch2 Let g be the output hidden state vector of the second branch LSTM unit. t The meaning is dynamic gating weight coefficient;

[0043] S314, residual block cross-layer connections alleviate gradient vanishing by superimposing the outputs and inputs of the first and second LSTM layers through skip connections, as shown in the formula:

[0044]

[0045] in, Let be the hidden state vector of the LSTM at time step t in the l-th layer. Let be the hidden state vector of the previous time step at time step t in layer l. Let [x, y, z] be the spatial coordinate vector at time step t. t , Let be the hidden state vector at time step t in layer (l-1). For the LSTM unit of layer l, l =1,2;

[0046] S315, a hybrid activation function, uses the sigmoid function for the forget gate and input gate, the tanh function for the candidate memory unit, and the Swish function f(x)=x·σ(x) for the output gate;

[0047] S316, a fully connected output layer, outputs the predicted boundary force or displacement values ​​for the next time step;

[0048] Preferably, step S35 includes:

[0049] S351, taking displacement loading as an example, converts the displacement of the next time step into a loading displacement command according to the displacement equivalence principle, the formula is:

[0050]

[0051] In the formula, D load For displacement command, D pred L represents the displacement at the next time step. phy L is the actual geometric span of the physical specimen. num For the equivalent span of the numerical model, L represents the large-scale model. phy =L num ;

[0052] S352, driven by an LSTM-PID intelligent control system, uses a self-balancing loading reaction frame. The formula for calculating the control quantity is:

[0053]

[0054] In the formula, γ is the feedforward gain coefficient, which is dynamically adjusted through fuzzy logic;

[0055] S353 dynamically maps the gated weights to the PID proportional coefficients to achieve adaptive loading.

[0056] Preferably, in step S3, the vehicle-fire-wind coupling test platform includes a self-balancing loading reaction frame, a multi-module T-shaped intelligent dynamic control combustion system, and a variable speed angle controllable fan system; wherein,

[0057] The multi-module T-shaped flame envelope combustion system includes a horizontal linear combustion module and a vertical annular combustion module. The horizontal linear combustion module includes several frame-type burners arranged linearly and equipped with vertical flame tubes. The frame-type burners are suspended and fixed in the middle test area of ​​a self-balancing loading reaction frame by several steel wire ropes. The vertical annular combustion module includes several layers of annular burners arranged vertically. The main body of the annular burner is an annular steel pipe. The annular steel pipes are fixedly connected by connecting pipes. The annular steel pipes are also equipped with several flame nozzles that can adjust the flame angle.

[0058] The variable speed and angle controllable fan system includes several highly dynamic axial flow fans arranged in a matrix topology on one side of the self-balancing loading reaction frame and capable of adjusting the wind direction and angle. The highly dynamic axial flow fans are connected to ultrasonic wind speed and direction sensors and pressure sensors to provide real-time feedback of wind field parameters to the LSTM-PID intelligent control system for dynamic error control. The highly dynamic axial flow fans are also equipped with integrated frequency converters to achieve continuous wind speed adjustment and precise tilt angle control.

[0059] Preferably, in step S5, the input layer of the LSTM-PID intelligent control system includes temperature field data, wind speed, angle, burner valve opening degree, and historical errors; the hidden layer contains 54-64 neurons; and the output layer outputs the burner valve opening degree increment Δu. gas and fan frequency f fan Control commands are generated according to the following formula:

[0060]

[0061] In the formula, Δu gas f is the increment of the burner valve opening. fan α represents the fan frequency; β and α are coupling coefficients, which are dynamically adjusted according to real-time operating conditions.

[0062] Finally, control commands are sent to the burner and fan frequency converter to synchronously trigger wind-fire coupling loading, thereby realizing the dynamic reproduction of the fire scene.

[0063] Preferably, in step S5, the control command generation step is as follows:

[0064] S51 uses sensors to collect temperature field data, wind speed angle data and historical sequence of burner valve opening in real time to construct a multi-dimensional time series input vector;

[0065] S52, the input vector is fed into the pre-trained LSTM prediction model to predict the burner valve opening increment and fan frequency at the next moment. The LSTM prediction model updates the weights online through a sliding time window, and the prediction error converges to <5%.

[0066] S53 uses the predicted burner valve opening increment and fan frequency as feedforward correction values, which are then superimposed on the feedback output of the PID controller to generate the final control command.

[0067] Preferably, in step S8, the time delay compensator is primarily an LSTM-based neural network containing a hidden layer of 16-32 neurons, and the input layer nodes contain the historical time delay sequence Δt. i The system's real-time load rate, ambient temperature gradient, and ambient wind speed gradient, where i = (1, 2, ..., n), are mapped by a fully connected layer to the predicted time delay Δt for the next analysis step. i+1 Specifically, it includes the following steps:

[0068] S81, Construct a time-delay dataset and collect historical time-delay sequences Δt. i The system's real-time load rate, ambient temperature gradient, and wind speed gradient data are used to construct an input vector, which is then input into an LSTM-based neural network. Here, Δt... i The time delay between the issuance of a control command and the achievement of the sensor's measured value.

[0069] S82, offline training and online optimization, uses data samples from the preloading stage to train an LSTM-based time-delay compensation model with the loss function being the mean absolute error; during the formal loading, after each analysis step, the data stream is captured based on a sliding window of 5-10 steps, and the network parameters are updated online through the Adam optimizer.

[0070] S83, dynamic time delay compensation, sends control commands Δt2 time in advance before loading in the second analysis step, and records the actual time delay Δt. real △t2 refers to the time difference between the issuance of the instruction in the first analysis step and the loading being completed;

[0071] S84, Error Correction and Convergence, Based on Δt i-real With △t i To correct the deviation, the weights of the LSTM-based time-delay compensation model are adjusted using the gradient descent method until the time-delay error of multiple consecutive analysis steps is ≤1 second; Δt i-real The time delay actually measured in each analysis step, i=(1,2,...,n).

[0072] Preferably, the method further includes: dynamically displaying the stress, deformation, and temperature field distribution under a full-bridge fire using three-dimensional animation software.

