Bridge cable large-scale substructure fire real-time mixing test method

Abstract

The invention discloses a real-time mixed fire test method for a large-scale substructure of a bridge cable, and belongs to the technical field of bridge fire resistance test. A full bridge is divided into a physical test substructure and an LSTM numerical substructure, the numerical substructure is constructed based on multi-scale gating residual LSTM, and second-level high-precision prediction of the boundary force or displacement of a cable system in a fire disaster is achieved through a parallel multi-branch network, a dynamic feature selection gate and residual cross-layer connection. By combining an LSTM-PID intelligent control system, displacement / stress and a temperature field are converted into loading signals according to a prediction instruction, a time delay compensation mechanism is synchronously integrated, a center hole jack, a combustor and a fan are driven in real time, a non-uniform wind-fire coupling scene is reproduced, and a full-bridge three-dimensional visual fire response is formed. According to the method, the speed bottleneck of traditional finite element calculation is broken through, second-level thermal coupling interaction is achieved, and a high-precision and high-reliability test method and platform are provided for large-span bridge fire resistance design.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge fire resistance performance testing, and in particular to a real-time hybrid test method for large-scale substructure fires of bridge cables. Background Art

[0002] In modern transportation networks, long-span bridges serve as core arteries connecting regional economies. Their structural safety in the event of a fire is directly related to public safety and the sustainable operation of this vital transportation network. Fire not only softens steel but also, through non-uniform temperature fields, triggers internal force redistribution, localized buckling, and even the cascading collapse of the entire bridge. This coupled thermal-mechanical catastrophic mechanism poses significant challenges to traditional fire-resistant design methods.

[0003] Current fire resistance research methods have significant flaws: traditional tests rely on uniform temperature heating in combustion furnaces or scaled models, making it difficult to reproduce the coupling effects of non-uniform fire loads, wind fields, and force loads. Although hybrid simulation technology is mature in the field of earthquake resistance, it is still in the exploratory stage in fire scenarios. Existing numerical tools, such as OpenSees and ABAQUS, can be used for modeling and thermal-mechanical coupling analysis, but their computational efficiency is low and cannot meet the needs of real-time hybrid testing. Some foreign studies have attempted to combine experimental substructures with numerical models, but due to computational delays and scale effects, high-fidelity analysis of full-scale structures cannot be achieved. Domestically, progress has been slow due to the lack of efficient numerical substructure calculation models and test platforms.

[0004] The particularities of fire resistance assessment for long-span bridges further exacerbate technical bottlenecks. First, the collaborative working mechanism of the cable system, pylons, and stiffening beams exhibits strong nonlinear characteristics during fire, making local component testing unable to reveal the failure path and evolution of the overall structure under fire. Second, the structural stiffness degradation and cable force redistribution caused by high fire temperatures exhibit significant spatiotemporal asynchrony, necessitating real-time coupling of thermal-mechanical boundary conditions. However, existing hybrid test systems lack the accuracy to meet the requirements of long-cycle loading. Third, existing systems rely on traditional finite element calculations, making it difficult to overcome the "speed-accuracy" trade-off, resulting in significant error accumulation under long-cycle loading. Furthermore, despite some research efforts to improve efficiency through GPU acceleration and model reduction, computational speeds still fall short of the requirements of real-time hybrid experiments.

[0005] Therefore, in response to the urgent need for fire safety assessment of large-span bridges, it is urgent to develop a hybrid test method that integrates large-scale substructure physical tests, fast calculation numerical models based on neural networks, and intelligent time-delay compensation, so as to break through the spatial scale limitations and real-time bottlenecks of traditional methods and provide technical support for bridge fire resistance design that combines scientific rigor and physical test operability. Summary of the Invention

[0006] The present invention aims to address, at least to some extent, one of the technical problems in the related art. To this end, the first objective of the present invention is to propose a real-time hybrid testing method for large-scale substructure fires in bridge cables. This method overcomes the speed bottleneck of traditional finite element calculations, achieves second-level thermal-mechanical coupling interaction, and provides a highly accurate and reliable testing method for the fire-resistant design of long-span bridges.

[0007] To achieve the above objectives, the first embodiment of the present invention proposes a real-time hybrid test method for large-scale substructure fires of bridge cables, comprising: S1, divide the cable-supported 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; S3, fabricate the test specimens corresponding to the test substructure and build the vehicle-fire-wind coupling test platform; S4, deploy fire scene environment monitoring system; S5, preload the specimen on the vehicle-fire-wind coupling test platform, obtain the initial time delay of each system, and input it into the LSTM-PID intelligent control system for feedforward compensation; In step S6, the computing server receives data from the fire environment monitoring system of the monitoring system; the LSTM-based numerical substructure fast computing model predicts the boundary displacement of the next time step in real time, and converts the predicted boundary displacement into force or displacement loading instructions for physical testing according to the displacement equivalence principle; S7, applying force load and open flame non-uniform thermal load on the specimen, sending the force or displacement loading instructions 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 collecting data and transmitting it back to the computing server; S8, start the LSTM-based time delay compensator for dynamic compensation, issue control instructions in advance by predicting the time delay and gradually reduce the time delay error; S9, repeating steps S6 to S8 until the deformation or stress of the specimen in a certain analysis step exceeds the limit, and then ending the loading; S10, using 3D animation software to dynamically display the stress, deformation and temperature field distribution of the entire bridge under fire.

[0008] Preferably, in step S1, when the prototype of the specimen is a main cable, its two ends are successively passed through the free-rotating anchoring end and the free-rotating loading end at the same height on the self-balancing loading reaction frame and then anchored using anchors respectively. A large-scale equivalent model is selected according to the need, with the diameter of the specimen being no less than 1:4 to the diameter of the actual bridge main cable; When a full-scale boom or sling needs to be tested, its upper end is fixed to the test model of the main cable through a cable clamp, and its lower end is connected to the hydraulic loading device embedded in the self-balancing loading reaction frame.

