Early damage identification and intelligent early warning method based on real-time networking strain

Abstract

The invention discloses an early damage identification and intelligent early warning method based on real-time networking strain. The method comprises the following steps: arranging a plurality of sensors to collect strain information; performing normalization, temperature drift compensation and standardization processing on the strain information to obtain a fusion data time sequence corresponding to the sensitive area; counting the features to obtain damage index results of the features; a corresponding damage label is obtained; constructing a neural network model for damage state classification; training to obtain a trained neural network model; and carrying out online damage identification to obtain the probability of the damage level and carrying out early warning interpretation. According to the invention, real-time identification, intelligent early warning and state evaluation of complex structure damage can be realized, so that the safety and reliability of the test are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent technology for structural health monitoring and mechanical testing. Specifically, it relates to an early damage identification and intelligent early warning method based on real-time networked strain signals. It is particularly suitable for application scenarios where large and complex structures are subjected to mechanical testing under dynamic load conditions, and where the changes in structural state are monitored in real time and damage is warned. Background Technology

[0002] During structural mechanics testing, especially in structures subjected to complex multi-source loads (including static pressure, impact loads, and alternating fatigue), the strain distribution in localized areas of the material may undergo microscopic changes with the load, leading to initial damage such as microcracks and interface separation. If this initial damage is not identified and controlled in a timely manner, it can easily expand into structural failure events, such as weld cracking, shell perforation, or stress-over-limit yielding, severely impacting test safety and the accuracy of structural service assessment.

[0003] Currently, structural damage detection mainly relies on non-destructive testing methods, such as visual inspection, ultrasonic testing, X-ray testing, and acoustic emission technology. These methods typically require interrupting the test for detection or analyzing the results after the test, lacking real-time capability and dynamic continuity, and failing to effectively track changes in the structure's response during loading, often missing the window of early damage warning.

[0004] Although some technologies have introduced strain sensor arrays to monitor the strain response of structures, most existing deployment methods adopt fixed-point and sparse distribution strategies, which makes it difficult to achieve full coverage monitoring of key stress areas in complex structures with multiple regions and scales (such as multi-cavity shells and large welded cylinders), and cannot capture typical damage evolution characteristics such as crack initiation paths and abrupt responses caused by stress concentration.

[0005] On the other hand, traditional strain data analysis methods are mostly based on static data or offline statistical models, which cannot handle the temporal, non-stationary, and abrupt characteristics of structural strain responses. These limitations make it difficult to effectively identify damage in the early stages, resulting in delayed model warnings or large identification errors.

[0006] Furthermore, most current damage identification models employ rule-based or linear regression strategies, lacking the ability to deeply mine and model multi-dimensional dynamic strain signals, and thus failing to meet the modeling requirements of nonlinear evolution processes under complex damage mechanisms. Especially in stress coupling and multi-scale strain responses, a systematic intelligent identification and response loop has not yet been formed.

[0007] Therefore, existing technologies have key shortcomings in the following aspects: 1. Insufficient spatial monitoring resolution: The number of strain sensors is limited and the deployment density is low, making it impossible to cover high-risk areas and difficult to form a global state perception. 2. Insufficient dynamic modeling capability: Unable to effectively extract key temporal features during structural loading, such as abrupt changes in strain rate and spectral drift; 3. Lack of intelligent early warning mechanism: Most models rely on manual threshold setting and lack data-driven prediction capabilities, resulting in insensitive early warnings; 4. Weak system feedback and adaptive capabilities: It lacks a real-time model correction and feedback mechanism and cannot automatically adjust the identification strategy according to new damage patterns.

[0008] In summary, there is an urgent need to develop a novel structural health monitoring method that possesses high-density network strain monitoring capabilities, integrates time-series intelligent modeling algorithms, and can identify and respond to structural damage states in real time and in a graded manner, in order to meet the safety and stability assessment requirements of complex structures during the testing process.

