Automatic Identification and Evaluation System for Cable-Stayed Bridge Cable Force Based on Dynamic Modeling

Abstract

This invention discloses an automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling, specifically relating to the field of dynamic modeling. The system includes: a sensor preprocessing module, a baseline modeling module, a hybrid prediction module, an evaluation and early warning module, and a self-optimizing iterative module. It proposes a pure data-driven AI model that integrates multi-source data fusion. This model fuses vibration signal data features with prior knowledge of string vibration through a dual-channel architecture and is trained using a composite loss function constrained by physics and rate of change, achieving intelligent identification of cable forces under complex working conditions. Furthermore, it constructs an intelligent evaluation framework with online cross-validation, trend prediction, and multi-source correlation, achieving model self-optimization and long-term adaptation through incremental learning. Addressing the problems of weak generalization, physical inconsistency, and poor adaptability of data-driven AI in engineering, this invention proposes a general architecture integrating multiple technologies to achieve accurate identification of tension states in multiple scenarios.

Description

Technical Field

[0001] This invention relates to the field of dynamic modeling, and more specifically, to an automatic identification and evaluation system for cable forces in cable-stayed bridges based on dynamic modeling. Background Technology

[0002] With the development of bridge structural health monitoring technology, existing technologies for identifying and assessing cable forces in cable-stayed bridges mainly include the frequency method based on string vibration theory, the inversion method based on finite element analysis, and intelligent identification methods based on data-driven algorithms. The traditional frequency method collects cable vibration signals and calculates the force using the theoretical relationship between natural frequencies and cable forces, offering advantages such as ease of implementation and wide engineering application. The finite element inversion method, on the other hand, establishes a structural numerical model and inversely calculates parameters from the measured response, allowing for consideration of more complex boundary conditions and structural characteristics. In recent years, with the development of artificial intelligence technology, neural network-based cable force identification methods have gradually emerged. By extracting features and learning patterns from vibration signals, these methods achieve automatic prediction and state assessment of cable forces. Some systems also combine multi-source sensor data to conduct structural state assessment and early warning analysis.

[0003] However, existing technologies still have certain shortcomings. On the one hand, traditional physical model-based calculation methods are highly dependent on the accuracy of boundary conditions and parameters. Under complex working conditions, environmental temperature changes, and long-term performance degradation, model errors are prone to accumulate, making it difficult to achieve long-term stable application. On the other hand, while purely data-driven artificial intelligence models have strong nonlinear fitting capabilities, they often lack physical constraints, resulting in insufficient generalization ability, inconsistencies between predicted results and actual mechanical laws, and weak adaptive capabilities when data distribution changes or structural state evolves, making it difficult to achieve long-term reliable online operation. In addition, existing systems mostly focus on single identification functions, with insufficient consideration for cable force change trends, anomaly root cause analysis, and model self-optimization mechanisms. A systematic solution integrating dynamic modeling, twin modeling, intelligent identification, and adaptive evaluation has not yet been formed, making it difficult to meet the needs of refined health monitoring throughout the entire life cycle of complex bridge structures. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the present invention provides an automatic identification and evaluation system for cable tension of cable-stayed bridges based on dynamic modeling, which solves the problems mentioned in the background art through the following scheme.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling, comprising: The sensor preprocessing module is used to collect and process the vibration and image data of the cable, and assign a unique quality level label to it based on the noise level and working conditions. The benchmark modeling module is used to calibrate the healthy benchmark cable force range and safe rate of change threshold of the cable based on the string vibration theory and the selected high-quality label data, and to construct a theoretical cable force calculation model that includes temperature, stiffness and real-time calibration correction coefficients. The hybrid prediction module is used to extract and fuse the data-driven features and physical prior features of the vibration signal. It identifies the cable force by using a hybrid driving model that integrates physical consistency loss and rate of change constraint loss, and performs online cross-validation and dynamic calibration using the theoretical cable force calculation model to output the final identified cable force. The assessment and early warning module is used to compare the final identified cable force with the healthy baseline cable force range and the safe rate of change threshold to achieve a three-level state assessment, and combine trend prediction and visual impairment identification to perform abnormal root cause correlation, generate a report and execute graded early warning. The self-optimizing iterative module monitors the model accuracy based on the three-level state evaluation results during system operation and the corresponding historical final identified cable force data. When performance degrades, it updates the hybrid driving model through incremental learning and coordinates the adjustment of the healthy benchmark cable force range, the safe rate of change threshold, and various correction coefficients in the theoretical cable force calculation model, while iteratively optimizing the system operating parameters.