[0073] As can be seen from the above technical solution, compared with the prior art, the present invention provides a real-time hybrid test method for fire in large-scale substructures of bridge cables, which has the following beneficial effects:

[0074] 1. Overcoming the computational speed bottleneck: Traditional finite element methods suffer from low computational efficiency in fire scenarios, making it difficult to meet the demands of real-time hybrid experiments. This invention significantly improves computational speed by introducing a multi-scale gated residual neural network based on LSTM, achieving second-level thermo-coupling interaction and solving the problem of significant error accumulation under long-period loading in traditional methods. The rapid fire temperature field generation model, based on an improved generative adversarial network (GAN), can generate high-precision non-uniform temperature field distributions in real time, further enhancing the real-time performance and accuracy of the experiment.

[0075] 2. High-precision prediction and control: The multi-scale gated residual LSTM network, through its parallel multi-branch structure, dynamic feature selection gate, and residual cross-layer connections, can effectively extract short-term local mutation features and long-term global evolution features of specimens under fire conditions, achieving high-precision, second-level prediction of boundary forces or displacements. The LSTM-PID intelligent control system combines the predictive power of LSTM with the stability of a PID controller, dynamically adjusting loading commands based on real-time feedback data to ensure precise control of the experimental process.

[0076] 3. Real-time Dynamic Compensation and Time Delay Handling: The time delay compensation mechanism, using an LSTM-based time delay compensator, dynamically predicts and compensates for the time delay between control commands and sensor measurements, ensuring the synchronization and accuracy of the test process. This mechanism can gradually reduce the time delay error until the time delay error is less than 1 second throughout the entire test cycle. During the test, through online incremental learning and real-time data feedback, the model can dynamically adjust weights to adapt to changes in the fire scenario, further improving prediction accuracy.

[0077] 4. High-fidelity fire scene reproduction: The multi-module T-shaped flame envelope combustion system and variable-speed, angle-controllable fan system can simulate the non-uniform fire field and wind-fire coupling environment in real fires, achieving high-fidelity reproduction of fire scenes in laboratory environments. The fire scene environment monitoring system, through the deployment of various sensors (such as armored thermocouples, strain gauges, fiber optic grating sensors, ultrasonic anemometers, etc.), can monitor the fire scene environment and parameters such as stress and deformation of the specimens in real time, providing comprehensive data support for the experiment.

[0078] 5. Visualization and Real-time Monitoring: The 3D visualization module binds LSTM predicted data or sensor measured data to the 3D model mesh nodes, dynamically visualizing the stress, deformation, and temperature field distribution changes of the entire bridge under fire conditions, providing intuitive visualization support for the test process. Simultaneously, the system can dynamically update the load-displacement curves, temperature-time curves, and failure warning threshold panel, facilitating real-time monitoring of the test status.

[0079] 6. Scientific Rigor and Operability: This invention combines the advantages of physical experimental substructures and numerical substructures. Large-scale substructure experiments verify the reliability of the numerical model, ensuring the scientific rigor of the experimental results. Simultaneously, this method provides a high-precision and high-reliability experimental method and platform for the fire-resistant design of long-span bridges, possessing strong operability and practical application value.

[0080] 7. Innovation and Practicality: This invention introduces advanced machine learning methods (such as LSTM and GAN) into the field of bridge fire resistance performance testing technology, breaking through the limitations of traditional experimental methods and providing new ideas and technical support for the fire resistance design of long-span bridges. This method can effectively simulate the structural response under real fire scenarios, providing important technical basis for bridge fire resistance design, evaluation, and optimization, and has broad application prospects.

[0081] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of a real-time fire mixing test process according to an embodiment of the present invention;

[0083] Figure 2 This is a diagram of a multi-scale gated residual LSTM neural network architecture according to an embodiment of the present invention;

[0084] Figure 3 This is a flowchart of a real-time fire mixing test according to an embodiment of the present invention;

[0085] Figure 4 This is a neural network architecture diagram of a GAN-based rapid fire temperature field generation model according to an embodiment of the present invention.

[0086] Figure 5 This is a schematic diagram of the overall structure of the real-time fire mixing test device according to an embodiment of the present invention;

[0087] Figure 6 This is a schematic diagram of the overall structure and experimental substructure arrangement according to an embodiment of the present invention;

[0088] Figure 7 This is a diagram of the LSTM neural network architecture in the LSTM-PID intelligent control system according to an embodiment of the present invention;

[0089] Figure 8 This is a neural network architecture diagram of an LSTM-based time delay compensator according to an embodiment of the present invention;

[0090] Figure 9This is a schematic diagram of the composite anchor ring and variable cross-section force transmission spindle according to an embodiment of the present invention;

[0091] Figure 10 This is a schematic diagram of the structure of the freely rotating loading end according to an embodiment of the present invention;

[0092] Figure 11 This is a schematic diagram of the structure of the free-rotating anchor end according to an embodiment of the present invention;

[0093] Figure 12 This is a schematic diagram of the structure of a vertical annular combustion module according to an embodiment of the present invention.

[0094] Explanation of reference numerals in the attached figures:

[0095] 1. Variable speed angle controllable fan system; 11. High dynamic axial flow fan; 2. Horizontal linear combustion module; 21. Frame burner; 3. Vertical annular combustion module; 31. Annular steel pipe; 32. Flame nozzle; 4. Self-balancing loading reaction frame; 41. Main crossbeam; 42. Column; 43. Support; 44. Upper crossbeam; 45. Base; 46. Box-type bottom beam; 47. Sliding crossbeam; 5. Freely rotating loading end; 51. Composite anchor ring; 52. Four-sided frustum force transmission section; 53. Cylindrical rotating section; 54. First through-hole jack; 55. Primary anchor at loading end; 56. Secondary anchor at loading end; 6. Freely rotating anchor end; 71. Second through-hole jack. Detailed Implementation

[0096] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0097] The following description, with reference to the accompanying drawings, describes the real-time hybrid test method for fire in large-scale substructures of bridge cables proposed in this invention.