[0009] Existing structural damage detection and early warning technologies have many shortcomings in dealing with mechanical tests of complex structures, mainly in the following aspects: Existing data analysis methods are mainly based on static strain values ​​for evaluation, lacking the ability to model the dynamic changes of strain signals with loading time, and thus failing to effectively identify early abnormal behaviors in the structural response process. Existing damage identification systems mostly rely on threshold judgments or empirical rules, lack sufficient intelligence, cannot achieve prediction and early warning based on real-time data, have delayed responses, and are difficult to avoid risks in a timely manner. Summary of the Invention

[0010] The technical problem solved by this invention is to overcome the shortcomings of the prior art. This invention provides an early damage identification and intelligent warning method based on real-time networked strain signals. This method is applicable to state monitoring and risk control in the process of structural mechanics testing. It can realize real-time identification, intelligent warning and state assessment of complex structural damage, thereby improving the safety and reliability of the test.

[0011] The technical solution of this invention is: The method for early damage identification and intelligent warning based on real-time networked strain includes the following steps: 1) Deploy multiple sensors in the sensitive area to collect strain information; 2) Normalize, compensate for temperature drift, and standardize the strain information to obtain the fused data time series corresponding to the sensitive area. ; 3) Time series data of fused data corresponding to sensitive areas Statistical characteristics are used to obtain damage index results for the features; the features include: fused data time series. strain mean, strain standard deviation, and overall rate of change Hurst exponent for dynamic feature extraction ; 4) Classify the damage index results into damage levels to obtain corresponding damage labels. ; 5) A three-layer bidirectional recurrent neural network is used to construct a neural network model for damage state classification; 6) Combine the damage index results obtained in step 3) with the damage labels obtained in step 4). The model is divided into a training set and a validation set, and the neural network model constructed in step 5) is trained to obtain a fully trained neural network model. ; 7) Use the neural network model trained in step 6) Online damage identification is performed on structural products to obtain the probability of damage level; 8) The probability of the damage level output in step 7) is used to provide early warning and judgment on the damage status of the structural product.

[0012] Preferably, step 2) involves obtaining the fused data time series corresponding to the sensitive area. The method is as follows: 21) The collected strain information is subjected to zero-point normalization processing to obtain the strain information after zero-point normalization. ;

[0013] in, The sensor number; Indicates the first Each sensor at time strain readings, For the first The nominal value of each sensor; 22) Strain information after zero-point normalization Temperature drift compensation is performed to obtain strain information after temperature drift compensation. ;

[0014] in, For temperature sensitivity coefficient, This represents the change in temperature between the sensor installation location and a reference temperature value. 23) The strain information from the temperature drift compensation process is standardized to obtain the fused data time series corresponding to the sensitive region. ;

[0015]

[0016]

[0017] in, To standardize data, The weighted fusion coefficient, This represents the total number of sensors.

[0018] Preferably, in step 3), the time series data is fused. strain mean, strain standard deviation, and overall rate of change Hurst exponent for dynamic feature extraction Specifically:

[0019]

[0020]

[0021]

[0022]

[0023] in, Let be the range of strain of the i-th sensor within the window. .

[0024] Preferably, in step 4) ;in, =1 indicates a lossless state. =2 indicates minor damage. =3 indicates severe damage; Determine damage labels The method is as follows: For minor damage: peak strain increment should be less than 5%, strain standard deviation. The rate of change is less than 30%; the rate of change of the power spectral density integral is less than 3%; Hurst exponent Less than 0.7, the spectral energy ratio in the high-frequency region of the waveform signal is less than 0.15; Moderate damage corresponds to: peak strain increment between 5% and 20%, strain standard deviation. The Hurst exponent increases by 30%-100% from its initial state. Increased to 0.7-0.85, the spectral energy ratio in the high-frequency region of the waveform signal is between 0.15 and 0.3; Severe damage corresponds to: peak strain increment exceeding 20%, strain standard deviation... A change rate greater than 100% indicates the Hurst exponent. The spectral energy ratio in the high-frequency region of the waveform signal is greater than 0.85, and the spectral energy ratio in the high-frequency region is greater than 0.3. If the characteristic damage index results belong to different damage levels, the damage label shall be defined according to the most severe damage level. .