[0006] The technical effects and advantages of this invention are as follows: 1. This invention constructs a neural network architecture that integrates data-driven and physical guidance, and designs a composite loss function that includes physical consistency constraints for training, successfully enabling the AI ​​model to have both strong data fitting ability and solid physical law reliability. 2. Based on the proposed high-reliability cable force identification model, this invention constructs a progressive intelligent assessment chain from three-level state assessment and short-term trend prediction to multi-source root cause correlation. This system can not only identify anomalies, but also use AI models to predict changing trends and integrate visual data to lock in potential damage through correlation reasoning, thus realizing an assessment leap from shallow alarm to deep intelligent diagnosis. 3. This invention designs a model self-optimization iteration module, which enables the system to have a continuous performance monitoring and incremental learning mechanism. When the model performance degradation is detected, it can automatically trigger the update of the core AI model parameters using new data, and can coordinately adjust the relevant physical parameters and thresholds, so that the entire system can adapt to the long-term evolution of the bridge state, ensuring the accuracy and reliability of the AI ​​model throughout its entire life cycle. 4. This invention uses a physical augmentation AI model as the intelligent core, organically integrating high-fidelity data perception, high-precision intelligent recognition, in-depth evaluation and self-optimization. Each module is closely connected around the core AI model, forming a standardized and implementable overall solution, providing a systematic methodology and architectural reference for the intelligent health monitoring of complex engineering structures. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0008] Figure 2 This is a schematic diagram of the data preprocessing and hierarchical structure of the present invention.

[0009] Figure 3 This is a schematic diagram of the hybrid model cable force recognition structure of the present invention.

[0010] Figure 4 This is a schematic diagram of the status assessment and early warning generation structure of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] refer to Figures 1-4 The automatic identification and evaluation system for cable tension of cable-stayed bridges based on dynamic modeling shown includes: a sensor preprocessing module, a benchmark modeling module, a hybrid prediction module, an evaluation and early warning module, and a model self-optimization and iteration module.

[0013] The sensing preprocessing module specifically includes: S101. Multi-source sensor data acquisition deployment: Through standardized sensor unit configuration, precise installation and positioning, and synchronous transmission design, a stable and reliable monitoring sensor network is constructed to provide high-fidelity, spatiotemporally synchronized source data support for subsequent cable force identification and status assessment, ensuring the matching of technical parameters between the data acquisition stage and subsequent preprocessing and model calculation.

[0014] Integrated intelligent sensing unit configuration: Each cable to be monitored is equipped with an integrated intelligent sensing unit. The unit is encapsulated in an IP67-rated sealed protective shell, which has the ability to withstand wind loads, corrosion, and temperatures (-40℃~85℃) for outdoor use. The shell is designed to be lightweight (weight ≤500g) to avoid additional impact on the vibration characteristics of the cable.

[0015] Accelerometer Sensing Module: A MEMS triaxial accelerometer is selected to collect the transverse vibration acceleration signal of the cable perpendicular to the cable axis. The sampling frequency is fixed at 100Hz as the sole frequency resolution control parameter. This parameter is verified by the Nyquist criterion and can completely cover the natural vibration frequency range of the cable from 1 to 50Hz, meeting the resolution requirements of subsequent string vibration theory modeling for frequency data. The sensor range is set to ±8g, and the resolution g≤0.001g, ensuring the capture of weak vibration signals while avoiding overload distortion.

[0016] Visual sensing module: Equipped with a high-definition CMOS visual sensor, the image resolution is fixed at 1920×1080px as the sole image quality control parameter, with a pixel density ≥100dpi. It can clearly identify damage features such as cracks and corrosion larger than 0.1mm on the cable surface, and is adapted to the input accuracy requirements of the subsequent YOLOv8 damage detection algorithm. The sensor frame rate is set to 10fps to balance real-time performance and data transmission bandwidth. The lens adopts an anti-reflective coating design to reduce image interference in strong light and backlight environments.

[0017] Precise installation location calibration: The installation location of the intelligent sensing unit is based on the effective length of the cable. Quantitative calibration is performed to ensure the spatial representativeness of the collected data and the effectiveness of vibration feature extraction.

[0018] Installation point layout: The first installation point is located at the midpoint of the cable ( The first installation point is located at the antinode of the cable's lateral vibration, where the vibration response amplitude is the largest, allowing for precise capture of each natural frequency. The second installation point is located at a distance from the lower anchor point of the cable. This location is a concentrated area of ​​vibration energy, which can supplement the vibration details not covered at the midpoint, forming a two-point complementary monitoring system; the distance between the two installation points is fixed at [missing information]. This ensures that the data is evenly distributed in space.

