Method for testing strength properties of jute fiber reinforced composites

Abstract

The present application relates to the technical field of composite material strength detection, and discloses a jute fiber reinforced composite material strength performance detection method.The method extracts multi-scale mechanical characteristics through multi-source sensing signals, and constructs a mapping network of the same and internal damage state.Innovatively, a virtual loading driver is introduced, dynamic loading is performed in a simulation environment based on feature input, and virtual damage field evolution is predicted, according to which physical loading control parameters are adjusted in real time.Fusion of actual loading feedback data and virtual evolution data is performed, the mapping network is corrected online, and the driver parameters are iteratively optimized, forming a closed-loop adaptive loading process until the data converges to a preset damage criterion threshold, and finally the strength performance index is output.The present application realizes real-time tracking of composite material damage evolution and adaptive regulation and control of loading strategy, and improves the accuracy and test efficiency of strength detection.

Description

Technical Field

[0001] This invention relates to the field of composite material strength testing technology, specifically a method for testing the strength properties of jute fiber reinforced composite materials. Background Technology

[0002] Current strength testing of jute fiber-reinforced composites typically employs uniaxial loading using physical testing machines, with macroscopic mechanical data recorded by sensors to assess performance. This process relies on a pre-set, fixed loading program, failing to identify and respond in real time to the complex nonlinear responses exhibited by the material due to fiber, interface, and matrix damage. Loading control and damage evolution are disconnected; even with online monitoring, the data is primarily used for post-hoc analysis, failing to establish a dynamic closed-loop adjustment of the loading strategy.

[0003] The main drawback of existing methods is the severe disconnect between the loading process and the actual damage state within the material. Fixed loading patterns cannot adapt to the accumulation process of multi-scale damage, potentially missing the damage critical point or inducing atypical failure, thus affecting the accuracy of the assessment. Furthermore, the model parameters describing damage typically need to be pre-calibrated through numerous independent experiments, making it impossible to update and self-correct online using real-time feedback during a single test, limiting the accuracy and efficiency of the methods.

[0004] A method is needed to achieve real-time synchronization and interaction between the loading process and the material damage evolution. This method must construct a virtual loading environment that reflects the damage state and drive it through real-time sensing signals. A mechanism based on physical feedback needs to be introduced to iteratively optimize the virtual environment and loading strategy online, enabling the actual loading to adaptively track the true path of internal material damage, thereby achieving accurate determination of strength performance. Summary of the Invention

[0005] The purpose of this invention is to provide a method for testing the strength properties of jute fiber reinforced composite materials, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for testing the strength properties of jute fiber reinforced composite materials, the method comprising:

[0007] The real-time physical field signal of the jute fiber reinforced composite material sample was obtained by a multi-source sensing unit, and the multi-scale mechanical features in the physical field signal were extracted.

[0008] Establish a mapping network between the multi-scale mechanical characteristics and the internal damage state of the specimen;

[0009] A virtual loading driver is constructed, and the virtual loading driver is simulated and calibrated according to a preset loading path to generate a set of driving parameters;

[0010] The virtual loading driver is imported into the mapping network, and the multi-scale mechanical features are used as input conditions to drive the virtual loading driver to perform dynamic loading in the simulation environment;

[0011] The evolution path of the virtual damage field inside the sample during the dynamic loading process is captured in real time.

[0012] The real-time adjustment amount of the loading control parameters is calculated based on the evolution path of the virtual damage field;

[0013] The physical loading device performs the actual loading operation corresponding to the loading control parameters, and simultaneously collects feedback mechanical data during the actual loading process.

[0014] The feedback mechanical data is fused and compared with the evolution data of the virtual damage field to correct the internal weights of the mapping network;

[0015] Based on the corrected mapping network, the driver parameter set of the virtual loading driver is iteratively optimized;

[0016] Repeat the process of dynamically loading the driving parameter set for iterative optimization until the feedback mechanical data converges to the preset damage criterion threshold, and output the final strength performance index.

[0017] Preferably, the step of acquiring the real-time physical field signal of the jute fiber reinforced composite material sample through a multi-source sensing unit and extracting the multi-scale mechanical features from the physical field signal specifically includes:

[0018] An array of acoustic emission sensors was deployed to capture the acoustic emission signals of the sample under small loads, while a distributed fiber optic sensor network was deployed to measure the full-field strain distribution on the sample surface.

[0019] The acoustic wave emission signal and the full-field strain distribution data are acquired synchronously to form a time-aligned physical field signal sequence;

[0020] Time-frequency joint analysis was performed on the physical field signal sequence to separate the characteristic frequency band energy of fiber breakage, matrix cracking and interface debonding from the acoustic emission signal;

[0021] Extract the gradient rate of change and strain energy density distribution cloud map of the strain concentration region from the full-field strain distribution data;

[0022] By spatially correlating the energy of the characteristic frequency band with the gradient change rate of the strain concentration region, the core point of damage initiation is identified.

[0023] Based on the strain energy density distribution cloud map and the damage initiation core point, the energy release rate trend of each damage region at the mesoscale is calculated.

[0024] The spatial coordinates of the damage initiation core point, the energy release rate trend, and the time-varying trajectory of the energy in the characteristic frequency band are summarized to form the multi-scale mechanical feature data set required for this detection.

[0025] Preferably, establishing the mapping network between the multi-scale mechanical characteristics and the internal damage state of the specimen includes:

[0026] Normalize each data item in the multi-scale mechanical feature data set to the same dimension space;

[0027] Construct a hybrid neural network model containing convolutional layers and long short-term memory layers, wherein the convolutional layers are used to process strain energy density distribution cloud map data with spatial topological relationships, and the long short-term memory layers are used to process time-varying trajectory data of the energy in the characteristic frequency band.

[0028] The spatial coordinates of the core point of damage initiation and the trend of energy release rate are used as guiding signals for the attention mechanism and input into the hybrid neural network model.

[0029] The hybrid neural network model was trained using mechanical characteristics and damage morphology images of jute fiber reinforced composite material specimens with known damage states from historical tests.

[0030] After training, the hybrid neural network model can output a quantitative damage state matrix that describes the internal microcrack density, crack propagation direction, and interface damage area of ​​the sample based on the input real-time multi-scale mechanical characteristics.

[0031] Preferably, the process of constructing a virtual load driver and calibrating the virtual load driver according to a preset load path to generate a driver parameter set includes:

[0032] Create a virtual loading driver model in a numerical simulation environment that matches the geometry and dynamic characteristics of the physical loading device;

[0033] Define the initial load-displacement control logic, actuator response delay time, and load holding accuracy parameters for the virtual load driver model;

[0034] Define a preset loading path that includes a linear loading segment, a constant load holding segment, and a cyclic loading segment;

[0035] In the numerical simulation environment, the virtual load driver model is applied to a standard reference sample model and runs along the preset load path;

[0036] Record the sequence of deviations between the actual output load curve and the theoretical command load curve of the virtual load driver model throughout the entire operation process;

[0037] Based on the deviation sequence, the internal control parameters of the virtual load driver model are dynamically adjusted using a reverse compensation algorithm until the actual output load curve and the theoretical command load curve match a set threshold.