[0009] Preferably, in step S2, constructing the fire temperature field rapid generation model includes: S21, build a GAN network, which includes a generator and a discriminator; The generator uses fire source power, normalized position coordinates, and wind speed as input conditions, extracts local fire dynamic features through a double-layer spatiotemporal convolution LSTM branch, and embeds the residual term of the heat diffusion equation to construct a physical constraint branch. After the dual-path features are fused through a fully connected layer, a normalized temperature field matrix is ​​output, which is finally denormalized into an actual temperature-time curve. The two branches of the double-layer spatiotemporal convolution LSTM have 64 and 128 channels respectively, and are activated using LeakyReLU. The output normalized temperature field matrix is ​​activated using Sigmoid, with a value range of [0,1]. The discriminator takes the generated or real temperature field matrix and the fire source parameters as joint input, 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 through global average pooling, outputs the authenticity probability of the temperature field, and calculates the residual of the heat diffusion equation as a physical rationality criterion through residual connection. The number of channels of the three-layer convolutional network is 128, 64, and 32 respectively, and ReLU is used for activation; the output layer uses Sigmoid activation; The residual of the heat diffusion equation is the difference between the generated temperature field and the predicted value of the theoretical equation, and the calculation formula is: Where, To generate the temperature field, is the ambient temperature, To generate the spatial second-order derivative of the temperature field, is the balance term between the heat source and the ambient temperature; is the L2 norm, which is used to measure the size of the residual; S22, based on fire dynamics software simulation, generates three-dimensional non-uniform temperature field data for the entire bridge, covering different combinations of fire source power, location, and wind speed. It extracts spatiotemporal temperature distribution data and constructs a training dataset, with fire source parameters as input and a gridded temperature field matrix as output label. It verifies data reliability through experimental substructure testing, reproduces the target temperature field of the fire dynamics software, and compares the measured and simulated temperature gradient distributions to ensure an error range of ≤5%. S23, offline training of a rapid generation model of fire temperature field based on a GAN network, and verification data based on fire dynamics software simulation and physical experiments; S24: Deploy the trained fire temperature field rapid generation model to the computing server and bind it to the fire source parameter input interface; input targets are fire source power, location, and wind speed, and generate a target temperature-time curve.

[0010] Preferably, in step S2, the LSTM-based numerical substructure rapid calculation model is a multi-scale gated residual LSTM neural network prediction model. The multi-scale gated residual LSTM neural network prediction model is constructed by introducing a parallel multi-branch structure, a dynamic feature selection gate, and a residual cross-layer connection on the basis of LSTM, and is used to predict the evolution process of the full bridge under fire according to the measured values ​​of the multi-dimensional monitoring sensor in the test substructure test and output the loading boundary force or displacement instruction in real time. The specific steps of training the LSTM-based numerical substructure rapid calculation model are as follows: S31, building a multi-scale gated residual LSTM neural network, wherein the multi-scale gated residual LSTM neural network includes a two-branch LSTM unit, i.e., a first-branch LSTM unit and a second-branch LSTM unit; S32: Construct a dataset for a fast computational model of a numerical substructure based on an LSTM. Specifically, the dataset includes: 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 thermal-mechanical coupling finite element model of the full bridge and the test substructure using finite element analysis software; fabricating a specimen corresponding to the test substructure, subjecting the specimen to thermal-mechanical coupling loading with a non-uniform temperature field, and verifying the accuracy of the specimen thermal-mechanical coupling model; and 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 split the dataset, uses Huber Loss as the loss function, uses Nadam as the optimizer, and has an initial learning rate of 0.001, which decays by 50% every 10 epochs. The first branch LSTM unit inputs the original step data, and the second branch LSTM unit inputs the downsampled data. The dynamic feature selection gate weights are optimized by backpropagation of historical errors. S34, online incremental learning, dynamically adapts to fire changes by intercepting sensor data in real time, fine-tuning only the model downsampling branch parameters and constraining the weight change amplitude to ≤0.01; S35, real-time prediction and control during real-time mixing experiments; 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.

[0011] Preferably, step S31 includes: S311, the input layer receives historical time series data, and the input features include position spatial coordinates, temperature field, boundary displacement, cable force and wind speed; S312, parallel dual-branch LSTM unit, where the first branch LSTM unit processes the original step-length time series data with a step size of 1, 128 hidden units, and an activation function of tanh, and is used to extract short-term local mutation features; the second branch LSTM unit processes the downsampled time series data with a step size of 5, 64 hidden units, and an activation function of tanh, and is used to extract long-term global evolution features; the first branch LSTM unit and the second branch ISTM unit both include a first layer LSTM and a second layer LSTM; S313, the dynamic feature selection gate calculates the branch weight through the sigmoid function, and the fusion formula is: Among them, h final is the hidden state vector after the fusion of the two branch LSTM units, h branch1 is the output hidden state vector of the first branch LSTM unit, h branch2 is the output hidden state vector of the second branch LSTM unit, g t The meaning of is the dynamic gating weight coefficient; S314, the residual block cross-layer connection alleviates the gradient disappearance, and the output and input of the first layer LSTM and the second layer LSTM are superimposed through the jump connection. The formula is: in, is the hidden state vector of the l-th layer LSTM at time step t, is the hidden state vector of the previous time step of time step t at layer l, is the spatial coordinate vector [x,y,z] at time step t t , is the hidden state vector at the l-1 layer at time step t, is the LSTM unit of layer l, l =1,2; S315, hybrid activation function, the forget gate and input gate use the sigmoid function, the candidate memory unit uses the tanh function, and the output gate uses the Swish function f(x)=x·σ(x); S316, fully connected output layer, outputs the boundary force prediction value or displacement prediction value of the next time step; Preferably, step S35 includes: S351, taking displacement loading as an example, convert the displacement of the next time step into loading displacement instructions according to the displacement equivalence principle. The formula is: Where D load is the displacement instruction, D pred is the displacement of the next time step; Lphy is the actual geometric span length of the physical specimen, L num is the equivalent span length of the numerical model, for large-scale models L phy =L num ; S352, the self-balancing loading reaction frame is driven by the LSTM-PID intelligent control system. The control quantity calculation formula is: Where γ is the feedforward gain coefficient, which is dynamically adjusted through fuzzy logic; S353, dynamic gating weight is mapped to PID proportional coefficient to achieve adaptive loading.

[0012] 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, 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 a plurality of linearly arranged frame-type burners provided with vertical flame tubes, and the frame-type burners are suspended and fixed in the middle test area of ​​the self-balancing loading reaction frame by a plurality of steel wire ropes; the vertical annular combustion module includes a plurality of layers of annular burners arranged in an upper and lower manner, and the main body of the annular burner is annular steel pipe, which is fixedly connected by a connecting pipe and is also provided with a plurality of flame nozzles capable of adjusting the flame spray angle; The variable speed angle controllable fan system includes several high-dynamic axial flow fans arranged in a matrix topology on one side of a self-balancing loading reaction frame and capable of adjusting the wind direction angle; the high-dynamic axial flow fan is connected to an ultrasonic wind speed and direction sensor and a pressure sensor to provide real-time feedback of wind field parameters to the LSTM-PID intelligent control system for dynamic error regulation; the high-dynamic axial flow fan is also equipped with a fan integrated frequency converter to achieve continuous wind speed adjustment and precise control of the tilt angle.