[0019] Installation process requirements: During installation, a "snap-on fixing + epoxy adhesive reinforcement" method should be used to ensure that the sensor is tightly attached to the cable surface with a gap of ≤0.1mm to avoid signal distortion caused by loose installation; the sensitive axis direction of the accelerometer must be calibrated with a level to be strictly consistent with the lateral vibration direction of the cable with a deviation of ≤±5; the vision sensor lens should be vertically aligned with the cable surface, and the shooting range should cover the cable area with a diameter of ≥50cm to ensure no blind spots.

[0020] Data synchronization and transmission mechanism: Construct a three-level data transmission architecture of "unit-edge-cloud" to ensure the spatiotemporal consistency and real-time performance of data collection.

[0021] Time synchronization configuration: All intelligent sensing units are connected to a unified time synchronization system, which prioritizes the use of the PTP protocol with a synchronization accuracy of ≤1ms. It is compatible with the NTP protocol in complex network environments to ensure that the clock deviation of all sensing units in the whole bridge is less than 10 milliseconds, thus meeting the spatiotemporal alignment requirements for cable force comparison analysis between multiple cables.

[0022] Data transmission design: Raw acceleration data is stored in CSV format, including timestamps and triaxial acceleration values, while image data is stored in JPEG format, along with shooting timestamps and installation point numbers. Both data are transmitted via wired (fiber optic) or wireless (5G / LoRa) networks. AES-128 encryption is used to ensure data security during wireless transmission. Edge computing nodes are deployed in the bridge control room to receive and temporarily store data (storage capacity ≥1TB), supporting network outage caching and automatic retransmission after network recovery. Raw data is pushed to the cloud data receiving end in real time, with a transmission latency ≤200ms, providing timely support for subsequent edge-side preprocessing and cloud-based model inference.

[0023] S102. Data Preprocessing and Grading: Through standardized data cleaning and feature preservation, noise and interference factors are eliminated. At the same time, a quantitative quality grading system is established to provide high-quality, distinguishable standardized data packages for subsequent model training, benchmark calibration and anomaly analysis, ensuring that the preprocessing results accurately match the input requirements of subsequent modules.

[0024] Vibration signal preprocessing: An adaptive wavelet threshold denoising algorithm is used to process the original acceleration signal. The core objective is to preserve the natural frequency components of the vibration signal and remove environmental noise. The specific process is as follows: Signal decomposition: The db4 wavelet was selected as the basis function to perform 5-level wavelet packet decomposition on the original acceleration signal. After decomposition, one set of low-frequency approximation coefficients (containing the main vibration energy) and 31 sets of high-frequency detail coefficients (mainly noise components) were obtained. The number of decomposition levels was verified by experiments and can completely separate the vibration signal of the cable from 1 to 50 Hz and the high-frequency noise.

[0025] Threshold quantization: A unique denoising intensity parameter is used for the high-frequency detail coefficients after decomposition. Soft threshold quantization is performed, where The standard deviation of the original signal is used. This threshold has been verified by a large amount of experimental data. It can remove noise while avoiding frequency distortion caused by excessive smoothing. Signal reconstruction: Wavelet packet reconstruction is performed on the quantized high-frequency coefficients and the original low-frequency approximation coefficients to obtain the denoised clean vibration time history data. The signal-to-noise ratio (SNR) of the reconstructed signal is ≥30dB to ensure the accuracy of subsequent extraction of features such as multi-scale energy entropy and natural frequency.

[0026] Image data preprocessing: The original image is standardized to eliminate interference from lighting, blur, etc., while preserving damage features and details. The specific process is as follows: Grayscale conversion: The original color image is subjected to weighted average grayscale conversion using fixed weight coefficients [0.299, 0.587, 0.114]. These weight coefficients conform to the characteristics of human vision and can maximize the preservation of the grayscale difference between the cable surface damage and the background, avoiding interference from color information in damage identification.

[0027] Gaussian deblurring: A 3×3 Gaussian kernel is used to perform a convolution operation on the grayscale image, where... The convolution stride is set to 1 to ensure that the image size remains unchanged. This operation can effectively eliminate slight blur caused by lens shake and atmospheric scattering, while preserving details such as cracks and rust, providing a high-definition input image for the subsequent YOLOv8 algorithm.

[0028] Data quality grading and labeling: Based on the preprocessed signal quality, operating conditions, and image analysis results, the system automatically assigns a unique quality grade label to each data point. The grading standards are clear and quantifiable, directly guiding the data usage rules of subsequent modules. Grade A data (high-quality data): vibration signal noise amplitude < 5% of signal peak value, monitoring period is during normal bridge operation, and image analysis did not identify any external damage; this type of data serves as the core data source for cable force benchmark range calibration, hybrid drive model training, and model adaptive updates.

[0029] Level B data (usable data): The vibration signal noise amplitude is between 5% and 10% of the signal peak value. The monitoring period is during the normal operation of the bridge (without abnormal conditions), and the image has no obvious damage. This type of data needs to be filtered twice before it can be used for online inference verification of cable force and for condition assessment reference.