[0025] Preferably, in step 5), the input to the neural network model is:

[0026] in, This represents the j-th dimension of the input feature at time t, where d is the dimension of the input feature; corresponding to the mean strain, standard deviation of strain, and overall rate of change. Hurst exponent for dynamic feature extraction When d equals 4.

[0027] Preferably, the output of the neural network model is:

[0028] in, For prediction The probability of.

[0029] Preferably, the loss function of the neural network model is cross-entropy.

[0030] Preferably, the method for performing early warning interpretation in step 8) is as follows: Level 1 Warning: 0.7 0.85, minor damage; Level II Warning: 0.85 0.95, moderate damage; Level 3 Warning: 0.95, severe damage.

[0031] in, for The probability of.

[0032] Compared with the prior art, the advantages of the present invention are mainly reflected in the following aspects: A data analysis method based on dynamic feature enhancement is employed to fuse and preprocess the collected strain data, and to extract key dynamic features by combining statistical indicators (such as variance and rate of change) and nonlinear characteristic parameters (such as the Hurst exponent). This method can accurately reflect the non-stationary response of the structure during loading and capture potential damage signs.

[0033] A recurrent neural network (RNN) is introduced for damage identification modeling. Leveraging its ability to model time-series data, strain feature sequences are trained to correspond with damage states, constructing a multi-classification model to output the current damage probability distribution of the structure. This model is capable of processing dynamic, multi-dimensional, and high-frequency data, improving the sensitivity for identifying early, minor damage.

[0034] A real-time intelligent early warning mechanism is established, setting graded thresholds based on model output results (e.g., 0.7 for mild damage, 0.85 for moderate damage, and 0.95 for severe damage) to achieve graded early warning responses. The system generates early warning prompts through a visual interface and provides operational suggestions based on the risk level, ensuring the safety of personnel and equipment during the experiment.

[0035] An online optimization mechanism for damage identification and assessment models is constructed. During the continuous experiment, new data collected in real time can be fed back for fine-tuning or retraining of model parameters, thereby improving the adaptability and robustness of the identification system and forming a closed-loop optimization path of model-data-feedback. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation To better describe the present invention, the present invention will be described in detail below with reference to schematic diagrams and examples.

[0037] This invention relates to an early damage identification and intelligent warning method based on real-time networked strain, such as... Figure 1 As shown, the steps are as follows: 1) Structural analysis and sensor deployment planning This step aims to develop a strain sensor placement plan based on the geometry, stress characteristics, and historical failure modes of the structure under test. High-risk areas of the structure (such as stress concentration zones, weld seams, and thin-walled sections) are identified through finite element analysis, engineering experience, or historical test data, and these areas are prioritized for strain sensor placement.

[0038] The implementation methods include: Perform 3D modeling of the target structure; Determine the main loading paths (axial compression, radial expansion, alternating bending, etc.); Use stress-strain simulation tools (such as ANSYS and Abaqus) to obtain preliminary strain field distribution maps; Sensing points are deployed in selected sensitive areas on the distribution map, with multiple sensors deployed in each sensitive area.

[0039] The output for each sensitive area is: the total number of sensors in the current sensitive area. Number of each sensor Spatial coordinates and deployment density (Number of points per unit length in the sensitive area), i∈[1,N].

[0040] Each sensor in the sensitive area is individually labeled. The purpose of this step is to ensure that the subsequently acquired data has spatial representativeness and damage response sensitivity.

[0041] 2) Strain sensor installation and network deployment After determining the deployment plan, multi-channel strain gauges were installed, and a data synchronous acquisition network was established. The selected sensors were resistance strain gauges or fiber Bragg gratings (FBGs), and the installation method (adhesive, wrap-around, or embedded) was chosen based on the material, shape, and environmental tolerance of the structural surface.