[0098] This invention discloses a real-time hybrid test method for fire in large-scale substructures of bridge cables, which uses a self-balancing loading reaction frame 4 for testing. (See attached document.) Figure 5 , Figures 9 to 12The self-balancing loading reaction frame 4 includes two sets of symmetrically arranged rectangular lattice frames. Each set of rectangular lattice frames includes four steel box columns 42, arranged in a rectangular pattern, with supports 43 between adjacent columns 42. Upper crossbeams 44 are bolted to the top of each column 42. The tops of the two sets of rectangular lattice frames are connected by two main crossbeams 41. Both sets of rectangular lattice frames have an integral base 45 formed by a box-shaped steel platform at their bottom. The two bases 45 are connected by a box-shaped bottom beam 46, with the height of the base 45 lower than that of the box-shaped bottom beam 46. A sliding crossbeam 47 is provided in the middle of each set of rectangular lattice frames. Driven by an electric lifting and locking system, the sliding crossbeam 47 slides freely and locks in the height direction within the rectangular lattice frame. Supports 43 are provided between adjacent columns 42, and the supports 43 are located at the bottom of the columns 42 without affecting the lifting and lowering of the sliding crossbeam 47. The electric lifting and locking system for the crossbeam includes a drive device mounted on the upper crossbeam 44. The drive device is connected to the sliding crossbeam via a conventional screw drive system. The sliding crossbeam and the column 42 are provided with an array of matching positioning holes for mounting positioning pins. The sliding crossbeam is positioned by the positioning pins.

[0099] The self-balancing loading reaction frame also includes a freely rotating loading end 5 and a freely rotating anchoring end 6.

[0100] Both the free-rotating loading end 5 and the free-rotating anchoring end 6 include a composite anchor ring 51. The composite anchor ring 51 has a square frame on the outside and a circular through hole on the inside. A variable cross-section force-transmitting main shaft is fixedly installed on both sides of the composite anchor ring 51. The variable cross-section force-transmitting main shaft includes a truncated quadrangular force-transmitting section 52 welded to the composite anchor ring 51 and a cylindrical rotating section 53 welded to the truncated quadrangular force-transmitting section 52. Radial stiffening ribs are evenly distributed on the outer surface of the truncated quadrangular force-transmitting section 52. The cylindrical rotating section 53 passes through the sliding beam 47 and rotates freely within the sliding beam 47 around the axis of the cylindrical rotating section 53 via a sleeve, thereby driving the composite anchor ring 51 to rotate freely within the sliding beam 47. Combined with the free sliding of the sliding beam 47, this allows for suspension tensioning at any angle during cable substructure loading.

[0101] Both the free-rotating loading end 5 and the free-rotating anchoring end 6 adopt a variable cross-section force-transmitting spindle and a composite anchor ring 51 design, which enables the cable model to rotate freely and anchor during the loading process, simulating the mechanical behavior under actual working conditions. The design of the variable cross-section force-transmitting spindle allows it to adaptively transmit force during loading, improving the flexibility and adaptability of the test device.

[0102] The free-rotating loading end 5 also includes a first through-hole jack 54 and a two-stage gradient anchoring unit. The first through-hole jack 54 is fixedly installed in the circular through-hole inside the composite anchor ring 51 via a flange, and the inner diameter of the circular through-hole matches the outer diameter of the first through-hole jack 54. The two-stage gradient anchoring unit includes a primary anchor 55 and a secondary anchor 56, which are respectively located on both sides of the first through-hole jack 54 to anchor the cable model. The outer diameter of the primary anchor 55 is smaller than the inner diameter of the first through-hole jack 54 and the inner diameter of the composite anchor ring 51 in the free-rotating anchoring end 6, and the outer diameter of the secondary anchor 56 is 1.2-1.5 times the outer diameter of the flange of the first through-hole jack 54. The main cable or cable-stayed model steel wire bundle extends outward through the primary anchor 55, passes through the inner cavity of the first through-hole jack 54, and is finally anchored by the secondary anchor 56.

[0103] The first through-hole jack 54 pushes the secondary anchor 56 at the loading end to achieve axial force loading on the cable model. During the loading process, the variable cross-section force-transmitting main shaft can rotate freely around its axis. The design of the dual-stage gradient anchoring unit at the loading end, through the synergistic action of the primary anchor 55 and the secondary anchor 56 at the loading end, achieves axial force loading on the cable model, improving the reliability and safety of the anchoring. The first through-hole jack 54, by pushing the secondary anchor 56 at the loading end, achieves axial force loading on the cable model, enabling precise control of the loading force and ensuring the accuracy of the experiment.

[0104] See Figures 1 to 8 The real-time hybrid test method for fire in large-scale substructures of bridge cables according to embodiments of the present invention may include the following steps:

[0105] S1 divides the cable-stayed bridge to be analyzed into experimental and numerical substructures. The experimental substructure includes the large-scale substructure of the cable system. Specifically, as follows... Figure 6 As shown.

[0106] S2 deploys the trained fire temperature field rapid generation model and the LSTM-based numerical substructure rapid calculation model to the computing server.

[0107] Among them, the rapid generation model of fire temperature field is based on generative adversarial network / GAN. It takes fire source power, location and wind speed as input parameters and outputs the temperature rise-time curve of local fire non-uniform temperature field in real time. Through the gas flow-temperature mapping model of multi-module T-shaped flame envelope combustion system, the generated temperature field curve is analyzed into valve opening, gas flow and fan frequency control command of burner spatial coordinates (x,y,z).

[0108] Furthermore, the full-bridge numerical substructure utilizes a multi-scale gated residual neural network based on LSTM for rapid computation. The weight file of the offline-trained LSTM-based numerical substructure fast computation model, including parallel branch parameters, dynamic gate weights, and residual connection coefficients, is imported into the real-time computing server. Input / output interfaces are initialized, binding sensor data streams for spatial coordinates (x, y, z), temperature (T), real-time displacement (D), cable force (F), and wind speed (V) at the substructure boundary target location. The required time window is set, and the spatial coordinates are meshed to ensure alignment of the temperature field, cable force, and the geometric position of the physical specimen. The sensor network and computing server are connected via a high-speed communication protocol, transmitting the displacement command D... load The instruction is pre-written into the control instruction queue, waiting to be called by the LSTM-PID intelligent control system to activate output limiting protection and force displacement instruction D. load The stroke shall not exceed the maximum stroke of the physical actuator.

[0109] S3. Fabricate the test specimens corresponding to the experimental substructure and build a vehicle-fire-wind coupling test platform.