[0030] Level C data (data to be analyzed): vibration signal noise amplitude > 10% of signal peak value, or the above-mentioned abnormal working conditions are detected, or visual damage is identified by image recognition; this type of data is not used for cable force calculation and model training, but is specifically used for abnormal correlation analysis to provide a basis for fault root cause location.

[0031] Output: After processing in this step, a standardized data package with a unique quality grade label, precise timestamp, and installation point number is output. It contains clean vibration time history data and preprocessed grayscale images. The data format is fully compatible with the input requirements of subsequent modules, ensuring the continuity and consistency of the data flow.

[0032] The benchmark modeling module specifically includes: S201. Cable Force Reference Range Calibration: Based on high-quality data collected during the system initialization phase, and combined with cable design parameters and string vibration theory, determine the reference cable force range and safe rate of change threshold for each cable under healthy conditions.

[0033] Precise screening of health status data: From the long-term monitoring data output by the preprocessing module, Class A data that meets the following criteria is screened to ensure high reliability of the data source: Data quality requirements: vibration signal noise amplitude < 5% of signal peak value, image analysis shows no visible damage, and the monitoring period does not involve abnormal conditions such as strong winds ≥ 8, vehicle impact loads, or sudden temperature changes ≥ 10℃ / h.

[0034] Time span requirement: During the initialization phase, data should be monitored continuously for 30 days, with at least 2 hours of valid data segments selected each day, and the cumulative valid data duration should be ≥60 hours to avoid randomness under a single working condition.

[0035] Data integrity requirements: Each data segment must have ≥10,000 consecutive sampling points, based on a 100Hz sampling frequency, with a corresponding duration of ≥100 seconds, and no data loss or disconnection, ensuring the accuracy of frequency identification.

[0036] Initial cable force multi-order fusion calculation: for each segment of the filtered Class A denoised vibration signal The calculation accuracy is improved by using multi-order frequency fusion to calculate the initial cable force. Inherent frequency identification: Perform a Fast Fourier Transform (FFT) with a Hanning window to suppress spectral leakage, and set the number of sampling points to [value missing]. Frequency resolution ≤ 0.0061Hz, accurately identifying the first four orders ( Natural vibration frequency The selection of the first four frequencies is based on the fact that these frequencies are less affected by cable sag and boundary conditions, and the vibration amplitude is stable, which can effectively reduce the cable force calculation deviation caused by the identification error of a single frequency.

[0037] Single-order cable force calculation: for each frequency order Based on string vibration theory, and using the inherent design parameters of the cable as the only input parameters, the estimated value of the single-order cable force is calculated. The formula is:

[0038] in, Mass per unit length of cable The effective length of the cable. This represents the mode order.

[0039] Multi-order data fusion: for four groups A preliminary screening for outliers was performed, removing outliers that deviated from the results of other third-order calculations by more than 5% (if any). The arithmetic mean of the remaining valid data was then used as the initial cable force value for that data segment. If all four sets of data are normal, then the arithmetic mean of the four sets is taken directly to ensure... Stability.

[0040] Quantitative determination of the benchmark cable force range: Through statistical analysis and outlier removal, a stable and reliable benchmark cable force range is constructed. Sample set construction: Collect all eligible Level A data segments. To form an initial cable force sample set The sample size is ≥50 groups.

[0041] Statistical parameter calculation: Calculate the mean of the sample set. and standard deviation ,in It reflects the normal fluctuation range of cable force under healthy conditions.

[0042] Outlier cleaning: using The criterion is used as the sole outlier removal parameter, and outliers are removed if they meet the criteria. This criterion covers 99.73% of normal data and effectively eliminates random deviations under extreme operating conditions.

[0043] Benchmark range calibration: Recalculate the mean based on the cleaned sample set. and standard deviation Finally, the range of reference cable force was determined. The calculation formula is:

[0044] This range covers 95.45% of the health status data, preserving normal fluctuations while effectively distinguishing abnormal states, serving as the sole benchmark quantitative parameter for subsequent status assessment.

[0045] Determination of the safe rate of change of cable force: Based on material properties and engineering specifications, determine the safe threshold for the rate of change of cable force. The allowable stress variation rate of the cable is obtained through mechanical testing of the cable material, and then converted using the cable cross-sectional area, or by directly referencing the limit requirements for cable force variation in bridge engineering design specifications; its maximum allowable cable force variation rate is determined. This parameter represents the maximum allowable change in cable force per unit time. As the only safety parameter for the rate of change, it is used to judge sudden changes in cable force and avoid rapid changes in cable force that could lead to cable fatigue damage or structural imbalance in the bridge.