[0042] The key points for network deployment are as follows: All sensors have the same serial number and are configured with a unique address; The data acquisition system uses a unified clock to ensure that the time synchronization error is less than 1 millisecond; The communication protocol can be wired (such as CAN, Modbus) or wireless (such as ZigBee, LoRa). The network topology can be bus, star, or mesh, with star topology recommended for situations where time synchronization is critical.

[0043] After installation, perform bridge zeroing and preliminary system verification to ensure zero drift stability. Within.

[0044] The output result is a sensor matrix that has been successfully networked. :

[0045] Using sensor matrix Used to verify the correctness of system deployment and communication, and as an index of signal channels during data acquisition.

[0046] 3) Strain data acquisition and preliminary calibration Sampling frequency set to The sampling period is ; The data format is a time-series vector, and each sampling point is:

[0047] in, Indicates the first Each sensor at time The strain reading.

[0048] The initial calibration uses zero-point normalization of the unloaded baseline:

[0049] Before correcting the data, it is necessary to collect the temperature change data at each sensor installation location. Temperature data is obtained by configuring temperature sensors that correspond one-to-one with the strain channels.

[0050] Finally, temperature drift compensation is performed according to the following formula:

[0051] Among them, temperature sensitivity coefficient The value is obtained through pre-experimental calibration. Specifically, the temperature of the structure is controlled to rise under no-load conditions, and the strain output is recorded. The rate of change is then fitted based on a linear relationship.

[0052] The output is a set of original strain data matrices after time synchronization and compensation. :

[0053] The data acquisition system records timestamps in each round of synchronous sampling. The sampling period is Among them, sampling frequency Set to during system initialization Therefore, a time series can be represented as:

[0054] in, .

[0055] Time series The matrix will serve as the basic input for subsequent feature extraction and modeling, and in the data preprocessing and fusion stages, the timestamp... Used for dynamic feature extraction operations such as synchronous filtering and frequency domain analysis.

[0056] 4) Data preprocessing and fusion computing The collected raw strain data is processed, including denoising, standardization, and multi-channel fusion, to improve data stability and anti-interference ability, and to lay the foundation for feature extraction. The strain data matrix obtained in step three is then extracted. and time series .

[0057] use First-order low-pass filter, cutoff frequency Used to eliminate high-frequency interference:

[0058] in, ω is the angular frequency.

[0059] Standardize each sensor data to Obtain at the sampling time Standardized data Specifically:

[0060] After standardization, each channel becomes Using weighted fusion, for the same sampling time... Standardized data from all sensors The resulting fused data time series Specifically:

[0061]

[0062] in, The weighted fusion coefficient, It can be determined based on the sensor's location, sensitivity, or experience.

[0063] After performing filtering, normalization, and data fusion, the original strain data matrix The dataset will be processed into a cleaned dataset. Its definition is as follows: After fusion processing, the fused data time series is obtained. That is, multi-channel synthesized values, and the final cleaned dataset. Defined as:

[0064] This step yields the cleaned and fused time series data. Standardized data of each sensor and the dataset used for feature computation .

[0065] 5) Feature extraction and parametric modeling input preparation Key statistical and dynamic features are extracted from the fused strain sequence data to construct an input feature vector, which serves as the training input for the subsequent neural network model.

[0066] Statistical features are used to extract damage indicators, including but not limited to:

[0067]

[0068] in, The mean strain sequence of the i-th sensor within the time window T; This represents the strain standard deviation of the i-th channel within the time window T.

[0069]

[0070] in, This represents the overall rate of change of the i-th sensor channel within the time window T.

[0071] Hurst exponent for dynamic feature extraction This is used to measure the long-term correlation of a sequence.