[0013] Preferably, in step S5, the input layer of the LSTM-PID intelligent control system includes temperature field data, wind speed, angle, burner valve opening and historical error, the hidden layer includes 54-64 neurons, and the output layer outputs the burner valve opening increment Δu gas and fan frequency f fan ; The control instructions are generated according to the following formula: Where Δu gas is the burner valve opening increment, f fan is the fan frequency; α and β are coupling coefficients, which are dynamically adjusted according to the real-time working conditions; Finally, the control instructions are sent to the burner and fan inverter, synchronously triggering the wind-fire coupling loading to achieve dynamic reproduction of the fire scene.

[0014] Preferably, in step S5, the control instruction generation step is: S51, using sensors to collect temperature field data, wind speed angle data and burner valve opening history series in real time to construct an input vector of a multidimensional time series; S52: Input the input vector 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 less than 5%. S53: The predicted burner valve opening increment and fan frequency are used as feedforward correction quantities and added to the feedback output of the PID controller to generate a final control instruction.

[0015] Preferably, in step S8, the main body of the time delay compensator is a neural network based on LSTM, which includes a hidden layer of 16-32 neurons, and the input layer node includes the historical time delay sequence Δt i , system real-time load rate, ambient temperature gradient, and ambient wind speed gradient, where i=(1, 2, ..., n), and the output layer is mapped to the time-delay prediction value Δt for the next analysis step through the fully connected layer. i+1 , specifically including the following steps: S81, build a time-lag data set and collect historical time-lag series Δt i , the data of the three dimensions of system real-time load rate, ambient temperature gradient and wind speed gradient are constructed into an input vector and input into the LSTM-based neural network, where Δt i Refers to the delay time from when the control command is issued to when the sensor's measured value reaches the standard; S82, offline training and online optimization, uses pre-loaded data samples to train an LSTM-based time-delay compensation model, with the loss function being the mean absolute error. During formal loading, after each analysis step, the data stream is intercepted using a sliding window of 5-10 steps, and the network parameters are updated online using the Adam optimizer. S83, dynamic time delay compensation, before the second analysis step loading, send the control command in advance △t2 time, and record the actual time delay △t real ,△t2 refers to the time difference between issuing the instruction in the first analysis step and loading into place; S84, Error Correction and Convergence, Based on Δt i-real and △t i The gradient descent method is used to correct the weights of the LSTM-based time-delay compensation model until the time-delay error of multiple consecutive analysis steps is ≤1 second; Δt i-realRefers to the time lag actually measured in each analysis step, i=(1, 2,...,n).

[0016] Preferably, the method further comprises: dynamically displaying the stress, deformation and temperature field distribution of the entire bridge under fire by means of three-dimensional animation software.

[0017] 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 large-scale substructure fires of bridge cables, which has the following beneficial effects: 1. Breaking through the computational speed bottleneck: Traditional finite element methods suffer from low computational efficiency in fire scenarios, making them inefficient for real-time hybrid testing. This invention significantly improves computational speed by introducing a multi-scale gated residual neural network based on LSTM, achieving second-level thermal-mechanical coupling and resolving the significant error accumulation seen in traditional methods under long-term loading. The fire temperature field rapid generation model, based on an improved generative adversarial network (GAN), generates highly accurate non-uniform temperature field distributions in real time, further enhancing the real-time performance and accuracy of the test.

[0018] 2. High-Precision Prediction and Control: The multi-scale gated residual LSTM network, through its parallel multi-branch structure, dynamic feature selection gates, and residual cross-layer connections, effectively extracts both short-term local mutation characteristics and long-term global evolutionary characteristics of the specimen under fire, enabling high-precision, second-level prediction of boundary forces or displacements. The LSTM-PID intelligent control system combines the predictive power of the LSTM with the stability of the PID controller, dynamically adjusting loading instructions based on real-time feedback data to ensure precise control of the test process.

[0019] 3. Real-time Dynamic Compensation and Time Delay Management: 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 synchronization and accuracy during the test process. This mechanism gradually reduces 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.

[0020] 4. High-fidelity fire scene reproduction: The multi-module T-shaped flame envelope combustion system and variable-speed angle controllable fan system simulate the non-uniform fire scene and wind-fire coupling environment found in real-world fires, achieving high-fidelity reproduction of fire scenes in laboratory environments. The fire scene environment monitoring system, equipped with a variety of sensors (such as armored thermocouples, strain gauges, fiber Bragg grating sensors, and ultrasonic anemometers), monitors the fire scene environment and specimen stress, deformation, and other parameters in real time, providing comprehensive data support for testing.

[0021] 5. Visualization and Real-Time Monitoring: The 3D visualization module binds LSTM-predicted data or sensor-measured data to the 3D model grid nodes, dynamically visualizing the stress, deformation, and temperature distribution changes of the entire bridge under fire, providing intuitive visualization support for the test process. The system also dynamically updates the load-displacement curve, temperature-time curve, and failure warning threshold panel, facilitating real-time monitoring of the test status.

[0022] 6. Scientific Rigor and Operability: This method combines the advantages of physical and numerical substructures. Large-scale substructure testing verifies the reliability of the numerical model, ensuring the scientific rigor of the test results. Furthermore, this method provides a highly accurate and reliable testing method and platform for the fire resistance design of long-span bridges, demonstrating strong operability and practical application value.

[0023] 7. Innovation and Practicality: This invention introduces advanced machine learning methods (such as LSTM and GAN) to bridge fire resistance testing technology, overcoming the limitations of traditional testing methods and providing new insights and technical support for the fire resistance design of long-span bridges. This method can effectively simulate structural responses under real-world fire scenarios, providing an important technical basis for bridge fire resistance design, assessment, and optimization, and has broad application prospects.

[0024] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of a real-time mixed fire test process according to an embodiment of the present invention; Figure 2 A diagram of the architecture of a multi-scale gated residual LSTM neural network according to an embodiment of the present invention; Figure 3 A flow chart of a real-time mixed fire test according to an embodiment of the present invention; Figure 4 A neural network architecture diagram of a rapid generation model of a fire temperature field based on GAN according to an embodiment of the present invention; Figure 5 Schematic diagram of the overall structure of a real-time mixed fire test device according to an embodiment of the present invention; Figure 6 A schematic diagram of the overall structure and experimental substructure arrangement according to an embodiment of the present invention; Figure 7 LSTM neural network architecture diagram in the LSTM-PID intelligent control system according to an embodiment of the present invention; Figure 8A neural network architecture diagram of a LSTM-based time-delay compensator according to an embodiment of the present invention; Figure 9 Schematic diagram of the structure of a composite anchor ring and a variable-section force transmission main shaft according to an embodiment of the present invention; Figure 10 is a schematic structural diagram of a free-rotating loading end according to an embodiment of the present invention; Figure 11 is a schematic structural diagram of a free-rotating anchoring end according to an embodiment of the present invention; Figure 12 Schematic diagram of the structure of a vertical annular combustion module according to an embodiment of the present invention.