[0038] Archive all internal control parameters of the virtual load driver model at this time to form the initial driver parameter set for this test.

[0039] Preferably, the step of importing the virtual loading driver into the mapping network, using the multi-scale mechanical features as input conditions, and driving the virtual loading driver to perform dynamic loading in the simulation environment includes:

[0040] The multi-scale mechanical feature data set acquired in the current detection cycle is input into the trained hybrid neural network model to obtain the quantitative damage state matrix at the current moment.

[0041] The quantitative damage state matrix and the current state parameters of the virtual loading driver model are input into a loading decision agent program.

[0042] The loading decision agent program is based on a reinforcement learning framework. According to the current damage state and the loading target, it selects a loading action from a preset action space. The loading action includes increasing the load, decreasing the load, or maintaining the current load.

[0043] The virtual loading driver model performs a one-step loading operation on a virtual specimen model that is consistent with the current specimen state in the numerical simulation environment, based on the loading action selected by the loading decision agent program and the initial driving parameter set.

[0044] After the loading operation is completed, the mechanical state of the virtual sample model is updated, and a new set of virtual multi-scale mechanical features corresponding to this step is simulated and generated.

[0045] Preferably, the real-time capture of the evolution path of the virtual damage field inside the sample during the dynamic loading process includes:

[0046] In the numerical simulation environment, a damage evolution calculation rule based on phase field theory is embedded into the virtual sample model;

[0047] After each loading operation is performed by the virtual loading driver model, the damage evolution calculation rule is triggered to perform an iterative calculation.

[0048] Each iteration outputs the damage variable values, main crack propagation length, and number of branch cracks at each integration point within the virtual specimen model described in this step.

[0049] The damage variable values ​​output from multiple consecutive loading steps are arranged in chronological order to form the evolution sequence of the virtual damage field in the time dimension;

[0050] The damage field data at each time point in the evolution sequence are spatially reconstructed in three dimensions to generate a series of time-indexed damage field volumes. The series of time-indexed damage field volumes coherently describe the evolution path of the complete virtual damage field from damage initiation to macroscopic destruction.

[0051] Preferably, the real-time adjustment amount of the loading control parameters calculated based on the evolution path of the virtual damage field includes:

[0052] The evolution path of the virtual damage field was analyzed to identify the critical stage of accelerated damage propagation.

[0053] Extract the virtual multi-scale mechanical features one step before the critical stage, and calculate the feature matching degree between them and the physical field signals actually obtained from the sample.

[0054] If the feature matching degree is lower than the preset tolerance, it is determined that the virtual prediction deviates from the actual state;

[0055] Based on the predicted damage expansion amount in the evolution path of the virtual damage field and the assessment results of the current actual damage state, the load increment or decrement required to resynchronize the two is calculated.

[0056] Combining the load control accuracy and response speed of the virtual load driver model, the load increment or decrement is converted into specific adjustment values ​​and timings for the load setpoint in the physical loading device control system. These specific adjustment values ​​and timings are the real-time adjustment amounts of the load control parameters.

[0057] Preferably, the step of executing the actual loading operation corresponding to the loading control parameters through the physical loading device, and simultaneously collecting feedback mechanical data during the actual loading process, includes:

[0058] The real-time adjustment amount of the loading control parameters, i.e. the specific adjustment value and adjustment timing, is sent to the servo controller that controls the physical loading device.

[0059] The servo controller drives the actuator of the physical loading device to apply mechanical load to the jute fiber reinforced composite material sample according to the adjusted load set point.

[0060] While applying the load, the multi-source sensing unit is triggered to start a new round of data acquisition to obtain the acoustic emission signal and full-field strain distribution data under this actual loading operation.

[0061] The acoustic wave emission signal and the full-field strain distribution data collected this time, together with the load value, displacement value and timestamp actually output by the physical loading device, are packaged together into the feedback mechanical data package for the current step.

[0062] Preferably, the step of fusing and comparing the feedback mechanical data with the evolution data of the virtual damage field to correct the internal weights of the mapping network includes:

[0063] New acoustic wave emission signals and full-field strain distribution data are parsed from the feedback mechanics data package, and new multi-scale mechanical features are extracted.

[0064] From the evolution path of the virtual damage field, extract the predicted virtual multi-scale mechanical features corresponding to the current actual loading step.

[0065] Calculate the difference vectors of the new multiscale mechanical features and the corresponding virtual multiscale mechanical features in each feature dimension;

[0066] The difference vector is used as a loss signal and fed back to the hybrid neural network model;

[0067] The backpropagation algorithm is used to fine-tune the connection weights of the convolutional layer and the long short-term memory layer in the hybrid neural network model based on the loss signal, so that the network's predicted output of the current sample damage state is closer to the state reflected by the actual feedback data.

[0068] Preferably, the iterative optimization of the driver parameter set of the virtual loading driver based on the modified mapping network includes:

[0069] The response of the virtual sample model in the next loading step is re-predicted using the weighted hybrid neural network model.

[0070] Based on the re-predicted damage state, evaluate the control performance of the current set of drive parameters used by the virtual load driver model;

[0071] If the control performance index decreases, the parameter optimization routine is initiated to perform a targeted search within the parameter space of the driving parameter set to find a new combination of parameters that can improve the control performance index.

[0072] The internal control parameters of the virtual load driver model are updated using the newly found parameter combinations to form an optimized set of driver parameters.

[0073] The optimized set of driving parameters is applied to the virtual load driver model to guide the calculation of real-time adjustments to subsequent load control parameters, thereby initiating a new round of load-feedback-correction cycle.

[0074] Compared with the prior art, the beneficial effects of the present invention are:

[0075] By constructing a virtual loading actuator that can be independently calibrated and imported, and integrating it with a mapping network based on multi-source sensor signals, synchronous dynamic pre-simulation of physical experiments in digital space was achieved. Driven by real-time extracted multi-scale mechanical features, this actuator can perform dynamic loading calculations in parallel with physical loading within the simulation environment, pre-simulating the possible evolution of the damage field. This allows loading control to move away from pre-set fixed procedures and instead be guided by a digital twin reflecting the real-time state of the material, transforming the traditional "loading-observation" mode into a "predictive-guided loading" mode, enhancing the foresight and controllability of the testing process.