[0110] The vehicle-fire-wind coupling test platform includes a self-balancing loading reaction frame, a multi-module T-shaped intelligent dynamic control combustion system, and a variable-speed, angle-controllable fan system. Large-scale cable system test substructure specimens were fabricated. The through-hole jacks in the self-balancing loading reaction frame are driven by an LSTM-PID intelligent control system, iteratively loading the cable system substructure according to the stress ratio equivalence principle. A T-shaped intelligent dynamic control combustion system is arranged along the main cable and hangers. The burner system's gas pipeline is equipped with intelligent valve positioners. Based on the target temperature curve and thermocouple feedback data, the LSTM-PID intelligent control system adjusts the gas flow in real time to realistically represent a non-uniform fire field in a laboratory environment. The variable-speed, angle-controllable fan system dynamically adjusts the wind speed and angle based on the measured values ​​of an ultrasonic anemometer, achieving real-time controllability of the fan speed and angle to simulate a real wind-fire coupling fire environment.

[0111] S4, Deploy a fire scene environmental monitoring system.

[0112] The fire scene environmental monitoring system includes armored thermocouples, strain gauges, fiber optic grating sensors, force sensors, laser displacement sensors, ultrasonic anemometers, and a high-temperature fire scene panoramic camera. Specifically, armored thermocouples, strain gauges, and fiber optic grating sensors are arranged along the main cable and suspenders; high-precision force sensors and laser displacement sensors are installed at the loading ends of the main cable and suspenders; ultrasonic anemometers and miniature pressure sensors are installed on the windward and leeward sides of the cable system test substructure; and a 360° high-temperature fire scene panoramic camera records the changes in the fire scene and specimen morphology in real time throughout the entire process.

[0113] S5. The specimen is preloaded on the vehicle-fire-wind coupling test platform, and the initial time delay of each system is obtained and input into the LSTM-PID intelligent control system for feedforward compensation.

[0114] Specifically, the test substructure was preloaded using a self-balancing loading reaction frame, a multi-module T-shaped flame envelope combustion system, and a variable-speed angle controllable fan system. The initial time delay of each system was calculated using the least squares method based on feedback data from force sensors, thermocouples, ultrasonic anemometers, and pressure sensors. The time delay values ​​were then input into the LSTM-PID intelligent control system, and the initial error was eliminated through feedforward compensation, thus completing the elimination of the initial time delay of each system before the formal loading of the mixed test.

[0115] S6, the computing server receives data from the fire environment monitoring system of the monitoring system; the numerical substructure fast calculation model based on LSTM predicts the boundary displacement of the next time step in real time, and converts the predicted boundary displacement into force or displacement loading commands for physical experiments according to the displacement equivalence principle.

[0116] Specifically, in the real-time computing server, the sensor network data deployed in step S4 is received via a high-speed communication protocol, and the multidimensional data is mapped according to a spatial grid to construct the input matrix [x,y,z,T,D,F,V]. t It performs standardization processing using pre-stored normalization parameters; and predicts the boundary displacement D of the target position in the next time step in real time. pred (x,y,z), single inference time ≤ 1 second; according to the displacement equivalence principle, the predicted displacement is converted into physical test loading command.

[0117] S7. Apply force load and open flame non-uniform heat load to the specimen. Send the force or displacement loading command of the test substructure boundary loading extracted in step S6 to the self-balancing loading reaction frame through the LSTM-PID intelligent control system. Simultaneously collect feedback data such as force and displacement and transmit it back to the calculation server through TCP communication. Record and plot the stress-strain curve of the test substructure under thermo-mechanical coupling in real time.

[0118] S8 initiates a time delay compensator based on LSTM for dynamic compensation, which predicts the time delay and issues control commands in advance to gradually reduce the time delay error.

[0119] Specifically, based on the initial time delay compensation in step S5, the LSTM-based time delay compensator initiates a dynamic compensation mechanism: First, it calculates the time difference Δt1 between the issuance of commands for the force, burner opening, and fan speed and angle in analysis step T1 and the arrival of the sensor measurements at the target values; the LSTM-based time delay compensator predicts the time delay Δt2 that will occur in the next analysis step, and issues the command Δt2 in advance during the loading of the second analysis step T2; the time delay Δt is calculated and recorded after each round of loading. iThe weights of the LSTM network are dynamically updated to gradually reduce the time delay error until the time delay error is less than 1s throughout the entire experimental period.

[0120] S9. Repeat steps S6 to S8 until the deformation or stress of a certain analysis specimen exceeds the limit, then end the loading.

[0121] S10 dynamically displays the stress, deformation, and temperature field distribution under a full-bridge fire using 3D animation software. The 3D animation software can be 3D visualization software such as Unity3D.

[0122] Specifically, the Unity3D real-time engine can bind LSTM prediction data or sensor measured data to the mesh nodes of the 3D model, and dynamically visualize the stress, deformation and temperature field distribution changes of the entire bridge under fire based on the numerical-color mapping algorithm; dynamically update the load-displacement curve, temperature-time curve and failure warning threshold panel.

[0123] According to one embodiment of the present invention, in step S1, the test substructure includes a main cable, a boom or sling section, and connecting components within the section; wherein, as Figure 5 As shown, when the prototype of the specimen is the main cable of a suspension bridge, the large-diameter cable test model disclosed in Chinese Invention Patent No. CN120064553A, "A Large-Diameter Cable Test Model Real Fire Test Device and Test Method", is selected. The two ends of the main cable pass through the free-rotation anchoring end and the free-rotation loading end of the same height on the self-balancing loading reaction frame, and are then anchored with anchors. As needed, a large-scale equivalent model with a specimen diameter not less than 1:4 compared with the actual bridge main cable diameter is selected.

[0124] The boom or sling is a full-scale model. When testing a full-scale boom or sling, its upper end is fixed to the test model of the main cable by a cable clamp, and its lower end is connected to a hydraulic loading device embedded in a self-balancing loading reaction frame. In this embodiment, the hydraulic loading device is a second through-hole jack 71, which is embedded in the internal cavity of the box-shaped bottom beam 46. The boom passes through the second through-hole jack 71 and is then anchored.

[0125] According to an embodiment of the present invention, in step S2, constructing a rapid generation model of the fire temperature field includes:

[0126] S21. Construct a GAN network, which includes a generator and a discriminator; among which...

[0127] The generator takes fire source power, normalized location coordinates (x,y), and wind speed as input conditions. It extracts local fire dynamic features through a two-layer spatiotemporal convolutional LSTM branch and embeds residual terms of the heat diffusion equation to construct a physical constraint branch. The two features are fused through a fully connected layer to output a normalized temperature field matrix, which is then denormalized to the actual temperature-time curve. The two branches of the two-layer spatiotemporal convolutional LSTM have 64 and 128 channels, respectively, and are activated using LeakyReLU. The output normalized temperature field matrix is ​​activated using Sigmoid and has a value range of [0,1].