[0026] Description of reference numerals: 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 beam; 42. Column; 43. Support; 44. Upper beam; 45. Base; 46. Box bottom beam; 47. Sliding beam; 5. Free-rotating loading end; 51. Composite anchor ring; 52. Tetrahedral force transmission section; 53. Cylindrical rotation section; 54. First through-hole jack; 55. Loading end primary anchor; 56. Loading end secondary anchor; 6. Free-rotating anchor end; 71. Second through-hole jack. DETAILED DESCRIPTION

[0027] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0028] The following describes a real-time hybrid test method for large-scale substructure fires of bridge cables proposed in an embodiment of the present invention with reference to the accompanying drawings.

[0029] The embodiment of the present invention discloses a real-time hybrid test method for large-scale substructure fire of bridge cables, which uses a self-balancing loading reaction frame 4 for testing. 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 distributed in a rectangular shape. Supports 43 are provided between adjacent columns 42. Upper crossbeams 44 are installed on the tops of the columns 42 by bolt groups. The tops of the two sets of rectangular lattice frames are connected by two main crossbeams 41. The bottoms of the two sets of rectangular lattice frames are both installed with an integral base 45 formed by a box-shaped steel platform. The two bases 45 are connected by a box-shaped bottom beam 46. The height of the base 45 is lower than the box-shaped bottom beam 46. A sliding crossbeam 47 is provided in the middle of the two sets of rectangular lattice frames. The sliding crossbeam 47 is driven by the electric lifting and locking system of the crossbeam and can slide and lock freely in the height direction of the rectangular lattice frame. Supports 43 are provided between adjacent columns 42. The supports 43 are provided at the bottom of the columns 42 and do not affect the lifting and lowering of the sliding crossbeam 47. The electric lifting and locking system of the beam includes a driving device arranged on the upper beam 44, which is connected to the sliding beam through a conventional screw transmission system. The sliding beam and the column 42 are provided with array-type positioning holes that are adapted to each other for installing positioning pins, and the sliding beam is positioned by the positioning pins.

[0030] The self-balancing loading reaction frame further includes a freely rotating loading end 5 and a freely rotating anchoring end 6 .

[0031] 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-section force transmission spindle is fixedly installed on both sides of the composite anchor ring 51; the variable-section force transmission spindle includes a quadrangular pyramid force transmission section 52 welded to the composite anchor ring 51 and a cylindrical rotation section 53 welded to the quadrangular pyramid force transmission section 52. The outer surface of the quadrangular pyramid force transmission section 52 is evenly distributed with radial stiffening ribs. The cylindrical rotation section 53 passes through the sliding beam 47 and rotates freely around the axis of the cylindrical rotation section 53 in the sliding beam 47 through a sleeve, thereby driving the composite anchor ring 51 to rotate freely in the sliding beam 47. In conjunction with the free sliding of the sliding beam 47, the cable substructure can be suspended and tensioned at any angle when loaded.

[0032] Both the freely rotating loading end 5 and the freely rotating anchoring end 6 utilize a variable-section force transmission spindle and a composite anchor ring 51. This design enables the cable model to freely rotate and anchor during loading, simulating the mechanical behavior under actual working conditions. The variable-section force transmission spindle design enables adaptive force transmission during loading, enhancing the flexibility and adaptability of the test apparatus.

[0033] The free-rotating loading end 5 also includes a first through-hole jack 54 and a loading-end dual-stage gradient anchoring unit. The first through-hole jack 54 is fixedly mounted via a flange within a circular through-hole within the composite anchor ring 51. The inner diameter of the circular through-hole matches the outer diameter of the first through-hole jack 54. The loading-end dual-stage gradient anchoring unit comprises a loading-end primary anchor 55 and a loading-end secondary anchor 56, which are positioned on either side of the first through-hole jack 54 to anchor the cable model. The outer diameter of the loading-end 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 within the free-rotating anchoring end 6. The outer diameter of the loading-end secondary anchor 56 is 1.2-1.5 times the outer diameter of the flange of the first through-hole jack 54. The steel strands of the main cable or stay cable model extend outward through the loading-end primary anchor 55, pass through the inner cavity of the first through-hole jack 54, and are finally anchored by the loading-end secondary anchor 56.

[0034] The first through-hole jack 54 pushes the loading-end secondary anchor 56 to achieve axial force loading of the cable model. During loading, the variable-section force transmission spindle can rotate freely about its axis. The design of the dual-stage gradient anchor unit at the loading end, through the synergistic action of the loading-end primary anchor 55 and the loading-end secondary anchor 56, achieves axial force loading of the cable model, improving the reliability and safety of the anchoring. The first through-hole jack 54 pushes the loading-end secondary anchor 56 to achieve axial force loading of the cable model, enabling precise control of the loading force and ensuring the accuracy of the test.

[0035] See also Figures 1 to 8 The real-time hybrid test method for large-scale substructure fire of bridge cables according to an embodiment of the present invention may include the following steps: S1, divide the cable-supported bridge to be analyzed into experimental substructure and numerical substructure. The experimental substructure includes the large-scale substructure of the cable system. Figure 6 shown.

[0036] S2 deploys the trained fire temperature field rapid generation model and the LSTM-based numerical substructure rapid calculation model to the computing server.

[0037] Among them, the fire temperature field rapid generation model is built based on the generative adversarial network / GAN, with the fire source power, location and wind speed as input parameters, and outputs the temperature rise-time curve of the local fire non-uniform temperature field in real time; through the gas flow-temperature mapping model of the multi-module T-shaped flame envelope combustion system, the generated temperature field curve is parsed into the valve opening, gas flow and fan inverter frequency control instructions of the burner spatial coordinates (x, y, z).

[0038] Furthermore, the full-bridge numerical substructure uses a multi-scale gated residual neural network based on LSTM for fast calculation. The weight file of the LSTM numerical substructure fast calculation model completed by offline training, including parallel branch parameters, dynamic gate weights and residual connection coefficients, is imported into the real-time calculation server; the input and output interfaces are initialized, and the sensor data streams of the spatial coordinates (x, y, z), temperature (T), real-time displacement (D), cable force (F), and wind speed (V) of the substructure boundary target position are bound. The required time window is set, and the spatial coordinates are grid-mapped to ensure that the temperature field, cable force and the geometric position of the physical specimen are aligned; the sensor network is connected to the calculation server through a high-speed communication protocol, and the displacement instruction D is sent to the real-time calculation server. load The command is pre-written into the control command queue, waiting for the LSTM-PID intelligent control system to call, activate the output limit protection, and force the displacement command D load The maximum stroke of the physical actuator is not exceeded.