[0076] By establishing a bidirectional data flow and iterative optimization mechanism between the mapping network and the virtual loading driver, a complete adaptive closed loop is formed. During dynamic loading, the feedback mechanical data collected by the physical device and the evolution data of the virtual damage field are fused and compared in real time. The differences are used to correct the internal weight parameters of the mapping network online. The corrected network is then used to iteratively optimize the driving parameter set of the virtual loading driver, thereby adjusting its simulation behavior. This allows the virtual driver to continuously approximate the actual mechanical response of the material, while simultaneously refining the physical loading strategy. The loading control parameters can be adaptively adjusted according to the real-time state of material damage evolution, realizing dynamic optimization of the testing process following the material's own characteristics, improving the accuracy of damage critical state determination and testing efficiency. Attached Figure Description

[0077] Figure 1 This is a schematic diagram illustrating the working principle of the method for testing the strength properties of jute fiber reinforced composite materials according to the present invention.

[0078] Figure 2 A flowchart for extracting multi-scale mechanical features;

[0079] Figure 3 A flowchart for establishing a mapping network;

[0080] Figure 4 The diagram shows the peak position of the strain gradient and the evolution of the strain energy density during the loading process of jute fiber reinforced composite material;

[0081] Figure 5 This is a comparison chart of the measured and predicted strain gradient change rate during loading of jute fiber reinforced composite materials. Detailed Implementation

[0082] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0083] Please see Figure 1This invention provides a method for testing the strength properties of jute fiber reinforced composite materials. The method includes: a multi-source sensing unit arranged on the surface of the sample begins operation to acquire real-time physical field signals of the sample during loading; multiple scale mechanical features containing damage information are further extracted from these raw signals; a mapping network is established to understand the complex relationship between these mechanical features and the microscopic damage inside the material; simultaneously, a virtual loading driver corresponding to the physical loading device is constructed in a numerical simulation environment and calibrated according to a preset loading path to generate an initial set of driving parameters; the calibrated virtual loading driver is imported into the virtual environment constituted by the aforementioned mapping network, using the multi-scale mechanical features acquired from the sample in real time as input conditions to drive the virtual loading driver to perform a dynamic loading process on the virtual sample in the simulation environment; during this simulated loading process, the complete evolution path of the virtual damage field inside the sample needs to be captured and recorded in real time; based on the analysis results of this virtual evolution path, the real-time adjustment amount of the loading control parameters of the physical loading device is calculated; the physical loading device receives and executes these adjusted control commands, applies the corresponding load to the actual sample, and simultaneously collects and packages the actual generated feedback mechanical data. The actual collected feedback mechanical data is fused and compared with the simulated virtual damage field evolution data. The differences generated by the comparison are used to correct the internal connection weights of the mapping network. Based on the improved performance of the mapping network, the driving parameter set used by the virtual loading driver is iteratively optimized. The complete loop process from dynamic loading of the virtual environment to optimization of the driving parameter set is repeated until the feedback mechanical data collected from the actual sample converges to the preset damage criterion threshold. At this point, the final strength performance index of the jute fiber reinforced composite material sample is output.

[0084] In one embodiment of the present invention, see [reference] Figure 2When acquiring real-time physical field signals of jute fiber-reinforced composite material samples through multi-source sensing units and extracting multi-scale mechanical features from the physical field signals, the implementation method is as follows: An array of acoustic emission sensors is arranged on the sample surface and key areas to capture the acoustic wave emission signals of the sample under small loads. Simultaneously, a distributed fiber optic sensor network is arranged throughout the entire test area to measure the full-field strain distribution on the sample surface. Acoustic wave emission signals and full-field strain distribution data are collected synchronously to form a time-aligned physical field signal sequence. Time-frequency joint analysis is performed on the physical field signal sequence to separate the corresponding fiber fracture and matrix characteristics from the acoustic wave emission signals. Characteristic frequency band energies of different damage modes such as cracking and interface debonding are extracted from the full-field strain distribution data. The gradient change rate of strain concentration areas and the strain energy density distribution cloud map are extracted. The characteristic frequency band energies and the gradient change rate of strain concentration areas are spatially correlated to identify the core point where damage is most likely to occur. Based on the information of the strain energy density distribution cloud map and the core point of damage initiation, the energy release rate trend of each potential damage area at the mesoscale is calculated. The spatial coordinates of the core point of damage initiation, the energy release rate trend, and the time-varying trajectory of the characteristic frequency band energies are summarized to form the multi-scale mechanical feature data set required for this detection cycle.

[0085] In practice, multi-source sensing units deployed on the surface of jute fiber-reinforced composite material samples begin to work collaboratively. Acoustic emission sensor arrays are installed on the sample surface in a specific array arrangement to capture transient elastic waves excited by internal damage events when the material is under load. A distributed fiber optic sensor network is attached to or embedded near the sample surface in a pre-designed grid pattern for continuous measurement of the full-field strain distribution on the sample surface. The synchronous acquisition of acoustic emission signals and full-field strain distribution data constitutes a time-aligned physical field signal sequence. This synchronization is achieved through a unified data acquisition system clock, ensuring that acoustic emission events and strain field changes correspond precisely on the time axis. Time-frequency joint analysis is performed on the physical field signal sequence. Wavelet transform or short-time Fourier transform methods are used to process the acoustic emission signals, separating characteristic frequency band energies corresponding to fiber fracture, matrix cracking, and interface debonding damage mechanisms from the complex time-domain waveform. For example, fiber fracture events may mainly concentrate in the high-frequency band, while matrix cracking energy may be distributed in the mid-to-low frequency band. From the full-field strain distribution data acquired by the distributed fiber optic sensor network, the gradient rate of change in the strain concentration region is extracted through spatial difference operations. Simultaneously, the strain energy density at each measurement point is calculated, generating a full-field strain energy density distribution cloud map. The energy of each characteristic frequency band separated from the acoustic emission signal is spatially correlated with the gradient rate of change in the strain concentration region. This correlation is achieved by matching the coordinates of the damage event occurrence determined by the acoustic emission sensor array with the coordinates of the strain gradient peak region, thereby identifying the most active core point among multiple potential damage initiation core points. The spatial three-dimensional coordinates of the damage initiation core points, the energy release rate trend values ​​corresponding to each core point, and the trajectory of energy changes in each characteristic frequency band over time are summarized to constitute the multi-scale mechanical feature data set required for this detection cycle. The multi-scale mechanical feature data set is organized in the form of a structured data matrix to facilitate the input and processing of the subsequent mapping network model. In some embodiments, the time-frequency joint analysis further includes pattern recognition of the acoustic emission signal to further distinguish the signal characteristics of different types of damage events. Optionally, the distributed optical fiber sensor network uses a fiber grating sensor array, whose measured full-field strain distribution data has high spatial resolution. It can be understood that the multi-scale mechanical feature data set fully integrates cross-scale information from microscopic damage emission signals to macroscopic full-field strain responses.