[0072] in, Let be the strain range of the i-th channel within the window. Standard deviation, The sequence length is the same as above. Construct the overall feature vector:

[0073] This step yields multidimensional feature vectors. With strain feature samples used for training ,in, The damage labels are obtained through a method described below for damage level classification. During the training data acquisition process, some structural samples were pre-defined with different degrees of damage, for example: This means indexing data from the sample dimension, representing the k-th sample.

[0074] Minor damage: Grooving of structural surfaces, design of localized stress concentration; Moderate damage: weld defects, embedded cracks, or shear cuts; Severe damage: through cracks, fractures, etc.; During the data acquisition phase, through loading tests and strain response characteristic analysis, based on the structural peak strain increment and strain standard deviation... Power spectral density integral change rate, Hurst exponent Damage indicators such as the spectral energy ratio in the high-frequency region of the waveform signal are used to classify the damage level of the samples, and the highest damage level described by the damage indicators is used as the final output damage label. .

[0075] The damage severity classification method is defined as follows: Minor damage: Peak strain increment should be less than 5%, strain standard deviation The rate of change is less than 30%; the rate of change of the power spectral density integral is less than 3%; Hurst exponent Less than 0.7, the spectral energy ratio in the high-frequency region of the waveform signal is less than 0.15.

[0076] Moderate damage: Peak strain increment should be between 5% and 20%, strain standard deviation The Hurst exponent increases by 30%-100% from its initial state. Increased to 0.7-0.85, the spectral energy ratio of the high-frequency region of the waveform signal is between 0.15 and 0.3.

[0077] Severe damage: Peak strain increment should exceed 20%, strain standard deviation A change rate greater than 100% indicates the Hurst exponent. The spectral energy ratio in the high-frequency region of the waveform signal is greater than 0.85, and the spectral energy ratio in the high-frequency region is greater than 0.3.

[0078]

[0079] In this system, 1 represents an undamaged state, 2 represents minor damage, and 3 represents severe damage. This label set is used during the training phase to supervise the neural network in learning the mapping relationship between structural states and strain features.

[0080] 6) Construct a neural network model for damage state classification, possessing deep learning capabilities for time-series strain features. Use a sample dataset. .

[0081] The model uses a three-layer bidirectional recurrent neural network (Bi-RNN) with tanh as the activation function and softmax as the output layer.

[0082] , in, This represents the j-th dimension of the input feature at time t (such as mean strain, standard deviation of strain, total rate of change). Hurst exponent for dynamic feature extraction (etc.), where d is the input feature dimension.

[0083] , in, Let be the hidden state vector at time step t. and These are the input weight matrix and the state transition matrix, respectively. This is a bias term.

[0084] , in, , To predict the probability of belonging to damage level c, c = 1, 2, 3.

[0085] Number of nodes per layer of the model: Sequence length: At time 1; the loss function is cross-entropy:

[0086] in, To predict probabilities for the model, This represents the total number of damage levels. The label component represents the true damage state. When the true damage level of the current sample is the i-th type of damage, ,otherwise .

[0087] This step yields the initialized network model structure. ; 7) Model training and performance evaluation Using labeled strain feature samples The neural network model is trained and its accuracy is evaluated using standard performance metrics.

[0088] Divide the data into training set and verification set .

[0089] The model was trained using the cross-entropy loss function constructed in step 6), with the Adam algorithm as the optimization algorithm and a learning rate set to [value missing]. The training rounds are Accuracy is evaluated once per round, and the accuracy evaluation method is as follows:

[0090] Meanwhile, the F1 score is used as a weighted harmonic average of precision and recall to determine the robustness of the model.

[0091] Standard for achievement: Accuracy F1 score .

[0092] Finally, the trained damage recognition model is obtained. ; 8) Online damage identification and early warning level output The trained model is deployed to a real-time strain acquisition system to achieve online damage identification and automatically trigger graded early warning signals based on the identification results. First, the model infers and outputs a damage probability vector:

[0093] This represents the feature vector extracted from the time series t, such as:

[0094] in, Indicates the prediction as damage label The probability of.