[0128] The discriminator takes the generated or real temperature field matrix and fire source parameters (power, location, wind speed) as joint inputs. It extracts spatiotemporal joint features through a three-layer convolutional network and embeds a gradient penalty term to constrain the smoothness of the generated data. The output layer is compressed into a single node by global average pooling and outputs the probability of the temperature field's authenticity. At the same time, the residual of the heat diffusion equation is calculated through residual connections as a criterion for physical rationality. The three layers of the convolutional network have 128, 64, and 32 channels, respectively, and use ReLU activation. The output layer uses Sigmoid activation.

[0129] The residual of the thermal diffusion equation is the difference between the generated temperature field and the predicted value of the theoretical equation, and is calculated using the following formula:

[0130]

[0131] In the formula, To generate a temperature field, For ambient temperature, To generate the spatial second derivative of the temperature field, This is the balance term between the heat source and the ambient temperature; The L2 norm is used to measure the magnitude of the residuals.

[0132] S22 uses fire dynamics software (such as FDS) to simulate and generate full-bridge three-dimensional non-uniform temperature field data, covering combinations of different fire source power, location, and wind speed. It extracts spatiotemporal temperature distribution data to construct a training dataset, with fire source parameters as input and a gridded temperature field matrix as output. The reliability of the data is verified through experimental substructure tests, reproducing the target temperature field from the fire dynamics software. The measured and simulated temperature gradient distributions are compared to ensure an error range of ≤5%.

[0133] S23, Offline training of a rapid fire temperature field generation model based on GAN network, and verification data based on fire dynamics software simulation and physical experiment.

[0134] S24. The trained fire temperature field rapid generation model is deployed to the computing server and bound to the fire source parameter input interface; the input targets are fire source power, location, and wind speed, generating the target temperature-time curve. The architecture of the GAN-based fire temperature field rapid generation model neural network is as follows: Figure 4 As shown.

[0135] According to an embodiment of the present invention, in step S2, the fast numerical substructure calculation model based on LSTM is a multi-scale gated residual LSTM neural network prediction model. This model is constructed by introducing a parallel multi-branch structure, dynamic feature selection gates, and residual cross-layer connections on top of LSTM. It is used to predict the evolution process under a full-bridge fire based on measured values ​​from multi-dimensional monitoring sensors during the experimental substructure test and to output loading boundary forces or displacement commands in real time. The specific steps for training the fast numerical substructure calculation model based on LSTM are as follows:

[0136] S31, Construct a multi-scale gated residual LSTM neural network, which includes a dual-branch LSTM unit, namely a first-branch LSTM unit and a second-branch LSTM unit. The specific construction process includes:

[0137] S311, the input layer receives historical time-series data, and the input features include location spatial coordinates, temperature field, boundary displacement, cable force and wind speed;

[0138] S312 is a parallel dual-branch LSTM unit. The first branch LSTM unit processes the original time-series data with a step size of 1, 128 hidden units, and a tanh activation function, and is used to extract short-term local mutation features. The second branch LSTM unit processes downsampled time-series data with a step size of 5, 64 hidden units, and a tanh activation function, and is used to extract long-term global evolution features. Both the first branch LSTM unit and the second branch ISTM unit contain a first-layer LSTM and a second-layer LSTM.

[0139] S313, the dynamic feature selection gate calculates branch weights using the sigmoid function, and the fusion formula is:

[0140]

[0141] Among them, h final h is the hidden state vector after fusing the two branch LSTM units. branch1 h is the output hidden state vector of the first branch LSTM unit. branch2 Let g be the output hidden state vector of the second branch LSTM unit. t The meaning is dynamic gating weight coefficient;

[0142] S314, residual block cross-layer connections alleviate gradient vanishing by superimposing the outputs and inputs of the first and second LSTM layers through skip connections, as shown in the formula:

[0143]

[0144] in, Let be the hidden state vector of the LSTM at time step t in the l-th layer. Let be the hidden state vector of the previous time step at time step t in layer l. Let [x, y, z] be the spatial coordinate vector at time step t. t , Let be the hidden state vector at time step t in layer (l-1). For the LSTM unit of layer l, l =1,2;

[0145] S315, a hybrid activation function, uses the sigmoid function for the forget gate and input gate, the tanh function for the candidate memory unit, and the Swish function f(x)=x·σ(x) for the output gate;

[0146] S316, a fully connected output layer, outputs the predicted boundary force or displacement values ​​for the next time step;

[0147] S32, Constructing a dataset for a fast computation model of numerical substructures based on LSTM, specifically including the following steps:

[0148] S321 uses fire dynamics software to simulate different fire scenarios, calculates the three-dimensional non-uniform temperature field distribution of the entire bridge for different combinations of fire source power, fire source location and wind speed, and extracts the temperature rise-time curves of key parts.

[0149] S322, based on finite element analysis software (such as ABAQUS), establishes a thermal finite element model of the whole bridge and substructure. The temperature field output by the fire dynamics software is interpolated to the finite element mesh nodes according to the time step. The thermal expansion effect is calculated by first applying temperature load through sequential coupling analysis method, and then mechanical load is superimposed to perform structural response analysis, outputting displacement, cable force and strain data.

[0150] S323, fabricate specimens corresponding to the experimental substructure, reproduce the standard temperature field of the fire dynamics software in the laboratory, apply equivalent mechanical load through a through-hole jack, measure displacement and cable force data, and compare with the simulation results of the finite element analysis software to verify the reliability of the finite element model;

[0151] S324 processes the verified simulation data according to the input requirements of multi-scale gated residual LSTM to construct a time series dataset. The input features include temperature, displacement, cable force, and wind speed at key locations, and the output features are the displacement and cable force at the next time step.

[0152] S33, offline pre-training, uses a sliding time window to split the dataset, the loss function is Huber Loss (δ=1.5), the optimizer is Nadam, the initial learning rate is 0.001, and it decays by 50% every 10 epochs; the first branch LSTM unit is input with the original step size data, the second branch LSTM unit is input with downsampled data, and the dynamic feature selection gate weights are optimized through backpropagation of historical errors.