[0086] In one embodiment of the present invention, see [reference] Figure 3When establishing a mapping network between multi-scale mechanical features and the internal damage state of the specimen, the data in the multi-scale mechanical feature dataset are normalized to the same dimensional space. A hybrid neural network model containing convolutional layers and long short-term memory layers is constructed. The convolutional layers are used to process strain energy density distribution cloud map data with spatial topological relationships, and the long short-term memory layers are used to process time-varying trajectory data of energy in characteristic frequency bands. The spatial coordinates of the core point of damage initiation and the trend of energy release rate are used as guiding signals for the attention mechanism and input into the hybrid neural network model. The hybrid neural network model is trained using mechanical features and damage morphology images of jute fiber reinforced composite specimens with known damage states from historical experiments. After training, the hybrid neural network model can output a quantitative damage state matrix describing the internal microcrack density, crack propagation direction, and interface damage area of ​​the specimen based on the input real-time multi-scale mechanical features.

[0087] A virtual load driver is constructed and calibrated using a preset loading path. When generating the drive parameter set, a virtual load driver model with the same geometry and dynamic characteristics as the physical loading device is created in the numerical simulation environment. Initial load-displacement control logic, actuator response delay time, and load holding accuracy parameters are defined for the virtual load driver model. A preset loading path including a linear loading segment, a constant load holding segment, and a cyclic loading segment is set. In the numerical simulation environment, the virtual load driver model is applied to a standard reference specimen model and runs along the preset loading path. The deviation sequence between the actual output load curve and the theoretical command load curve of the virtual load driver model during the entire operation is recorded. Based on the deviation sequence, the internal control parameters of the virtual load driver model are dynamically adjusted using a reverse compensation algorithm until the consistency between the actual output load curve and the theoretical command load curve exceeds a set threshold. All internal control parameters of the virtual load driver model at this point are archived to form the initial drive parameter set used for this test.

[0088] In practical implementation, the process of establishing a mapping network between multi-scale mechanical characteristics and the internal damage state of the specimen begins in the data preprocessing stage. The multi-scale mechanical characteristic data set obtained from the above embodiments contains various types and dimensions of data, such as the spatial coordinates of the core point of damage initiation, the trend of energy release rate, and the time-varying trajectory of energy in the characteristic frequency band. It is necessary to normalize all data items in the multi-scale mechanical characteristic data set to a unified, dimensionless or dimensionless numerical space through linear or nonlinear transformations. The normalization operation aims to eliminate the bias caused by the differences in dimensions and orders of magnitude of different physical quantities on the subsequent training of the network model. A specially designed hybrid neural network model is constructed. The structure of the hybrid neural network model includes convolutional layers for processing spatial topological data and long short-term memory layers for processing time-series data. The convolutional layers are configured to receive and process strain energy density distribution cloud map data with clear two-dimensional or three-dimensional spatial relationships, while the long short-term memory layers are configured to receive and process time-varying trajectory data of energy in the characteristic frequency band. The time-varying trajectory data of energy in the characteristic frequency band reflects the dynamic process of energy evolution over time in different damage modes. The spatial coordinates of the damage initiation core point are fused and encoded with the energy release rate trend information to generate an attention weight guidance signal. The attention weight guidance signal is input into the corresponding layer of the hybrid neural network model to guide the network to focus on key spatial regions and time segments related to the identified core point and high energy release rate trend when processing strain energy density distribution cloud map and time-varying energy trajectories of characteristic frequency bands.

[0089] The hybrid neural network model was trained under supervision using mechanical characteristics and damage morphology images of jute fiber-reinforced composite specimens with known damage states from historical tests. The input to each sample in the training dataset was a set of multi-scale mechanical characteristic data collected and extracted from historical specimens. Each sample's label was a quantitative damage morphology image of the corresponding historical specimen obtained after the test via micro-CT or metallographic analysis. The damage morphology image was parsed into a damage state matrix containing information on microcrack density, crack propagation direction, and interface damage area. During training, the hybrid neural network model calculated and predicted the output through forward propagation, and adjusted its internal parameters by minimizing the difference between the predicted damage state matrix and the actual damage state matrix. The adjustment of the weight parameters followed the following relationship for iterative updates:

[0090] ;

[0091] in: This represents the updated neural network connection weights. This represents the neural network connection weights before the update. This represents the preset learning rate parameter. Represents the loss function Regarding weight gradient, loss function The degree of difference between the predicted damage state matrix and the actual damage state matrix was measured. After training, the hybrid neural network model has the ability to infer and output a damage state matrix that quantitatively describes the microcrack density distribution, main crack propagation direction vector, and interface damage area ratio inside the current sample based on the newly input, real-time acquired multi-scale mechanical feature data set.

[0092] In practical implementation, the process of constructing a virtual load driver and calibrating it according to a preset loading path is initiated in a numerical simulation environment. A three-dimensional virtual load driver model with the same geometry and dynamic characteristics as the physical loading device is created in the numerical simulation environment. The geometric dimensions, joint degrees of freedom, and actuator stroke of the virtual load driver model are strictly consistent with the physical loading device. Initial load-displacement control logic, actuator response delay time parameters, and load holding accuracy parameters are defined for the virtual load driver model. The load-displacement control logic determines the servo control law of the virtual load driver model when receiving displacement or load commands. The actuator response delay time parameter simulates the time lag from receiving a command to generating an actual action in the physical actuator. The load holding accuracy parameter defines the ability of the virtual load driver model to maintain load stability during the constant load phase. A preset loading path is defined, including a linear loading segment, a constant load holding segment, and a cyclic loading segment. The preset loading path is given in the form of a time-load or time-displacement function and serves as the command input for the virtual load driver model.

[0093] In the numerical simulation environment, a virtual load driver model is applied to a standard reference specimen model, driving the virtual load driver model to run strictly according to a preset loading path. The standard reference specimen model is established using a fully validated constitutive model and damage criterion for jute fiber reinforced composite materials. The deviation sequence between the actual load-time curve output by the virtual load driver model and the theoretical load-time curve (as input command) is recorded throughout the entire operation. This deviation sequence contains system error information caused by inaccuracies in the internal control parameters of the virtual load driver model. Based on the deviation sequence, a reverse compensation algorithm is used to dynamically adjust the internal control parameters of the virtual load driver model. The reverse compensation algorithm analyzes the pattern of the deviation sequence, calculates and corrects parameters such as gain and feedforward in the control logic in reverse, until, after multiple iterations, the actual output load curve and the theoretical command load curve achieve a 95% match with a set threshold across all loading segments. All internal control parameters of the virtual load driver model at this point, including but not limited to proportional gain, integral gain, derivative gain, and feedforward compensation coefficients, are archived and encapsulated to form the initial driving parameter set used for this strength performance testing of the jute fiber reinforced composite material. In some embodiments, the numerical simulation environment employs a coupled platform of multibody dynamics simulation software and finite element analysis software. Optionally, the material parameters of the standard reference specimen model are derived from standardized mechanical property test data of jute fiber reinforced composite materials. It can be understood that the simulation calibration process enables the virtual loading driver model to reproduce the dynamic characteristics of the physical loading device in the numerical environment, and the initial driving parameter set carries this calibration result.