[0046] S202. Construction of Theoretical Cable Force Calculation Model (B-Model): Construct a parameterized theoretical cable force calculation model based on physical mechanisms as a physical reference benchmark for data-driven AI models, and improve the adaptability of the model in actual engineering environments through multi-dimensional correction coefficients.

[0047] Establishment of the basic physical mechanism model: Based on the theory of string vibration, a basic model for calculating cable force is constructed to ensure the physical rationality of the model. Core formula: Theoretical cable force base value The calculation formula is:

[0048] in, For real-time identification of the cable fundamental frequency, This represents the mode order.

[0049] Applicable conditions: The model is suitable for scenarios where the cable sag is ≤1 / 10 of the effective length. If the cable sag is large, it can be compensated by the sag correction coefficient already included in the design parameters, without the need to introduce an additional complex model.

[0050] Introduction of multi-physical factor correction coefficients: To compensate for various influencing factors in the actual environment, three updatable correction coefficients with clear physical meanings are introduced to form a complete B-model:

[0051] in, The corrected real-time theoretical tension is calculated using three correction coefficients to compensate for environmental temperature, material aging, and system errors, respectively. Each coefficient is a unique adaptation or calibration parameter.

[0052] Correction factor for the influence of ambient temperature : ,in The coefficient of thermal expansion of the cable material is _____. This represents the difference between the real-time temperature and the standard temperature (20℃). As the temperature rises, the cable elongates, and the cable tension decreases. Less than 1; as the temperature decreases, the cable shortens and the cable force increases. Greater than 1.

[0053] Cable stiffness attenuation correction factor : During the system initialization phase, the cables are in a healthy state with no stiffness reduction. The initial value is set to 1.0; as the only stiffness attenuation adaptation parameter, it is only updated as a whole (such as adjusted to 0.98, 0.96, etc.) when it is determined through long-term data analysis that the overall stiffness of the cable has irreversibly decreased. The update cycle is no less than 1 year to avoid frequent adjustments that may cause model instability. The value ranges from 0.9 to 1.0. If the calculated attenuation coefficient is less than 0.9, it indicates that the cable stiffness has been severely attenuated, triggering a dangerous state warning, rather than simply updating the coefficient.

[0054] Real-time model calibration factor : During the system initialization phase, the model has no systematic errors. The initial value is set to 1.0; as the sole real-time error calibration parameter, it is used daily based on the AI ​​model output from the previous 24 hours during routine operation. With B-model output relative deviation time series mean It is dynamically updated.

[0055] Update rules: , in , This represents the number of valid data points for the day; each update increment is limited to ±0.01 to avoid over-correction leading to model oscillation; if Then pause updates. This triggers the anomaly investigation process.

[0056] The hybrid prediction module specifically includes: S301. Cross-modal feature extraction and calibration: Automatically extract and fuse two types of complementary features from the preprocessed vibration signal to provide rich and well-calibrated input for the subsequent hybrid model.

[0057] Data-driven feature extraction: For denoised vibration signals marked as Grade A or Grade B Wavelet packet decomposition is performed. The number of decomposition levels is fixed at S=5 (as the sole decomposition depth parameter), and the db4 wavelet basis is used. This method uniformly subdivides the signal frequency band, enabling precise capture of the energy distribution in different frequency bands.

[0058] Calculate the wavelet packet coefficients for all 32 sub-bands at level 5. Extract two key features: Multi-scale energy entropy: Calculates the proportion of energy in each sub-band to the total energy and its Shannon entropy. This feature characterizes the complexity of the signal energy distribution in the frequency domain; Wavelet coefficient mean: Calculate the absolute mean of the wavelet coefficients for each sub-band, reflecting the signal strength of that frequency band.

[0059] The features extracted from all sub-bands (64 dimensions in total) are concatenated into a vector, and Z-score normalization is performed to form a standardized data-driven feature vector. .

[0060] Physical prior feature extraction: For the same signal Power spectral density (PSD) analysis was performed to automatically identify and record the first four natural frequencies. And the peak power spectral density corresponding to each frequency. These frequencies are the most direct physical basis for cable force calculation.

[0061] In the time domain, the signal is calculated directly. The four statistical measures are: mean (close to 0), standard deviation (reflecting vibration intensity), peak factor (the ratio of peak value to RMS value), and kurtosis (reflecting the sharpness of the signal distribution).

[0062] The above 4 frequencies, 4 spectral peaks, and 4 time-domain statistics (a total of 12 dimensions) are concatenated to form a physical prior feature vector. .