[0153] S34, online incremental learning, dynamically adapts to changes in fire conditions by capturing sensor data in real time, fine-tuning the parameters of the model's downsampling branch and constraining the weight change amplitude to ≤0.01, thereby improving prediction accuracy while maintaining model stability;

[0154] S35 performs real-time prediction and control during the real-time mixed experiment. The specific process is as follows:

[0155] S351, taking displacement loading as an example, converts the displacement of the next time step into a loading displacement command according to the displacement equivalence principle, the formula is:

[0156]

[0157] In the formula, D load For displacement command, D pred L represents the displacement at the next time step. phy L is the actual geometric span of the physical specimen. num For the equivalent span of the numerical model, L represents the large-scale model. phy =L num ;

[0158] S352, driven by an LSTM-PID intelligent control system, uses a self-balancing loading reaction frame. The formula for calculating the control quantity is:

[0159]

[0160] In the formula, γ is the feedforward gain coefficient, which is dynamically adjusted through fuzzy logic;

[0161] S353, Dynamic Gating Weights g t Mapping to PID proportional coefficient K p To achieve adaptive loading;

[0162] S36 iteratively updates the thermal-displacement / force boundary conditions until the specimen deformation or stress exceeds the limit, outputting the failure mode and critical threshold. The architecture of the multi-scale gated residual LSTM neural network is as follows: Figure 2 As shown.

[0163] According to one embodiment of the present invention, in step S3, the vehicle-fire-wind coupling test platform includes a self-balancing loading reaction frame 4, a multi-module T-shaped intelligent dynamic control combustion system, and a variable speed angle controllable fan system 1; wherein,

[0164] The multi-module T-shaped flame envelope combustion system includes a horizontal linear combustion module 2 and a vertical annular combustion module 3. The horizontal linear combustion module 2 includes several frame-type burners 21 arranged linearly and equipped with vertical flame tubes. The frame-type burners are suspended and fixed in the middle test area of ​​the self-balancing loading reaction frame 4 by several steel wire ropes. The vertical annular combustion module 3 includes several layers of annular burners arranged vertically. The main body of the annular burner is an annular steel pipe 31. The annular steel pipes 31 are fixedly connected by connecting pipes. The annular steel pipes 31 are also equipped with several flame nozzles 32 that can adjust the flame angle.

[0165] The variable speed and angle controllable fan system 1 includes several high dynamic axial flow fans 11 arranged in a matrix topology on one side of the self-balancing loading reaction frame 4 and capable of adjusting the wind direction and angle. The high dynamic axial flow fans 11 are connected to ultrasonic wind speed and direction sensors and pressure sensors to provide real-time feedback of wind field parameters to the LSTM-PID intelligent control system for dynamic error control. The high dynamic axial flow fans are also equipped with integrated frequency converters to achieve continuous wind speed adjustment and precise tilt angle control.

[0166] According to an embodiment of the present invention, in step S5, the input layer of the LSTM-PID intelligent control system includes temperature field data, wind speed, angle, burner valve opening degree, and historical error; the hidden layer includes 54-64 neurons, wherein the number of neurons is preferably 64; and the output layer outputs the burner valve opening degree increment Δu. gas and fan frequency f fan Control commands are generated according to the following formula:

[0167]

[0168] In the formula, Δu gas f is the increment of the burner valve opening. fan α represents the fan frequency; β and α are coupling coefficients, which are dynamically adjusted according to real-time operating conditions. It is an existing PID control equation, which is improved by adding the burner valve opening increment Δu to the original PID control method. gas With the frequency f of the fan fan The item implements the calculation of control commands;

[0169] Finally, control commands are sent to the burner and fan inverters to synchronously trigger wind-fire coupled loading, achieving dynamic reproduction of the fire scenario. The architecture of the LSTM neural network in the LSTM-PID intelligent control system is as follows: Figure 7 As shown.

[0170] According to an embodiment of the present invention, in step S5, the step of generating control instructions is as follows:

[0171] S51 uses sensors to collect temperature field data, wind speed angle data and historical sequence of burner valve opening in real time to construct a multi-dimensional time series input vector;

[0172] S52, the input vector is fed into the pre-trained LSTM prediction model to predict the burner valve opening increment and fan frequency at the next moment. The LSTM prediction model updates the weights online through a sliding time window, and the prediction error converges to <5%.

[0173] S53 uses the predicted burner valve opening increment and fan frequency as feedforward correction values, which are then superimposed on the feedback output of the PID controller to generate the final control command.

[0174] According to an embodiment of the present invention, in step S8, the LSTM time delay compensator is mainly an LSTM-based neural network containing a hidden layer of 16-32 neurons, wherein the number of neurons is preferably 32, and the input layer nodes contain the historical time delay sequence Δt. i The system's real-time load rate, ambient temperature gradient, and ambient wind speed gradient, where i = (1, 2, ..., n), are mapped by a fully connected layer to the predicted time delay Δt for the next analysis step. i+1 Specifically, it includes the following steps:

[0175] S61, Construct a time-delay dataset and collect historical time-delay sequences Δt i The system's real-time load rate, ambient temperature gradient, and wind speed gradient data are used to construct an input vector, which is then input into an LSTM-based neural network. Here, Δt... i The time delay between the issuance of a control command and the achievement of the sensor's measured value.

[0176] S62, offline training and online optimization, uses data samples from the preloading stage to train an LSTM-based time-delay compensation model with the loss function being the mean absolute error; during the formal loading, after each analysis step, the data stream is intercepted based on a sliding window of 5-10 steps, and the network parameters are updated online through the Adam optimizer, where the preferred step size is 10.

[0177] S63, dynamic time delay compensation, sends control commands Δt2 time in advance before loading in the second analysis step, and records the actual time delay Δt. real △t2 refers to the time difference between the issuance of the instruction in the first analysis step and the loading being completed;

[0178] S64, Error Correction and Convergence, Based on Δt i-real With △t i To correct the deviation, the weights of the LSTM-based time-delay compensation model are adjusted using the gradient descent method until the time-delay error of multiple consecutive analysis steps is ≤1 second; Δt i-real This refers to the actual time delay measured in each analysis step, i = (1, 2, ..., n). The architecture of the neural network based on the LSTM time delay compensator is as follows: Figure 8 As shown.