[0039] S3, manufacture the test specimens corresponding to the test substructure and build a vehicle-fire-wind coupling test platform.

[0040] 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. A large-scale cable system test substructure specimen was fabricated. The through-type jack in the self-balancing loading reaction frame was driven by an LSTM-PID intelligent control system, which iteratively loaded the cable system substructure according to the stress ratio equivalence principle. A T-shaped intelligent dynamic control combustion system was arranged along the main cable and boom, and an intelligent valve positioner was installed on the burner system gas pipeline. Based on the target temperature curve and thermocouple feedback data, the LSTM-PID intelligent control system adjusted the gas flow in real time, achieving a realistic representation of the non-uniform fire scene in a laboratory environment. The variable-speed angle controllable fan system dynamically adjusted the wind speed and angle according to the actual measured values ​​of the ultrasonic anemometer, achieving real-time controllable fan speed and angle, simulating a real-world wind-fire coupled fire environment.

[0041] S4, deploy fire scene environment monitoring system.

[0042] The fire scene environment monitoring system includes armored thermocouples, strain gauges, fiber Bragg grating sensors, force sensors, laser displacement sensors, ultrasonic anemometers, and a high-temperature fire scene panoramic camera. Specifically, armored thermocouples, strain gauges, and fiber Bragg grating sensors are deployed along the main cable and boom; high-precision force sensors and laser displacement sensors are installed at the loading ends of the main cable and boom; ultrasonic anemometers and micro pressure sensors are installed on the windward and leeward sides of the cable system test substructure; and a 360-degree high-temperature fire scene panoramic camera records the morphological changes of the fire scene and the test specimen in real time.

[0043] In 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.

[0044] Specifically, the test substructure was preloaded through 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 lag of each system was calculated using the least squares method based on the feedback data from the force sensor, thermocouple, ultrasonic anemometer, and pressure sensor. The time lag value was 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 lag of each system before the formal loading of the hybrid test.

[0045] In step S6, the computing server receives data from the fire environment monitoring system of the monitoring system; the LSTM-based numerical substructure fast calculation model predicts the boundary displacement of the next time step in real time, and converts the predicted boundary displacement into force or displacement loading instructions for physical tests according to the displacement equivalence principle.

[0046] Specifically, in the real-time computing server, the sensor network data arranged in step S4 is received through a high-speed communication protocol, the multi-dimensional data is mapped into a spatial grid, and the input matrix [x, y, z, T, D, F, V] is constructed. t , and normalize it by pre-stored normalization parameters; predict the boundary displacement D of the target position at the next time step in real time pred (x, y, z), single inference time ≤ 1 second; the predicted displacement is converted into physical test loading instructions according to the displacement equivalence principle.

[0047] S7, applying force load and open flame non-uniform thermal load to the specimen, sending the force or displacement loading instruction of the test substructure boundary loading extracted in step S6 to the self-balancing loading reaction frame through the LSTM-PID intelligent control system, synchronously collecting feedback data such as force and displacement and transmitting them back to the computing server through TCP communication, and recording and drawing the stress-strain curve of the test substructure under thermal-mechanical coupling in real time; S8, starts the LSTM-based time delay compensator for dynamic compensation, issues control instructions in advance by predicting the time delay and gradually reduces the time delay error.

[0048] Specifically, based on the initial time lag compensation in step S5, the LSTM-based time lag compensator starts the dynamic compensation mechanism: first, the time difference Δt1 from the issuance of the command of the T1 analysis step force, burner opening, and fan speed and angle to the sensor measurement value reaching the target value is calculated; the LSTM-based time lag compensator predicts the time lag Δt2 that will appear in the next analysis step, and when the second analysis step T2 is loaded, the command is issued in advance by Δt2; the time lag Δt is calculated and recorded after each round of loading. i, dynamically update the LSTM network weights, and gradually reduce the time lag error until the time lag error of the entire test cycle is less than 1s.

[0049] S9, repeat steps S6 to S8 until the deformation or stress of a certain analysis specimen exceeds the limit, and then end the loading.

[0050] S10, dynamically displaying the stress, deformation, and temperature field distribution of the entire bridge under fire using 3D animation software. The 3D animation software may be 3D visualization software such as Unity3D.

[0051] Specifically, the Unity3D real-time engine can be used to bind LSTM prediction data or sensor measured data to the three-dimensional model grid nodes, and the stress, deformation and temperature field distribution changes of the entire bridge under fire can be dynamically visualized based on the value-color mapping algorithm; the load-displacement curve, temperature-time curve and failure warning threshold panel can be dynamically updated.

[0052] According to one embodiment of the present invention, in step S1, the test substructure includes the main cable, the boom or sling segment and the connection components involved in the segment; wherein, Figure 5 As shown, when the prototype of the test specimen is the main cable of a suspension bridge, a large-diameter cable test model disclosed in the Chinese invention patent publication number CN120064553A, a real-fire test device and test method for a large-diameter cable test model, is selected. The two ends of the main cable are successively passed through the free-rotating anchor end and the free-rotating loading end at the same height on the self-balancing loading reaction frame and then anchored using anchors. A large-scale equivalent model is selected according to the need, with the diameter of the test specimen being no less than 1:4 compared to the actual bridge main cable diameter. The boom or sling is a full-scale model. When testing a full-scale boom or sling, its upper end is secured to the test model of the main cable via a cable clamp, while its lower end is connected to a hydraulic loading device embedded in the self-balancing loading reaction frame. In this embodiment, the hydraulic loading device is a second through-hole jack 71 embedded in the cavity within the box-type bottom beam 46. The boom is anchored after passing through the second through-hole jack 71.