[0094] In one embodiment of the present invention, a virtual loading driver is imported into a mapping network. Using multi-scale mechanical features as input conditions, when the virtual loading driver performs dynamic loading in a simulation environment, the set of multi-scale mechanical feature data acquired in the current detection cycle is input into the trained hybrid neural network model to obtain the quantitative damage state matrix at the current moment. The quantitative damage state matrix and the current state parameters of the virtual loading driver model are jointly input into a loading decision agent program. The loading decision agent program is based on a reinforcement learning framework and selects a loading action from a preset action space according to the current damage state and loading target. The loading action includes increasing the load, decreasing the load, or maintaining the current load. The virtual loading driver model performs a one-step loading operation on a virtual specimen model consistent with the current specimen state in the numerical simulation environment according to the loading action selected by the loading decision agent program and the initial driving parameter set. After the one-step loading operation is completed, the mechanical state of the virtual specimen model is updated, and a new set of virtual multi-scale mechanical features corresponding to this step is simulated and generated.

[0095] When capturing the evolution path of the virtual damage field inside the specimen during dynamic loading in real time, a damage evolution calculation rule based on phase field theory is embedded into the virtual specimen model in the numerical simulation environment. After each loading operation is executed by the virtual loading driver model, the damage evolution calculation rule is triggered to perform an iterative calculation. Each iteration outputs the damage variable values, main crack propagation length, and number of branch cracks at each integration point inside the virtual specimen model under this step. The damage variable values ​​output by multiple consecutive loading steps are arranged in chronological order to form an evolution sequence of the virtual damage field in the time dimension. The damage field data at each time point in the evolution sequence are spatially reconstructed in three dimensions to generate a series of time-indexed damage field volume data. This series of time-indexed damage field volumes coherently describes the evolution path of the complete virtual damage field from damage initiation to macroscopic failure.

[0096] In practice, the process of importing a virtual loading actuator into a mapping network and driving dynamic loading using multi-scale mechanical features as input conditions begins with real-time sensor data acquired from the physical sample. The set of multi-scale mechanical feature data acquired in the current detection cycle is completely input into a pre-trained hybrid neural network model. The hybrid neural network model performs forward propagation calculations on the multi-scale mechanical feature data set and outputs a quantitative damage state matrix. The quantitative damage state matrix and the current state parameters of the virtual loading actuator model are jointly transmitted to a loading decision agent program. The loading decision agent program is built based on a reinforcement learning framework and contains a value function that evaluates the relationship between the current state and the loading target. Based on the current damage state described by the quantitative damage state matrix and the pre-set loading target, the loading decision agent program selects a loading action from a preset action space. The preset action space includes three discretized basic instructions: increasing the load, decreasing the load, or maintaining the current load.

[0097] In the numerical simulation environment, a one-step loading operation is performed on a virtual specimen model that maintains the same state as the actual jute fiber reinforced composite specimen. The one-step loading operation in the numerical simulation environment is represented by a short time integration step. Within this step, the virtual loading driver model adjusts the applied load according to the loading action and applies corresponding mechanical boundary conditions to the virtual specimen model.

[0098] In practical implementation, the process of capturing the evolution path of the virtual damage field inside the specimen during dynamic loading is synchronized with the virtual loading operation. In the numerical simulation environment, a damage evolution calculation rule based on phase field theory is embedded into the virtual specimen model. Phase field theory characterizes the degree of material damage by introducing a continuous order parameter, the damage variable. After each loading operation is executed by the virtual loading driver model, the damage evolution calculation rule is triggered to perform an iterative calculation. Each iteration solves the coupled system of the phase field control equation and the mechanical equilibrium equation. Each iteration outputs the damage variable values ​​of all finite element integration points or mesh nodes inside the virtual specimen model at this step, the main crack propagation length identified based on the damage variable distribution, and the number of branch cracks generated. The main crack propagation length is calculated by extracting the connected regions where the damage variables exceed the critical threshold.

[0099] The damage variable values ​​output from multiple consecutive loading steps are arranged in time step order to form an evolution sequence of the virtual damage field in the time dimension. This evolution sequence is a four-dimensional data array with three spatial dimensions and one temporal dimension. The full-field damage variable data corresponding to each specific time point in the evolution sequence are spatially reconstructed in three dimensions to generate a series of time-indexed damage field volume data. Each frame of damage field volume data is a three-dimensional spatial snapshot of the internal damage state of the material at a specific moment. This series of time-indexed damage field volume data coherently describes the evolution path of the virtual specimen model from an initial undamaged state to damage initiation, expansion, and ultimately macroscopic failure. In some embodiments, the governing equations of the phase-field damage evolution calculation law include damage driving force and energy dissipation terms, and damage variables... The evolution is governed by the following relationship:

[0100]

[0101] in: The functional represents the total damage driving energy of the system. Represents the spatial gradient of the damage variable. This represents the anisotropy tensor related to the fracture properties of a material. This represents a function describing the change in material fracture energy density with damage. It can be understood that by solving such governing equations, a continuous damage field evolution process conforming to physical laws can be obtained. Optionally, the 3D reconstruction process uses volume rendering or isosurface extraction techniques to transform discrete damage variable data into a visualized 3D model.

[0102] In one embodiment of the present invention, when calculating the real-time adjustment amount of the loading control parameters based on the evolution path of the virtual damage field, the evolution path of the virtual damage field is analyzed, the critical stage of accelerated damage propagation is identified, the virtual multi-scale mechanical features one step before the critical stage are extracted, and the feature matching degree is calculated with the physical field signal actually obtained from the sample. If the feature matching degree is lower than the preset tolerance, it is determined that the virtual prediction deviates from the actual state. According to the predicted damage propagation amount of the next step in the evolution path of the virtual damage field and the evaluation result of the current actual damage state, the load increment or decrement required to resynchronize the two is calculated. Combining the load control accuracy and response speed of the virtual loading driver model, the load increment or decrement is converted into a specific adjustment value and adjustment timing for the load setpoint in the physical loading device control system. The specific adjustment value and adjustment timing are the real-time adjustment amount of the loading control parameters.

[0103] When the physical loading device performs the actual loading operation corresponding to the loading control parameters and simultaneously collects feedback mechanical data during the actual loading process, the real-time adjustment amount of the loading control parameters, i.e., the specific adjustment value and adjustment timing, is sent to the servo controller controlling the physical loading device. The servo controller drives the actuator of the physical loading device to apply mechanical load to the jute fiber reinforced composite material sample according to the adjusted load set point. At the same time as applying the load, the multi-source sensing unit is triggered to start a new round of data acquisition, acquiring the acoustic emission signal and full-field strain distribution data under this actual loading operation. The acquired acoustic emission signal and full-field strain distribution data, together with the load value, displacement value and timestamp actually output by the physical loading device, are packaged together into the feedback mechanical data package of the current step.