[0179] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A real-time hybrid test method for fire in large-scale substructures of bridge cables, characterized in that, include: S1, divide the cable-bearing bridge to be analyzed into experimental substructure and numerical substructure; S2, deploy the trained fire temperature field rapid generation model and the LSTM-based numerical substructure rapid calculation model to the computing server. In step S2, constructing the fire temperature field rapid generation model includes: S21. Construct a GAN network, which includes a generator and a discriminator; among them... The generator takes fire source power, normalized location coordinates and wind speed as input conditions, extracts local fire dynamic features through a two-layer spatiotemporal convolution LSTM branch, and embeds the residual terms of the heat diffusion equation to construct a physical constraint branch; after the two features are fused by a fully connected layer, the normalized temperature field matrix is ​​output, and finally denormalized into the actual temperature-time curve. The discriminator takes the generated or real temperature field matrix and fire source parameters as joint inputs, extracts spatiotemporal joint features through a three-layer convolutional network, and embeds gradient penalty terms to constrain the smoothness of the generated data; the output layer is compressed into a single node through global average pooling, outputting the probability of the authenticity of the temperature field, and at the same time calculates the residual of the heat diffusion equation through residual connection as a criterion for physical rationality. The residual of the thermal diffusion equation is the difference between the generated temperature field and the predicted value of the theoretical equation. The calculation formula is as follows: In the formula, To generate a temperature field, For ambient temperature, To generate the spatial second derivative of the temperature field, This is the balance term between the heat source and the ambient temperature; The L2 norm is used to measure the magnitude of the residuals. S22: Based on fire dynamics software simulation, three-dimensional non-uniform temperature field data of the entire bridge is generated. Spatiotemporal temperature distribution data is extracted, and a training dataset is constructed. The reliability of the data is verified through experimental substructure tests. The measured and simulated temperatures are compared to ensure that the error range is ≤5%. S23, Offline training of a rapid fire temperature field generation model based on GAN network, and verification data based on fire dynamics software simulation and physical experiment. S24, quickly deploy the trained fire temperature field generation model to the computing server; S3, fabricate the test specimens corresponding to the experimental substructure, and build a vehicle-fire-wind coupling test platform; S4, Deploy a fire scene environmental monitoring system; S5. The specimen is preloaded on the vehicle-fire-wind coupling test platform, and the initial time delay of each system is obtained and input into the LSTM-PID intelligent control system for feedforward compensation. S6, the computing server receives data from the fire environment monitoring system of the monitoring system; the numerical substructure fast calculation model based on LSTM predicts the boundary displacement of the next time step in real time, and converts the predicted boundary displacement into force or displacement loading command of physical test according to the displacement equivalence principle. S7. Apply force load and open flame non-uniform heat load to the specimen. Send the force or displacement loading command of the test substructure boundary extracted in step S6 to the self-balancing loading reaction frame through the LSTM-PID intelligent control system and synchronously collect data and transmit it back to the calculation server. S8 activates the LSTM-based time delay compensator for dynamic compensation, which predicts the time delay and issues control commands in advance to gradually reduce the time delay error. S9. Repeat steps S6 to S8 until the deformation or stress of the specimen exceeds the limit in a certain analysis step, then end the loading. S10 uses 3D animation software to dynamically display the stress, deformation, and temperature field distribution under a full-bridge fire.

2. The real-time hybrid test method for fire in large-scale substructures of bridge cables according to claim 1, characterized in that, In step S1, When the prototype of the specimen is the main cable, its two ends pass through the free rotation anchorage end and the free rotation loading end at the same height on the self-balancing loading reaction frame, and are then anchored using anchorages. As needed, a large-scale equivalent model with a specimen diameter not less than 1:4 compared to the actual bridge main cable diameter is selected. When it is necessary to test a full-scale boom or sling, its upper end is fixed to the test model of the main cable by a cable clamp, and its lower end is connected to the hydraulic loading device embedded in the self-balancing loading reaction frame.

3. The real-time hybrid test method for fire in large-scale substructures of bridge cables according to claim 1, characterized in that, In step S2, the LSTM-based numerical substructure fast calculation model is a multi-scale gated residual LSTM neural network prediction model. This model is constructed by introducing a parallel multi-branch structure, dynamic feature selection gates, and residual cross-layer connections on top of LSTM. It is used to predict the evolution process under a full-bridge fire based on measured values ​​from multi-dimensional monitoring sensors during the experimental substructure test and to output loading boundary forces or displacement commands in real time. The specific steps for training the LSTM-based numerical substructure fast calculation model are as follows: S31, Construct a multi-scale gated residual LSTM neural network, wherein the multi-scale gated residual LSTM neural network includes a dual-branch LSTM unit, namely a first branch LSTM unit and a second branch LSTM unit. S32, Construct a dataset for a fast numerical substructure calculation model based on LSTM, specifically including: using fire dynamics software to simulate the three-dimensional non-uniform temperature field distribution within the full-bridge space under different fire scenarios, and extracting temperature-time curves; establishing a thermo-mechanical coupling finite element model of the full-bridge and the experimental substructure based on finite element analysis software; fabricating specimens corresponding to the experimental substructure, subjecting the specimens to non-uniform temperature field thermo-mechanical coupling loading, and verifying the accuracy of the specimen thermo-mechanical coupling model; using the verified simulation data to train a multi-scale gated residual LSTM neural network model; S33, offline pre-training, uses a sliding time window to segment the dataset. The first branch LSTM unit is input with the original step size data, and the second branch LSTM unit is input with the downsampled data. The dynamic feature selection gate weights are optimized through backpropagation of historical errors. S34, online incremental learning, dynamically adapts to changes in fire conditions by capturing sensor data in real time, fine-tuning the parameters of the model's downsampling branch and constraining the weight change amplitude to ≤0.01; S35 performs real-time prediction and control during real-time mixed testing; S36 iteratively updates the thermal-displacement / force boundary conditions until the specimen deformation or stress exceeds the limit, and outputs the failure mode and critical threshold.