[0053] According to one embodiment of the present invention, in step S2, constructing a fire temperature field rapid generation model includes: S21, build a GAN network, which includes a generator and a discriminator; The generator takes the fire source power, normalized position coordinates (x, y), and wind speed as input. It extracts local fire dynamic features through a two-layer spatiotemporal convolutional LSTM branch and embeds the residual term of the heat diffusion equation to construct a physical constraint branch. After the two-way features are fused through a fully connected layer, the output is a normalized temperature field matrix, which is ultimately denormalized into the actual temperature-time curve. The two branches of the two-layer spatiotemporal convolutional LSTM have 64 and 128 channels, respectively, and use LeakyReLU activation. The output normalized temperature field matrix uses Sigmoid activation with a value range of [0, 1]. The discriminator takes the generated or real temperature field matrix and fire source parameters (power, location, wind speed) as joint inputs, extracts spatiotemporal 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 through global average pooling, outputting the probability of temperature field authenticity. At the same time, the residual connection is used to calculate the residual of the heat diffusion equation as a physical plausibility criterion. The three-layer convolutional network has 128, 64, and 32 channels, respectively, and uses ReLU activation. The output layer uses Sigmoid activation. The residual of the heat diffusion equation is the difference between the generated temperature field and the predicted value of the theoretical equation. The calculation formula is: Where, To generate the temperature field, is the ambient temperature, To generate the spatial second-order derivative of the temperature field, is the balance term between the heat source and the ambient temperature; is the L2 norm, which is used to measure the size of the residual; S22, based on fire dynamics software (such as FDS), simulates and generates three-dimensional non-uniform temperature field data for the entire bridge, covering different combinations of fire source power, location, and wind speed. Extracts spatiotemporal temperature distribution data and constructs a training dataset with fire source parameters as input and a gridded temperature field matrix as output label. Verifies data reliability through experimental substructure testing, reproduces the target temperature field of the fire dynamics software, and compares the measured and simulated temperature gradient distributions to ensure an error range of ≤5%. S23, offline training of a rapid generation model of fire temperature field based on a GAN network, and verification data based on fire dynamics software simulation and physical experiments; S24, deploy the trained fire temperature field rapid generation model to the computing server, bind the fire source parameter input interface; input the target as fire source power, location, and wind speed, and generate the target temperature-time curve. Among them, the architecture of the neural network of the fire temperature field rapid generation model based on GAN is as follows: Figure 4 shown.

[0054] According to one embodiment of the present invention, in step S2, the LSTM-based numerical substructure rapid calculation model is a multi-scale gated residual LSTM neural network prediction model. The multi-scale gated residual LSTM neural network prediction model is constructed by introducing a parallel multi-branch structure, a dynamic feature selection gate, and a residual cross-layer connection on the basis of LSTM. It is used to predict the evolution process of the entire bridge under fire based on the measured values ​​of the multi-dimensional monitoring sensors in the test substructure test and output the loading boundary force or displacement instruction in real time. The specific steps of training the LSTM-based numerical substructure rapid calculation model are as follows: S31: Build a multi-scale gated residual LSTM neural network. The multi-scale gated residual LSTM neural network includes a two-branch LSTM unit, i.e., a first-branch LSTM unit and a second-branch LSTM unit. The specific construction process includes: S311, the input layer receives historical time series data, and the input features include position spatial coordinates, temperature field, boundary displacement, cable force and wind speed; S312, parallel dual-branch LSTM unit, where the first branch LSTM unit processes the original step-length time series data with a step size of 1, 128 hidden units, and an activation function of tanh, and is used to extract short-term local mutation features; the second branch LSTM unit processes the downsampled time series data with a step size of 5, 64 hidden units, and an activation function of tanh, and is used to extract long-term global evolution features; the first branch LSTM unit and the second branch ISTM unit both include a first layer LSTM and a second layer LSTM; S313, the dynamic feature selection gate calculates the branch weight through the sigmoid function, and the fusion formula is: Among them, h final is the hidden state vector after the fusion of the two branch LSTM units, h branch1 is the output hidden state vector of the first branch LSTM unit, h branch2 is the output hidden state vector of the second branch LSTM unit, g t The meaning of is the dynamic gating weight coefficient; S314, the residual block cross-layer connection alleviates the gradient disappearance, and the output and input of the first layer LSTM and the second layer LSTM are superimposed through the jump connection. The formula is: in, is the hidden state vector of the LSTM layer at time step t, is the hidden state vector of the previous time step of time step t at layer l, is the spatial coordinate vector [x,y,z] at time step t t , is the hidden state vector at the l-1 layer at time step t, is the LSTM unit of layer l, l =1,2; S315, hybrid activation function, the forget gate and input gate use the sigmoid function, the candidate memory unit uses the tanh function, and the output gate uses the Swish function f(x)=x·σ(x); S316, fully connected output layer, outputs the boundary force prediction value or displacement prediction value of the next time step; S32, constructing a dataset for a LSTM-based numerical substructure fast calculation model, specifically including the following steps: 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; S322: Build a thermal-mechanical finite element model of the entire bridge and substructures using finite element analysis software (such as ABAQUS). Interpolate the temperature field output by the fire dynamics software to the finite element mesh nodes according to the time step. Use a sequential coupling analysis method to first apply temperature loads to calculate thermal expansion effects, then superimpose mechanical loads to analyze the structural response, and output displacement, cable force, and strain data. S323: Fabricate test specimens corresponding to the test substructure and reproduce the daily temperature field of the fire dynamics software in the laboratory. Apply equivalent mechanical loads using a through-hole jack, measure displacement and cable tension data, and compare them with the simulation results of the finite element analysis software to verify the reliability of the finite element model. S324, processing the verified simulation data according to the multi-scale gated residual LSTM input requirements 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 displacement and cable force at the next time step. 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 inputs the original step data, and the second branch LSTM unit inputs the downsampled data. The dynamic feature selection gate weights are optimized by historical error backpropagation; S34, online incremental learning, intercepts sensor data in real time, fine-tunes only the model's downsampling branch parameters and constrains the weight change amplitude to ≤0.01, dynamically adapting to fire changes and improving prediction accuracy while maintaining model stability; S35 performs real-time prediction and control during the real-time mixing test. The specific process is as follows: S351, taking displacement loading as an example, convert the displacement of the next time step into loading displacement instructions according to the displacement equivalence principle. The formula is: Where D load is the displacement instruction, D pred is the displacement of the next time step; L phy is the actual geometric span length of the physical specimen, L num is the equivalent span length of the numerical model, for large-scale models L phy =L num ; S352, the self-balancing loading reaction frame is driven by the LSTM-PID intelligent control system. The control quantity calculation formula is: Where γ is the feedforward gain coefficient, which is dynamically adjusted through fuzzy logic; S353, dynamic gating weight g t Mapped to PID proportional coefficient K p , to achieve adaptive loading; S36, iteratively updates the thermal-displacement / force boundary conditions until the specimen is deformed or the stress exceeds the limit, and outputs the failure mode and critical threshold. Among them, the architecture of the multi-scale gated residual LSTM neural network is as follows Figure 2 shown.