[0104] In practical implementation, the process of calculating the real-time adjustment of loading control parameters based on the evolution path of the virtual damage field begins with an in-depth analysis of the evolution path of the virtual damage field generated in the numerical simulation environment. This analysis identifies the critical stage of accelerated damage propagation, defined as the turning point where the rate of increase of the damage variable with the loading step exceeds a preset threshold. Virtual multi-scale mechanical features preceding the critical stage are extracted. These features include the predicted acoustic emission characteristic frequency band energy distribution, the gradient change rate of the strain concentration region, and the strain energy density cloud map. The extracted virtual multi-scale mechanical features are then compared with the actual physical field signals obtained from the jute fiber-reinforced composite material sample through a multi-source sensing unit. The matching degree calculation is performed for each feature dimension separately, such as comparing the amplitude of the predicted and measured acoustic emission energy in the fiber fracture characteristic frequency band, and comparing the coordinates of the predicted and measured strain gradient peak positions. Feature matching degree is quantified by a predefined similarity metric function. If the calculated overall feature matching degree is lower than the preset tolerance of 85%, it is determined that the virtual predicted state deviates from the actual physical state and loading control intervention is required.

[0105] In practice, the process of executing the actual loading operation corresponding to the loading control parameters through the physical loading device and simultaneously acquiring feedback mechanical data begins with the issuance of real-time adjustments to the loading control parameters. The specific adjustment values ​​and timing information contained in these real-time adjustments are sent to the servo controller controlling the physical loading device via a real-time communication interface. The servo controller drives the actuators of the physical loading device to apply mechanical loads to the jute fiber reinforced composite material sample according to the adjusted load setpoint. The load application process follows a preset loading rate limit to ensure smooth loading. Simultaneously with the load application, a new round of data acquisition is triggered by the multi-source sensing unit. The acoustic emission sensor array and distributed fiber optic sensor network in the multi-source sensing unit work synchronously to acquire the acoustic emission signals and full-field strain distribution data generated by the sample under this actual loading operation. The acquisition process is strictly synchronized with the load application process, triggered by a unified clock. The original waveforms of the acquired acoustic emission signals and the original full-field strain distribution data, along with the actual load values, piston displacement values, and precise timestamp information output by the physical loading device, are packaged together into a feedback mechanical data package for the current loading step. This feedback mechanical data package is stored in a standardized data structure for subsequent comparison and analysis with virtual data. Within a single decision-making cycle, refer to Table 1 for a comparison of the matching degree between virtual predicted features and actual feedback features.

[0106] Table 1: Feature Matching Degree Comparison Table

[0107]

[0108] See Figure 4 Throughout the loading process of testing the strength performance of jute fiber reinforced composite materials, the dynamic evolution relationship between the peak positions of the principal strain gradient (X and Y directions) and the maximum strain energy density was simultaneously presented, with the critical stage starting point of accelerated damage propagation marked by a red dashed line. Specifically, the peak position of the principal strain gradient X (yellow curve) generally showed a trend of first decreasing and then increasing, while the peak position of the principal strain gradient Y (purple curve) showed an initial increase followed by a fluctuating decrease in the middle stage; the maximum strain energy density (green curve) showed a fluctuating increase in the first half of the loading step, and then a phased decrease after the critical stage as the damage evolved. The core value of the figure lies in its intuitive reflection of the spatiotemporal correlation of multi-scale mechanical characteristics: the spatial migration of the peak position of the strain gradient and the change of the strain energy density correspond to the positional shift and energy release process of the core point of damage initiation. The abrupt changes in characteristics before and after the critical stage can serve as key reference nodes for calculating the matching degree between the virtual damage field and the actual physical field (as shown in Table 1, the high matching degree between the virtual predicted values ​​and the actual feedback values ​​of the peak positions of the principal strain gradient X and Y can be verified by the curve fluctuations of the corresponding step size in the figure). In terms of parameters, the unit for the peak position of the strain gradient is mm, the unit for the maximum strain energy density is MJ / m³, and the loading step size covers the entire cycle from the initial loading to the critical damage stage.

[0109] In one embodiment of the present invention, when the feedback mechanical data and the evolution data of the virtual damage field are fused and compared, and the internal weights of the mapping network are corrected, new acoustic emission signals and full-field strain distribution data are parsed from the feedback mechanical data packet. The virtual multi-scale mechanical features predicted at the step corresponding to the current actual loading step are extracted from the evolution path of the virtual damage field. The difference vectors between the new multi-scale mechanical features and the corresponding virtual multi-scale mechanical features in each feature dimension are calculated. The difference vectors are used as loss signals and fed back to the hybrid neural network model. The error backpropagation algorithm is used to fine-tune the connection weights of the convolutional layers and long short-term memory layers in the hybrid neural network model according to the loss signals, so that the network's predicted output of the current sample damage state is closer to the state reflected by the actual feedback data.

[0110] Based on the corrected mapping network, when iteratively optimizing the driving parameter set of the virtual loading driver, a weighted hybrid neural network model is used to re-predict the response of the virtual sample model in the next loading step. Based on the re-predicted damage state, the control effectiveness of the driving parameter set currently used by the virtual loading driver model is evaluated. If the control effectiveness index decreases, a parameter optimization routine is initiated to perform a directional search in the parameter space of the driving parameter set to find a new parameter combination that can improve the control effectiveness index. The found new parameter combination is used to update the internal control parameters of the virtual loading driver model, forming an optimized driving parameter set. The optimized driving parameter set is applied to the virtual loading driver model to guide the calculation of the real-time adjustment of subsequent loading control parameters, thereby starting a new round of loading-feedback-correction loop.

[0111] In practice, the process of fusing and comparing feedback mechanical data with the evolution data of the virtual damage field and correcting the weights within the mapping network begins with the parsing of the feedback mechanical data packet. From this packet, the acoustic emission signal and the original full-field strain distribution data acquired under the new actual loading step are extracted. The new multi-scale mechanical features include characteristic frequency band energy separated from the latest acoustic emission signal, strain gradient change rate and strain energy density cloud map calculated from the latest full-field strain distribution data, and the coordinates of the latest damage initiation core point and the energy release rate trend identified accordingly. From the evolution path data of the virtual damage field, the virtual multi-scale mechanical features predicted by the virtual environment corresponding to the current actual loading step are extracted. These virtual multi-scale mechanical features are completely consistent with the new multi-scale mechanical features in terms of data structure and feature dimensions.