4. The real-time hybrid test method for fire in large-scale substructures of bridge cables according to claim 3, characterized in that, Step S31 includes: S311, the input layer receives historical time-series data, and the input features include location spatial coordinates, temperature field, boundary displacement, cable force and wind speed; S312 is a parallel dual-branch LSTM unit, where the first branch LSTM unit processes the original step-size time series data to extract short-term local mutation features; the second branch LSTM unit processes downsampled time series data to extract long-term global evolution features; both the first branch LSTM unit and the second branch ISTM unit contain a first-layer LSTM and a second-layer LSTM. S313, the dynamic feature selection gate calculates branch weights using the sigmoid function, and the fusion formula is: Among them, h final h is the hidden state vector after fusing the two branch LSTM units. branch1 h is the output hidden state vector of the first branch LSTM unit. branch2 Let g be the output hidden state vector of the second branch LSTM unit. t The meaning is dynamic gating weight coefficient; S314, residual block cross-layer connections alleviate gradient vanishing by superimposing the outputs and inputs of the first and second LSTM layers through skip connections, as shown in the formula: in, Let be the hidden state vector of the LSTM at time step t in the l-th layer. Let be the hidden state vector of the previous time step at time step t in layer l. Let [x, y, z] be the spatial coordinate vector at time step t. t , Let be the hidden state vector at time step t in layer (l-1). For the LSTM unit of layer l, l =1,2; S315, a hybrid activation function; S316, a fully connected output layer, outputs the predicted boundary force or displacement values ​​for the next time step.

5. The real-time hybrid test method for fire in large-scale substructures of bridge cables according to claim 3, characterized in that, Step S35 includes: S351, taking displacement loading as an example, converts the displacement of the next time step into a loading displacement command according to the displacement equivalence principle, the formula is: In the formula, D load For displacement command, D pred L represents the displacement at the next time step. phy L is the actual geometric span of the physical specimen. num For the equivalent span of the numerical model, L represents the large-scale model. phy =L num ; S352, driven by an LSTM-PID intelligent control system, uses a self-balancing loading reaction frame. The formula for calculating the control quantity is: In the formula, γ is the feedforward gain coefficient, which is dynamically adjusted through fuzzy logic; S353 dynamically maps the gated weights to the PID proportional coefficients to achieve adaptive loading.

6. The real-time hybrid test method for fire in large-scale substructures of bridge cables according to claim 1, characterized in that, In step S3, the vehicle-fire-wind coupling test platform includes a self-balancing loading reaction frame, a multi-module T-shaped intelligent dynamic control combustion system, and a variable speed angle controllable fan system; wherein, The multi-module T-shaped flame envelope combustion system includes a horizontal linear combustion module and a vertical annular combustion module. The horizontal linear combustion module includes several frame-type burners arranged linearly and equipped with vertical flame tubes. The frame-type burners are suspended and fixed in the middle test area of ​​a self-balancing loading reaction frame by several steel wire ropes. The vertical annular combustion module includes several layers of annular burners arranged vertically. The main body of the annular burner is an annular steel pipe. The annular steel pipes are fixedly connected by connecting pipes. The annular steel pipes are also equipped with several flame nozzles that can adjust the flame angle. The variable speed and angle controllable fan system includes several highly dynamic axial flow fans arranged in a matrix topology on one side of the self-balancing loading reaction frame and capable of adjusting the wind direction and angle. The highly dynamic axial flow fans are connected to ultrasonic wind speed and direction sensors and pressure sensors to provide real-time feedback of wind field parameters to the LSTM-PID intelligent control system for dynamic error control. The highly dynamic axial flow fans are also equipped with integrated frequency converters to achieve continuous wind speed adjustment and precise tilt angle control.

7. The real-time hybrid test method for fire in large-scale substructures of bridge cables according to claim 1, characterized in that, In step S5, the input layer of the LSTM-PID intelligent control system includes temperature field data, wind speed, angle, burner valve opening degree, and historical errors; the hidden layer contains 54-64 neurons; and the output layer outputs the burner valve opening degree increment and fan frequency. The control command is generated according to the following formula: In the formula, Δu gas f is the increment of the burner valve opening. fan α represents the fan frequency; β are coupling coefficients, which are dynamically adjusted according to real-time operating conditions. Finally, control commands are sent to the burner and fan frequency converter to synchronously trigger wind-fire coupling loading, thereby realizing the dynamic reproduction of the fire scene.

8. The real-time hybrid test method for fire in large-scale substructures of bridge cables according to claim 7, characterized in that, In step S5, the control command generation steps are as follows: S51 uses sensors to collect temperature field data, wind speed angle data and historical sequence of burner valve opening in real time to construct a multi-dimensional time series input vector; S52, the input vector is fed into the pre-trained LSTM prediction model to predict the burner valve opening increment and fan frequency at the next moment. The LSTM prediction model updates the weights online through a sliding time window, and the prediction error converges to <5%. S53 uses the predicted burner valve opening increment and fan frequency as feedforward correction values, which are then superimposed on the feedback output of the PID controller to generate the final control command.

9. The real-time hybrid test method for fire in large-scale substructures of bridge cables according to claim 1, characterized in that, In step S8, the time delay compensator is mainly an LSTM-based neural network containing a hidden layer of 16-32 neurons, and the input layer nodes contain the historical time delay sequence Δt. i The system's real-time load rate, ambient temperature gradient, and ambient wind speed gradient, where i = (1, 2, ..., n), are mapped by a fully connected layer to the predicted time delay Δt for the next analysis step. i+1 Specifically, it includes the following steps: S81, Construct a time-delay dataset and collect historical time-delay sequences Δt. i The system's real-time load rate, ambient temperature gradient, and wind speed gradient data are used to construct an input vector, which is then input into an LSTM-based neural network. Here, Δt... i The time delay between the issuance of a control command and the achievement of the sensor's measured value. S82, offline training and online optimization, uses data samples from the preloading stage to train an LSTM-based time-delay compensation model with the loss function being the mean absolute error; during the formal loading, the network parameters are updated online through the optimizer after each analysis step. S83, dynamic time delay compensation, sends control commands Δt2 time in advance before loading in the second analysis step, and records the actual time delay Δt. real △t2 refers to the time difference between the issuance of the instruction in the first analysis step and the loading being completed; S84, Error Correction and Convergence, Based on Δt i-real With △t i To correct the deviation, the weights of the LSTM-based time-delay compensation model are adjusted using the gradient descent method until the time-delay error of multiple consecutive analysis steps is ≤1 second; Δt i-real The time delay actually measured in each analysis step, i=(1,2,...,n).

Images