[0055] 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, 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 a plurality of linearly arranged frame-type burners 21 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 a plurality of steel wire ropes. The vertical annular combustion module 3 includes a plurality of layers of annular burners arranged in an upper and lower layers. The main body of the annular burner is annular steel pipe 31, which is fixedly connected by connecting pipes and is also provided with a plurality of flame nozzles 32 capable of adjusting the flame spray angle. The variable speed 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 angle; the high-dynamic axial flow fan 11 is connected to an ultrasonic wind speed and direction sensor and a pressure sensor 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 fan is also equipped with a fan integrated frequency converter to achieve continuous wind speed adjustment and precise control of the tilt angle.

[0056] According to one 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 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 increment Δu gas and fan frequency f fan ; The control instructions are generated according to the following formula: Where Δu gas is the burner valve opening increment, f fan is the fan frequency; α and β are coupling coefficients, which are dynamically adjusted according to the real-time working conditions; The existing PID control equation is obtained by adding the burner valve opening increment Δu to the original PID control method. gas and fan frequency f fan Item realizes the calculation of control instructions; Finally, the control instructions are sent to the burner and fan inverter, synchronously triggering the wind-fire coupling loading to achieve dynamic reproduction of the fire scene. Among them, the architecture of the LSTM neural network in the LSTM-PID intelligent control system is as follows: Figure 7 shown.

[0057] According to one embodiment of the present invention, in step S5, the control instruction generation step is: S51, using sensors to collect temperature field data, wind speed angle data and burner valve opening history series in real time to construct an input vector of a multidimensional time series; S52: Input the input vector 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 less than 5%. S53: The predicted burner valve opening increment and fan frequency are used as feedforward correction quantities and added to the feedback output of the PID controller to generate a final control instruction.

[0058] According to one embodiment of the present invention, in step S8, the LSTM time lag compensator is a neural network based on LSTM, comprising a hidden layer of 16-32 neurons, wherein the number of neurons is preferably 32, and the input layer node comprises a historical time lag sequence Δt i , system real-time load rate, ambient temperature gradient, and ambient wind speed gradient, where i=(1, 2, ..., n), and the output layer is mapped to the time-delay prediction value Δt for the next analysis step through the fully connected layer. i+1 , specifically including the following steps: S61, build a time-lag data set and collect historical time-lag series Δti , the data of the three dimensions of system real-time load rate, ambient temperature gradient and wind speed gradient are constructed into an input vector and input into the LSTM-based neural network, where Δt i Refers to the delay time from when the control command is issued to when the sensor's measured value reaches the standard; S62, offline training and online optimization, uses pre-loaded data samples to train the LSTM-based time-delay compensation model, with the loss function being the mean absolute error. During formal loading, after each analysis step, the data stream is intercepted based on a sliding window with a step length of 5-10, and the network parameters are updated online using the Adam optimizer, with a step length of 10 being preferred. S63, dynamic time delay compensation, before the second analysis step loading, send the control command in advance △t2 time, and record the actual time delay △t real ,△t2 refers to the time difference between issuing the instruction in the first analysis step and loading into place; S64, Error Correction and Convergence, Based on Δt i-real and △t i The gradient descent method is used to correct the weights of the LSTM-based time-delay compensation model until the time-delay error of multiple consecutive analysis steps is ≤1 second; Δt i-real Refers to the time delay actually measured in each analysis step, i=(1, 2, ..., n). The architecture of the neural network based on LSTM time delay compensator is as follows: Figure 8 shown.

[0059] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A real-time hybrid test method for large-scale substructure fire of bridge cables, characterized by: include: S1, divide the cable-supported 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; S3, fabricate the test specimens corresponding to the test substructure and build the vehicle-fire-wind coupling test platform; S4, deploy fire scene environment monitoring system; S5, preload the specimen on the vehicle-fire-wind coupling test platform, obtain the initial time delay of each system, and input it into the LSTM-PID intelligent control system for feedforward compensation; In step S6, the computing server receives data from the fire environment monitoring system of the monitoring system; the LSTM-based numerical substructure fast computing model predicts the boundary displacement of the next time step in real time, and converts the predicted boundary displacement into force or displacement loading instructions for physical testing according to the displacement equivalence principle; S7, applying force load and open flame non-uniform thermal load on the specimen, sending the force or displacement loading instructions 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 collecting data and transmitting it back to the computing server; S8, start the LSTM-based time delay compensator for dynamic compensation, issue control instructions in advance by predicting the time delay and gradually reduce the time delay error; S9, repeating steps S6 to S8 until the deformation or stress of the specimen in a certain analysis step exceeds the limit, and then ending the loading; S10, using 3D animation software to dynamically display the stress, deformation and temperature field distribution of the entire bridge under fire.

2. The real-time hybrid test method for large-scale substructure fire of bridge cables according to claim 1 is characterized in that: In step S1, When the prototype of the specimen is the main cable, its two ends are successively passed through the free rotation anchorage end and the free rotation loading end at the same height on the self-balancing loading reaction frame and then anchored using anchors respectively. A large-scale equivalent model with a diameter of the specimen not less than 1:4 to the actual bridge main cable diameter is selected as needed; When a full-scale boom or sling needs to be tested, its upper end is fixed to the test model of the main cable through 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 large-scale substructure fire of bridge cables according to claim 1 is characterized in that: In step S2, constructing the fire temperature field rapid generation model includes: S21, build a GAN network, which includes a generator and a discriminator; The generator takes the fire source power, normalized position coordinates, and wind speed as input conditions, extracts local fire dynamic features through a two-layer spatiotemporal convolution LSTM branch, and embeds the residual term of the heat diffusion equation to construct a physical constraint branch. After the two-way features are fused through a fully connected layer, the normalized temperature field matrix is ​​output, which is finally denormalized into the actual temperature-time curve. The discriminator takes the generated or real temperature field matrix and the fire source parameters as joint input, 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 through global average pooling, outputs the authenticity probability of the temperature field, and calculates the residual of the heat diffusion equation through residual connection as a physical rationality criterion; The residual of the heat diffusion equation is the difference between the generated temperature field and the predicted value of the theoretical equation, and the calculation formula is: ； S22: Generate three-dimensional non-uniform temperature field data for the entire bridge based on fire dynamics software simulation, extract spatiotemporal temperature distribution data, construct a training data set, verify data reliability through test substructure tests, and compare measured and simulated temperatures to ensure an error range of ≤5%; S23, offline training of a rapid generation model of fire temperature field based on a GAN network, and verification data based on fire dynamics software simulation and physical experiments; S24, deploying the trained fire temperature field rapid generation model to the computing server.