[0112] The difference vectors between the new multi-scale mechanical features and the corresponding virtual multi-scale mechanical features are calculated across each feature dimension. Each component of the difference vector corresponds to the deviation between the measured and predicted values ​​in a specific feature dimension, such as the energy deviation in the fiber fracture frequency band or the coordinate deviation of the peak position of the principal strain gradient. The calculated difference vectors are used as the loss signal and fed back into the hybrid neural network model. The loss signal quantifies the degree of inconsistency between the hybrid neural network model's previous predictions and the actual physical response. An error backpropagation algorithm is employed to fine-tune the connection weights of the convolutional layers and long short-term memory layers in the hybrid neural network model based on the loss signal. The error backpropagation algorithm calculates the gradient of the loss signal with respect to each adjustable weight parameter in the network using the chain rule. The update of the weight parameters aims to minimize the loss signal, and its incremental adjustment follows the following relationship:

[0113]

[0114] in: Represents weight parameters The adjustment amount, This parameter represents the learning rate used during the network fine-tuning phase, and its value is typically smaller than the learning rate used during the initial training phase. Represents the loss function Regarding weight parameters gradient, loss function It is a norm of the aforementioned difference vector, used to comprehensively measure the overall prediction error. By iteratively executing this feedback-comparison-backpropagation process, the predicted output of the hybrid neural network model on the current damage state of the jute fiber reinforced composite sample is made closer to the state reflected by the actual feedback data, thereby achieving online adaptive correction of the mapping network.

[0115] In practice, the process of iteratively optimizing the driving parameter set of the virtual loading driver based on the modified mapping network is immediately followed by a weight correction step. Using a weight-corrected hybrid neural network model, the mechanical response and damage evolution of the virtual sample model in the next loading step are re-predicted. Based on the re-predicted damage state, the control effectiveness of the currently used driving parameter set of the virtual loading driver model is evaluated. Control effectiveness can be quantified by indicators such as the tracking error between the virtual loading path and the ideal loading path, or the time required to reach the target damage state. If the evaluation results show a decrease in control effectiveness, such as an increase in tracking error, a parameter optimization routine is initiated. A directional search is performed within the parameter space of the driving parameter set. This directional search can employ gradient-based optimization methods or heuristic algorithms to find new parameter combinations that improve control effectiveness. The found new parameter combinations are used to update the internal control parameters of the virtual loading driver model, forming an optimized driving parameter set. The driving parameter set includes, but is not limited to, the proportional, integral, and derivative gains of the load control loop, as well as the feedforward compensation coefficients. The optimized set of driving parameters is applied to the virtual load driver model to guide the calculation of the real-time adjustment of subsequent load control parameters. The optimized set of driving parameters enables the virtual load driver to control the behavior more accurately in the simulation environment, and the real-time adjustment of the load control parameters calculated based on its evolution path is also more suitable for the control of the physical loading device.

[0116] See Figure 5In the multi-scale mechanical characteristic analysis stage of strength performance testing of jute fiber reinforced composite materials, the evolution of the measured strain gradient change rate (blue solid line), the virtual predicted strain gradient change rate (red dashed line), and the difference between the two (gray filled area) under the loading steps are presented. Specifically, the horizontal axis of the figure represents the loading steps, and the vertical axis represents the strain gradient change rate (unit: 1 / mm). The measured curve reflects the actual mechanical response of the strain concentration area of ​​the sample during physical loading, while the virtual predicted curve is the simulated response generated by the mapping network combined with the virtual loading driver. The gray area intuitively quantifies the degree of deviation between the measured and predicted values ​​at different loading stages (e.g., near loading step 10, the deviation between the two increases significantly, corresponding to the mechanical characteristic fluctuations in the sample damage initiation stage). The core value of the figure lies in intuitively presenting the "measured-virtual" matching state of the strain gradient change rate in multi-scale mechanical characteristics. It is a key visualization basis for subsequent fusion comparison and correction of the mapping network weights. By analyzing the deviation distribution at different loading stages, the processing weights of strain gradient characteristics in the hybrid neural network model can be optimized in a targeted manner, improving the accuracy of virtual damage field evolution prediction.

Claims

1. A method for testing the strength properties of jute fiber reinforced composite materials, characterized in that, The method includes: The real-time physical field signal of the jute fiber reinforced composite material sample was obtained by a multi-source sensing unit, and the multi-scale mechanical features in the physical field signal were extracted. Establish a mapping network between the multi-scale mechanical characteristics and the internal damage state of the specimen; A virtual loading driver is constructed, and the virtual loading driver is simulated and calibrated according to a preset loading path to generate a set of driving parameters; The virtual loading driver is imported into the mapping network, and the multi-scale mechanical features are used as input conditions to drive the virtual loading driver to perform dynamic loading in the simulation environment; The evolution path of the virtual damage field inside the sample during the dynamic loading process is captured in real time. The real-time adjustment amount of the loading control parameters is calculated based on the evolution path of the virtual damage field; The physical loading device performs the actual loading operation corresponding to the loading control parameters, and simultaneously collects feedback mechanical data during the actual loading process. The feedback mechanical data is fused and compared with the evolution data of the virtual damage field to correct the internal weights of the mapping network; Based on the corrected mapping network, the driver parameter set of the virtual loading driver is iteratively optimized; Repeat the process of dynamically loading the driving parameter set for iterative optimization until the feedback mechanical data converges to the preset damage criterion threshold, and output the final strength performance index. The process of acquiring real-time physical field signals of jute fiber reinforced composite material samples through multi-source sensing units and extracting multi-scale mechanical features from the physical field signals specifically includes: An array of acoustic emission sensors was deployed to capture the acoustic emission signals of the sample under small loads, while a distributed fiber optic sensor network was deployed to measure the full-field strain distribution on the sample surface. The acoustic wave emission signal and the full-field strain distribution data are acquired synchronously to form a time-aligned physical field signal sequence; Time-frequency joint analysis was performed on the physical field signal sequence to separate the characteristic frequency band energy of fiber breakage, matrix cracking and interface debonding from the acoustic emission signal; Extract the gradient rate of change and strain energy density distribution cloud map of the strain concentration region from the full-field strain distribution data; By spatially correlating the energy of the characteristic frequency band with the gradient change rate of the strain concentration region, the core point of damage initiation is identified. Based on the strain energy density distribution cloud map and the damage initiation core point, the energy release rate trend of each damage region at the mesoscale is calculated. The spatial coordinates of the core point of damage initiation, the trend of the energy release rate, and the time-varying trajectory of the energy in the characteristic frequency band are summarized to form the multi-scale mechanical feature data set required for this detection. The process of establishing the mapping network between the multi-scale mechanical characteristics and the internal damage state of the specimen includes: Normalize each data item in the multi-scale mechanical feature data set to the same dimension space; Construct a hybrid neural network model containing convolutional layers and long short-term memory layers, wherein the convolutional layers are used to process strain energy density distribution cloud map data with spatial topological relationships, and the long short-term memory layers are used to process time-varying trajectory data of the energy in the characteristic frequency band. The spatial coordinates of the core point of damage initiation and the trend of energy release rate are used as guiding signals for the attention mechanism and input into the hybrid neural network model. The hybrid neural network model was trained using mechanical characteristics and damage morphology images of jute fiber reinforced composite material specimens with known damage states from historical tests. After training, the hybrid neural network model can output a quantitative damage state matrix that describes the internal microcrack density, crack propagation direction, and interface damage area of ​​the sample based on the input real-time multi-scale mechanical characteristics.