[0095] Set warning level: Level 1 Warning: 0.7 0.85, minor damage, warrants attention.

[0096] Level II Warning: 0.85 0.95, moderate damage, examination recommended.

[0097] Level 3 Warning: 0.95, severe damage, immediate response.

[0098] Output signals to the control interface or alarm module, indicating the damage level, prediction confidence level, and current sensing area.

[0099] This invention systematically solves the problems of insufficient sensing range, poor timeliness, weak intelligence, and lack of feedback loop in traditional structural monitoring schemes by comprehensively applying dynamic data processing, intelligent time sequence recognition, and hierarchical response mechanisms. It has significant novelty and inventiveness and can be widely used in structural testing and health management tasks in high-safety-level fields.

[0100] Example Early damage intelligent identification and early warning test of a certain type of high-pressure cylinder This embodiment uses a high-pressure sealed cylinder as the test object to verify the practical application effect of the early damage identification and intelligent early warning method based on real-time networked strain described in this invention. The cylinder is used for high-carrier wave testing of a hydraulic system; it is made of 7075 aluminum alloy, has a wall thickness of approximately 6 mm, and operates at a pressure of up to 40 MPa.

[0101] I. Sensor Deployment and Network Configuration: The structural 3D modeling was performed using SolidWorks, and the load analysis was performed using ANSYS. Based on the simulation results, 12 strain sensors were arranged along the circumference and installed using an adhesive method. A sensor point is set at every 30° to cover the upper and lower welding areas of the cylinder; Using a wireless acquisition system (ZigBee protocol), sampling frequency Time synchronization error ; Each sensor is equipped with a thermistor to achieve temperature monitoring and compensation.

[0102] II. Strain Data Acquisition and Preprocessing: Initial no-load zeroing calibration; Temperature drift compensation coefficient (Obtained through calibration); Using a fourth-order low-pass filter, the cutoff frequency is... ; After normalization, feature vectors are constructed and then fused in real time.

[0103] III. Damage Preset and Labeling: Minor damage: Microgrooves on the cylinder surface, depth ; Moderate damage: Embedded cracks (not penetrating) in the weld defect area; Severe damage: Axial artificial through-crack, approximately [length missing] ; Each type of damage was repeated 5 times and used as a labeled sample for training.

[0104] IV. Model Training and Deployment: Network structure: Three-layer Bi-RNN, 128 hidden nodes, sequence length ; Optimizer: Adam, learning rate 0.001, training epochs 100; The model achieved an accuracy of 93.7% and an F1 score of 0.91 on the validation set. Real-time deployment allows for deployment after damage occurs. The system can identify and issue warnings within the system.

[0105] V. Early Warning Results and Verification: For moderate damage, the system output This triggered a Level II warning. In cases of severe injury, This triggers a Level 3 warning and indicates the sensor number area; Compared with post-procedural ultrasound, the accuracy rate reached Misidentification is concentrated at the boundary between minor damage and normal state; the entire process is delayed. To meet the engineering requirements.

[0106] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any person skilled in the art can make possible variations and modifications to the technical solutions of the present invention using the disclosed methods and techniques without departing from the spirit and scope of the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the content of the technical solutions of the present invention, shall fall within the protection scope of the present invention. Where there is no conflict, the embodiments of this application and the technical features thereof can be combined with each other.