[0063] Feature fusion and adaptive calibration: Data-driven feature vectors With physical prior eigenvectors The features are concatenated to obtain the original mixed feature vector; To address the issue of two types of features having different scales and importance, a learnable cross-modal attention weight vector is introduced. During model training, its Each element in It will automatically learn to weight corresponding features, highlight important features, and suppress redundant or noisy features; The weighted feature vector is the final input feature after calibration. The criterion for judging the effectiveness of calibration is: on the training set, the intra-class variance of the weighted and calibrated feature vector among different samples should be reduced by ≥30% compared with that before calibration, to ensure that the features are more compact and discriminative.

[0064] S302. Hybrid-driven model (M-model) construction and training: Construct an innovative dual-channel neural network architecture and train it through a loss function that integrates multi-objective constraints, so that the model has both data learning ability and physical consistency.

[0065] Model architecture design: The data-driven channel is responsible for learning complex latent patterns from raw features. This channel consists of a one-dimensional convolutional neural network (1DCNN) connected in series with a Transformer encoder. The 1DCNN has a fixed number of convolutional kernels (C=64) to capture local correlations and primary features of the signal. The Transformer encoder has a fixed number of attention heads (H=8) to capture long-range dependencies and global context of the signal.

[0066] The physics-guided channel is a fully connected neural network whose inputs are primarily physical prior features. Instead of free fitting, this network is designed with fundamental constraints from string vibration theory embedded to ensure a strong correlation between its output and physical principles.

[0067] The outputs of the two channels are fused at a higher level, and then the final cable force prediction is output through a regression head. .

[0068] Improved multi-objective loss function and training: The loss function is key to the model's "physical augmentation" property, and it is defined as:

[0069] (Data Loss): Mean Squared Error (MSE), a measure of the model's predicted values. High-precision calibration of measured cable force values The differences between them. The data used for training consisted of N ≥ 1000 sets of high-quality Grade A data and their calibrated cable forces.

[0070] (Physical consistency loss): a metric The theoretical values ​​calculated by the B-model in the baseline modeling module under the same input conditions. The difference between them (e.g., MSE). This loss forces AI predictions to align with physical laws. Its weighting coefficients The weights are determined through cross-validation within the range [0.3, 0.7] and are the only physical constraint weights.

[0071] (Rate of Change Constraint Loss): Measures whether the instantaneous rate of change of cable force predicted by the model exceeds the safe rate of change determined by S201. This penalty applies to the portion exceeding the limit. This loss improves the smoothness and physical plausibility of the model's output. Its weight coefficients... The weight is determined within the interval [0.1, 0.3] and is the only weight constrained by the rate of change.

[0072] Training convergence condition: When the total loss is less than or equal to the value on the independent validation set... It no longer decreases after 10 consecutive training cycles, and its value is less than Training should be stopped when the time comes.

[0073] S303. Dual Redundancy Online Inference and Cross-Validation: The trained M-model and B-model are deployed as online services. Through the dual-model parallel inference and cross-validation mechanism, the reliability, real-time performance and self-calibration capability of the online recognition process are guaranteed.

[0074] Cloud-based redundant engine deployment: Two parallel inference threads are deployed on a cloud server to form a redundant engine; Thread A (Main Inference): Loads the trained M-model. Receives the real-time calibrated feature vectors from the S301 and directly outputs the AI-predicted cable force value. .

[0075] Thread B (Benchmark Verification): Runs the B-model built by S202. Receives the real-time calculated fundamental vibration frequency. and temperature Calculate and output the theoretical reference cable force value. .

[0076] Cross-validation and dynamic calibration process: The system calculates the relative deviation between the outputs of the two models in real time: .

[0077] Normal operating condition handling: In most cases, It should be maintained within a small range (e.g., <5%). The system will... The final identified cable tension value is output. Simultaneously, the system will also compile statistics from the past 24 hours. The moving average value is used as feedback to automatically fine-tune the real-time calibration factor in the B-model if the mean is stable and non-zero. This ensures that the long-term output trends of the two models remain consistent, enabling the models to self-calibrate.

[0078] Abnormal operating condition handling: When This indicates a discrepancy in the predictions of the two models, possibly due to abnormal interference on one side. The system immediately takes two measures: 1) Triggering an alert; 2) Activate a pre-trained, lighter backup inference model (such as a pure CNN model) and backtrack to call the 10 most recent historical A-level data sets and the current data for rapid comparative analysis to determine whether it is the beginning of a sudden change in environmental noise, sensor failure, or a real structural anomaly.

[0079] System performance assurance: from data input to final cable force value Output: End-to-end latency of the entire online inference process The time must be ≤500ms to meet the requirements of near real-time monitoring.