2. The method for testing the strength properties of jute fiber reinforced composite materials as described in claim 1, characterized in that, The process of constructing a virtual load driver and calibrating the virtual load driver according to a preset load path to generate a driver parameter set includes: Create a virtual loading driver model in a numerical simulation environment that matches the geometry and dynamic characteristics of the physical loading device; Define the initial load-displacement control logic, actuator response delay time, and load holding accuracy parameters for the virtual load driver model; Define a preset loading path that includes a linear loading segment, a constant load holding segment, and a cyclic loading segment; In the numerical simulation environment, the virtual load driver model is applied to a standard reference specimen model and runs along the preset load path; Record the sequence of deviations between the actual output load curve and the theoretical command load curve of the virtual load driver model throughout the entire operation process; Based on the deviation sequence, the internal control parameters of the virtual load driver model are dynamically adjusted using a reverse compensation algorithm until the actual output load curve and the theoretical command load curve match a set threshold. Archive all internal control parameters of the virtual load driver model at this time to form the initial driver parameter set for this test.

3. The method for testing the strength properties of jute fiber reinforced composite materials as described in claim 2, characterized in that, The step of importing the virtual loading driver into the mapping network, using the multi-scale mechanical features as input conditions, and driving the virtual loading driver to perform dynamic loading in the simulation environment includes: The multi-scale mechanical feature data set acquired in the current detection cycle is input into the trained hybrid neural network model to obtain the quantitative damage state matrix at the current moment. The quantitative damage state matrix and the current state parameters of the virtual loading driver model are input into a loading decision agent program. The loading decision agent program is based on a reinforcement learning framework. According to the current damage state and the loading target, it selects a loading action from a preset action space. The loading action includes increasing the load, decreasing the load, or maintaining the current load. The virtual loading driver model performs a one-step loading operation on a virtual specimen model that is consistent with the current specimen state in the numerical simulation environment, based on the loading action selected by the loading decision agent program and the initial driving parameter set. After the loading operation is completed, the mechanical state of the virtual sample model is updated, and a new set of virtual multi-scale mechanical features corresponding to this step is simulated and generated.

4. The method for testing the strength properties of jute fiber reinforced composite materials as described in claim 3, characterized in that, The real-time capture of the evolution path of the virtual damage field inside the sample during the dynamic loading process includes: In the numerical simulation environment, a damage evolution calculation rule based on phase field theory is embedded into the virtual sample model; After each loading operation is performed by the virtual loading driver model, the damage evolution calculation rule is triggered to perform an iterative calculation. Each iteration outputs the damage variable values, main crack propagation length, and number of branch cracks at each integration point within the virtual specimen model described in this step. The damage variable values ​​output from multiple consecutive loading steps are arranged in chronological order to form the evolution sequence of the virtual damage field in the time dimension; The damage field data at each time point in the evolution sequence are spatially reconstructed in three dimensions to generate a series of time-indexed damage field volumes. The series of time-indexed damage field volumes coherently describe the evolution path of the complete virtual damage field from damage initiation to macroscopic destruction.

5. The method for testing the strength properties of jute fiber reinforced composite materials as described in claim 4, characterized in that, The real-time adjustment amount of the loading control parameters calculated based on the evolution path of the virtual damage field includes: The evolution path of the virtual damage field was analyzed to identify the critical stage of accelerated damage propagation. Extract the virtual multi-scale mechanical features one step before the critical stage, and calculate the feature matching degree between them and the physical field signals actually obtained from the sample. If the feature matching degree is lower than the preset tolerance, it is determined that the virtual prediction deviates from the actual state; Based on the predicted damage expansion amount in the evolution path of the virtual damage field and the assessment results of the current actual damage state, the load increment or decrement required to resynchronize the two is calculated. Combining the load control accuracy and response speed of the virtual load driver model, the load increment or decrement is converted into specific adjustment values ​​and timings for the load setpoint in the physical loading device control system. These specific adjustment values ​​and timings are the real-time adjustment amounts of the load control parameters.

6. The method for testing the strength properties of jute fiber reinforced composite materials as described in claim 5, characterized in that, The step of executing the actual loading operation corresponding to the loading control parameters through a physical loading device, and simultaneously collecting feedback mechanical data during the actual loading process, includes: The real-time adjustment amount of the loading control parameters, i.e. the specific adjustment value and adjustment timing, is sent to the servo controller that controls the physical loading device. The servo controller drives the actuator of the physical loading device to apply mechanical load to the jute fiber reinforced composite material sample according to the adjusted load set point. While applying the load, the multi-source sensing unit is triggered to start a new round of data acquisition to obtain the acoustic emission signal and full-field strain distribution data under this actual loading operation. The acoustic wave emission signal and the full-field strain distribution data collected this time, together with the actual load value, displacement value and timestamp output by the physical loading device, are packaged together into the feedback mechanical data package for the current step.

7. The method for testing the strength properties of jute fiber reinforced composite materials as described in claim 6, characterized in that, The step of fusing and comparing the feedback mechanical data with the evolution data of the virtual damage field to correct the internal weights of the mapping network includes: New acoustic wave emission signals and full-field strain distribution data are parsed from the feedback mechanics data package, and new multi-scale mechanical features are extracted. From the evolution path of the virtual damage field, extract the predicted virtual multi-scale mechanical features corresponding to the current actual loading step. Calculate the difference vectors of the new multiscale mechanical features and the corresponding virtual multiscale mechanical features in each feature dimension; The difference vector is used as a loss signal and fed back to the hybrid neural network model; The backpropagation algorithm is used to fine-tune the connection weights of the convolutional layer and the long short-term memory layer in the hybrid neural network model based on the loss signal, so that the network's predicted output of the current sample damage state is closer to the state reflected by the actual feedback data.

8. The method for testing the strength properties of jute fiber reinforced composite materials as described in claim 7, characterized in that, The iterative optimization of the driver parameter set of the virtual load driver based on the modified mapping network includes: The response of the virtual sample model in the next loading step is re-predicted using the weighted hybrid neural network model. Based on the re-predicted damage state, evaluate the control performance of the current set of drive parameters used by the virtual load driver model; If the control performance index decreases, the parameter optimization routine is initiated to perform a targeted search within the parameter space of the driving parameter set to find a new combination of parameters that can improve the control performance index. The internal control parameters of the virtual load driver model are updated using the newly found parameter combinations to form an optimized set of driver parameters. The optimized set of driving parameters is applied to the virtual load driver model to guide the calculation of real-time adjustments to subsequent load control parameters, thereby initiating a new round of load-feedback-correction cycle.

Images