[0107] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A method for early damage identification and intelligent early warning based on real-time networked strain, characterized in that, The steps include the following: 1) Deploy multiple sensors in the sensitive area to collect strain information; 2) Normalize, compensate for temperature drift, and standardize the strain information to obtain the fused data time series corresponding to the sensitive area. ; 3) Time series data of fused data corresponding to sensitive areas Statistical characteristics are used to obtain damage index results for the characteristics. Features include: fused data time series strain mean, strain standard deviation, and overall rate of change Hurst exponent for dynamic feature extraction ; 4) Classify the damage index results into damage levels to obtain corresponding damage labels. ; 5) A three-layer bidirectional recurrent neural network is used to construct a neural network model for damage state classification; 6) Combine the damage index results obtained in step 3) with the damage labels obtained in step 4). The model is divided into a training set and a validation set, and the neural network model constructed in step 5) is trained to obtain a fully trained neural network model. ; 7) Use the neural network model trained in step 6) Online damage identification is performed on structural products to obtain the probability of damage level; 8) The probability of the damage level output in step 7) is used to provide early warning and judgment on the damage status of the structural product.

2. The method for early damage identification and intelligent early warning based on real-time networked strain according to claim 1, characterized in that, Step 2) Obtain the fused data time series corresponding to the sensitive area. The method is as follows: 21) The collected strain information is subjected to zero-point normalization processing to obtain the strain information after zero-point normalization. ; in, The sensor number; Indicates the first Each sensor at time strain readings, For the first The nominal value of each sensor; 22) Strain information after zero-point normalization Temperature drift compensation is performed to obtain strain information after temperature drift compensation. ; in, For temperature sensitivity coefficient, This represents the change in temperature between the sensor installation location and a reference temperature value. 23) The strain information from the temperature drift compensation process is standardized to obtain the fused data time series corresponding to the sensitive region. ; in, To standardize data, The weighted fusion coefficient, This represents the total number of sensors.

3. The method for early damage identification and intelligent early warning based on real-time networked strain according to claim 2, characterized in that, Step 3) Merging data time series strain mean, strain standard deviation, and overall rate of change Hurst exponent for dynamic feature extraction Specifically: in, Let be the range of strain of the i-th sensor within the window. .

4. The method for early damage identification and intelligent early warning based on real-time networked strain according to claim 3, characterized in that, Step 4) ;in, =1 indicates a lossless state. =2 indicates minor damage. =3 indicates severe damage; Determine damage labels The method is as follows: For minor damage: peak strain increment should be less than 5%, strain standard deviation. The rate of change is less than 30%; the rate of change of the power spectral density integral is less than 3%; Hurst exponent Less than 0.7, the spectral energy ratio in the high-frequency region of the waveform signal is less than 0.15; Moderate damage corresponds to: peak strain increment between 5% and 20%, strain standard deviation. The Hurst exponent increases by 30%-100% from its initial state. Increased to 0.7-0.85, the spectral energy ratio in the high-frequency region of the waveform signal is between 0.15 and 0.3; Severe damage corresponds to: peak strain increment exceeding 20%, strain standard deviation... A change rate greater than 100% indicates the Hurst exponent. The spectral energy ratio in the high-frequency region of the waveform signal is greater than 0.85, and the spectral energy ratio in the high-frequency region is greater than 0.

3. If the characteristic damage index results belong to different damage levels, the damage label shall be defined according to the most severe damage level. .

5. The method for early damage identification and intelligent early warning based on real-time networked strain according to any one of claims 1-4, characterized in that, In step 5), the input to the neural network model is: in, This represents the j-th dimension of the input feature at time t, where d is the dimension of the input feature; corresponding to the mean strain, standard deviation of strain, and overall rate of change. Hurst exponent for dynamic feature extraction When d equals 4.

6. The method for early damage identification and intelligent early warning based on real-time networked strain according to claim 5, characterized in that, The output of the neural network model is: in, For prediction The probability of.

7. The method for early damage identification and intelligent early warning based on real-time networked strain according to claim 6, characterized in that, The loss function of a neural network model is cross-entropy.

8. The method for early damage identification and intelligent early warning based on real-time networked strain according to claim 7, characterized in that, Step 8) The method for interpreting early warnings is as follows: Level 1 Warning: 0.7 0.85, minor damage; Level II Warning: 0.85 0.95, moderate damage; Level 3 Warning: 0.95, severe damage. in, for The probability of.

Images