[0080] The assessment and early warning module specifically includes: S401. Real-time Cable Stress Assessment: Based on pre-established absolute benchmarks and dynamic rules, the real-time identified cable stress values ​​are automatically classified and graded. The cable stress values ​​identified at the current moment (…) ), and the reference range of the cable's healthy tension. and safety change rate threshold As the core input, the evaluation process employs deterministic, multi-level logical rules to ensure that the results are objective, consistent, and traceable.

[0081] The system executes the following hierarchical determination logic: Normal state determination: When the identified cable force value meets the requirements And its instantaneous rate of change At this point, the system determines that the cable is in a "normal state." This indicates that the cable's stress level and dynamic changes are within a healthy historical fluctuation range, which is a safe operating condition.

[0082] Warning Status Determination: The system determines that it has entered a "warning status" when one of the following two situations occurs: Amplitude deviation: Cable force value has exceeded the healthy baseline range ( or However, its rate of change remains relatively slow. This usually indicates that the average stress level of the cables may have shifted systematically, and attention should be paid to whether there is long-term prestress loss or temperature compensation deviation; Abnormal changes: Although the cable tension value is still within the reference range ( However, its instantaneous rate of change exceeded the safety threshold. This indicates that the cable may be under abnormal instantaneous load or that a redistribution of internal forces has occurred due to localized damage.

[0083] Danger State Determination: When the monitoring data meets the following more stringent conditions, the system will determine it to be in a "danger state," and the representation structure may face immediate risks: Severely exceeding limits: The cable tension value deviates significantly from the reference range, for example, below 80% of the lower limit. ) or more than 120% of the upper limit ( (Unique dangerous amplitude threshold).

[0084] Rapid deterioration: The rate of change of cable tension is drastic, reaching twice or more the safe rate of change. (Unique danger rate of change threshold).

[0085] S402. Multi-step trend prediction and anomaly root cause correlation analysis: By predicting future trends and combining multi-source information, actively explore the physical causes behind anomalies, thereby improving the predictability and interpretability of the assessment.

[0086] Short-term cable force trend prediction: This function is implemented using a time series analysis model based on a long short-term memory network. The model uses the historical cable force identification results of the most recent 30 consecutive time points as the input sequence, and learns the variation pattern of cable force under the influence of traffic flow, daily temperature cycle, etc. through its internal memory units.

[0087] The model has a fixed configuration of 128 hidden layer units and can predict and output the trajectory and possible fluctuation range of cable force changes within the next hour. This prediction result helps determine whether the current anomaly is a temporary disturbance or a trend that will continue to worsen, providing crucial time-based foresight for maintenance decisions.

[0088] Intelligent correlation between abnormal patterns and visual damage: When the status assessment result is "warning" or "dangerous," the system automatically triggers multi-source data fusion analysis to locate the root cause. On one hand, the system calls up high-definition images acquired simultaneously from the "sensor preprocessing module" and uses a visual recognition algorithm based on the YOLOv8 architecture to detect damage such as corrosion and cracks on the cable surface and anchorage area. The result is only adopted when the detection confidence level is higher than 70%. On the other hand, the system accesses the built-in "abnormality-damage" correlation knowledge base, which encapsulates domain expert experience and historical cases.

[0089] The system combines real-time cable force anomaly patterns, trend characteristics, and visual recognition results for joint reasoning, and finally outputs a list of potential root causes sorted by probability, thereby transforming abstract "abnormal" signals into specific suspected fault points that can guide on-site inspections.

[0090] S403. Structured Report Generation and Tiered Early Warning Execution: Integrate all assessment and analysis results to generate standardized decision support documents and execute differentiated early warning actions based on the status level, completing the final closed loop from data perception to maintenance action.

[0091] Structured Assessment Report Generation: The system automatically generates a comprehensive assessment report. The report, centered on data and charts, clearly includes real-time cable stress values, assessment status level, cable stress prediction curves for the next hour, identified suspected damage types and image evidence, and preliminary diagnostic opinions based on correlation analysis. The standardized report format ensures accurate, complete, and efficient information delivery.

[0092] Implementation of tiered early warning and response mechanism: The system executes a preset tiered response strategy based on the assessed status level, and all early warnings must be issued within 1 second after the status determination is completed.

[0093] In the event of an early warning, the system pushes detailed text alarm notifications to the monitoring platform and mobile terminals of maintenance management personnel, and archives the report to the pending list, prompting planned inspections and analyses. In case of dangerous situations, the system will immediately trigger the audible and visual alarm devices at the monitoring center while simultaneously sending text messages. The alarm volume will be no less than 85 decibels to ensure that on-site personnel are aware of the situation. Emergency response suggestions will be prominently included in the report. Every alert initiation, receipt, and subsequent manual response feedback is fully recorded, forming a traceable closed-loop archive, providing real-world learning samples for the system's continuous self-optimization.