4. The real-time hybrid test method for large-scale substructure fire of bridge cables according to claim 1 is characterized in that: In step S2, the LSTM-based numerical substructure rapid calculation model is a multi-scale gated residual LSTM neural network prediction model. The multi-scale gated residual LSTM neural network prediction model is constructed by introducing a parallel multi-branch structure, a dynamic feature selection gate, and a residual cross-layer connection on the basis of LSTM. It is used to predict the evolution process of the full bridge under fire based on the measured values ​​of the multi-dimensional monitoring sensors in the test substructure test and output the loading boundary force or displacement instruction in real time. The specific steps of training the LSTM-based numerical substructure rapid calculation model are as follows: S31, building a multi-scale gated residual LSTM neural network, wherein the multi-scale gated residual LSTM neural network includes a two-branch LSTM unit, i.e., a first-branch LSTM unit and a second-branch LSTM unit; S32, constructing a data set for a rapid calculation model of a numerical substructure 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 thermal-mechanical coupling finite element model of the full bridge and the test substructure based on finite element analysis software; making a test specimen corresponding to the test substructure, subjecting the test specimen to thermal-mechanical coupling loading with a non-uniform temperature field, and verifying the accuracy of the test specimen thermal-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 split the data set. The first branch LSTM unit inputs the original step data, and the second branch LSTM unit inputs the downsampled data. The dynamic feature selection gate weight is optimized by historical error back propagation; S34, online incremental learning, dynamically adapts to fire changes by intercepting sensor data in real time, fine-tuning only the model downsampling branch parameters and constraining the weight change amplitude to ≤0.01; S35, real-time prediction and control during real-time mixing experiments; 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.

5. The real-time hybrid test method for large-scale substructure fire of bridge cables according to claim 4 is characterized in that: Step S31 includes: S311, the input layer receives historical time series data, and the input features include position spatial coordinates, temperature field, boundary displacement, cable force and wind speed; S312, a parallel dual-branch LSTM unit, wherein the first branch LSTM unit processes the original step time series data to extract short-term local mutation features; the second branch LSTM unit processes the downsampled time series data to extract long-term global evolution features; the first branch LSTM unit and the second branch ISTM unit both include a first layer LSTM and a second layer LSTM; S313, the dynamic feature selection gate calculates the branch weight through the sigmoid function, and the fusion formula is: ； S314, the residual block cross-layer connection alleviates the gradient disappearance, and the output and input of the first layer LSTM and the second layer LSTM are superimposed through the jump connection. The formula is: ； S315, hybrid activation function; S316, the fully connected output layer, outputs the boundary force prediction value or displacement prediction value of the next time step.

6. The real-time hybrid test method for large-scale substructure fire of bridge cables according to claim 4 is characterized in that: Step S35 includes: S351, taking displacement loading as an example, convert the displacement of the next time step into loading displacement instructions according to the displacement equivalence principle. The formula is: ； S352, the self-balancing loading reaction frame is driven by the LSTM-PID intelligent control system. The control quantity calculation formula is: ； S353, dynamic gating weight is mapped to PID proportional coefficient to achieve adaptive loading.

7. The real-time hybrid test method for large-scale substructure fire of bridge cables according to claim 1 is 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 a plurality of linearly arranged frame-type burners provided with vertical flame tubes, and the frame-type burners are suspended and fixed in the middle test area of ​​the self-balancing loading reaction frame by a plurality of steel wire ropes; the vertical annular combustion module includes a plurality of layers of annular burners arranged in an upper and lower manner, and the main body of the annular burner is annular steel pipe, which is fixedly connected by a connecting pipe and is also provided with a plurality of flame nozzles capable of adjusting the flame spray angle; The variable speed angle controllable fan system includes several high-dynamic axial flow fans arranged in a matrix topology on one side of a self-balancing loading reaction frame and capable of adjusting the wind direction angle; the high-dynamic axial flow fan is connected to an ultrasonic wind speed and direction sensor and a pressure sensor to provide real-time feedback of wind field parameters to the LSTM-PID intelligent control system for dynamic error regulation; the high-dynamic axial flow fan is also equipped with a fan integrated frequency converter to achieve continuous wind speed adjustment and precise control of the tilt angle.

8. The real-time hybrid test method for large-scale substructure fire of bridge cables according to claim 1 is 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 and historical error, the hidden layer includes 54-64 neurons, and the output layer outputs the burner valve opening increment and fan frequency; the control command is generated according to the following formula: ； Finally, the control instructions are sent to the burner and fan inverter, synchronously triggering the wind-fire coupling loading to achieve dynamic reproduction of the fire scene.

9. The real-time hybrid test method for large-scale substructure fire of bridge cables according to claim 8 is characterized in that: In step S5, the steps of generating the control instruction are: S51, using sensors to collect temperature field data, wind speed angle data and burner valve opening history series in real time to construct an input vector of a multidimensional time series; S52: Input the input vector 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 less than 5%. S53: The predicted burner valve opening increment and fan frequency are used as feedforward correction quantities and added to the feedback output of the PID controller to generate a final control instruction.

10. The real-time hybrid test method for large-scale substructure fire of bridge cables according to claim 1 is characterized in that: In step S8, the main body of the time delay compensator is a neural network based on LSTM, which contains a hidden layer of 16-32 neurons, and the input layer node contains the historical time delay sequence Δt i , system real-time load rate, ambient temperature gradient, and ambient wind speed gradient, where i=(1, 2, ..., n), and the output layer is mapped to the time-delay prediction value Δt for the next analysis step through the fully connected layer. i+1 , specifically including the following steps: S81, build a time-lag data set and collect historical time-lag series Δt i , the data of the three dimensions of system real-time load rate, ambient temperature gradient and wind speed gradient are constructed into an input vector and input into the LSTM-based neural network, where Δt i Refers to the delay time from when the control command is issued to when the sensor's measured value reaches the standard; S82, offline training and online optimization, uses pre-loading data samples to train the LSTM-based time-delay compensation model, with the loss function being the mean absolute error. During formal loading, the network parameters are updated online through the optimizer after each analysis step. S83, dynamic time delay compensation, before the second analysis step loading, send the control command in advance △t2 time, and record the actual time delay △t real ,△t2 refers to the time difference between issuing the instruction in the first analysis step and loading into place; S84, Error Correction and Convergence, Based on Δt i-real and △t i The gradient descent method is used to correct the weights of the LSTM-based time-delay compensation model until the time-delay error of multiple consecutive analysis steps is ≤1 second; Δt i-real Refers to the time lag actually measured in each analysis step, i=(1, 2,...,n).

Images