[0094] The model self-optimization iteration module specifically includes: S501. Core Model Performance Monitoring and Degradation Early Warning: Establish a quantitative performance tracking system to continuously evaluate the online recognition accuracy of the core artificial intelligence model and provide timely warnings when measurable performance degradation occurs, triggering optimization processes.

[0095] The system has an automated performance monitoring program deployed in the background. This program silently collects all recent batches of monitoring data generated under normal operating conditions at fixed intervals.

[0096] For each batch of data, the system will identify the cable tension value online. This is compared to a high-confidence reference value. The acquisition of this reference value is lagging and is primarily achieved through two methods: One method is to obtain the absolute true value by periodically calibrating a high-precision hydraulic sensor on-site. Secondly, the system selects identification results from periods with extremely high sensor signal quality and relatively stable environments, and uses their statistical average as a temporary, highly reliable benchmark.

[0097] Based on the above comparison data, the system calculates the core evaluation indicator—the average relative error of the prediction result. The calculation formula is as follows:

[0098] in, This represents the total number of samples used for comparison within the current statistical period. This metric is a core quantitative parameter for measuring the accuracy of the model in practical applications.

[0099] The system dynamically maintains a historical baseline for model performance. When the calculated values ​​for two consecutive monitoring periods... If all values ​​are greater than or equal to 5%, and after automatic diagnosis has ruled out external interference such as temporary sensor malfunctions and extreme weather events, the system determines that the core AI model has experienced significant performance degradation. At this point, the system will automatically generate a high-priority model optimization task and immediately trigger the subsequent incremental learning and update process. This 5% threshold is the sole quantitative criterion and triggering condition for determining whether model performance needs optimization.

[0100] S502. Incremental Learning-Based Model and Parameter Update: Automatically activated upon receiving a performance degradation warning. Its goal is to progressively optimize the core prediction model and its associated physical parameters by introducing new data on the current state of the bridge, in order to restore and continuously improve the overall prediction capability of the system.

[0101] The optimization process begins with the construction of a high-quality, timely training dataset. The system automatically filters out all vibration-cable force data pairs that meet the "Grade A" quality standard from recent historical databases, requiring a cumulative total of at least 2000 pairs. These data represent the latest behavioral characteristics of the bridge structure at the current stage. Simultaneously, any new manually calibrated data during this period are added to the dataset as crucial "golden samples."

[0102] The core model update is performed through an incremental learning algorithm. Specifically, the system loads the current online service's hybrid-driven model as a pre-trained model and continues training it using the newly constructed dataset. During this process, the learning rate is set to a small, fixed value. .

[0103] At the same time, the system will adaptively and collaboratively adjust the physical reference parameters: The system analyzes the overall distribution trend of cable force values ​​in long-term historical data. If statistical tests reveal a baseline range... The center value has drifted slowly. After obtaining expert confirmation or triggering internal rules, the system will perform a one-time calibration update on the range to objectively reflect the changes in inherent properties caused by material relaxation, long-term creep and other effects.

[0104] The cable stiffness attenuation correction factor in the theoretical model It will also make minor adjustments based on the trend analysis results of data over many years.

[0105] Furthermore, the key balancing parameter in the model loss function is the physical constraint weight. Weights constrained by rate of change It will also be re-optimized on new datasets to ensure that the loss function can still guide the model to reach the optimal balance under the current data distribution.

[0106] S503. System-level Operating Parameters and Strategy Iteration: Based on actual operation and maintenance feedback and accumulated performance, parameters in data acquisition, processing logic, and early warning strategies are dynamically optimized to improve overall system performance. This mainly covers the strategic iteration of three types of operating parameters: Data acquisition strategy optimization: The system continuously analyzes the baseline level of environmental noise. When it detects that the environmental noise amplitude remains high due to permanent changes in the bridge's surrounding environment, thus affecting data quality, the system can automatically or, with authorization, adjust the sampling frequency of the vibration signal. Increased from the standard 100Hz to 120Hz (maximum sampling frequency threshold). This "oversampling" strategy provides more information redundancy for subsequent signal processing algorithms, which helps to extract purer structural response signals in complex noise backgrounds.

[0107] Early warning push strategy optimization: To avoid causing a "storm of early warnings" during brief periods of strong interference and thus inconveniencing management personnel, the system has introduced a minimum early warning push interval. parameter; When the same cable triggers the same level of warning multiple times in a short period of time, the system will only push a warning immediately upon the first trigger. Subsequent repeated warnings will be pushed out of this time interval, or when the status level is upgraded. This time interval can be flexibly configured according to actual management feedback.

[0108] Periodic review of assessment thresholds: Considering that the long-term aging of bridge materials and performance is a slow process, the system sets a fixed review cycle of 3 years. At the end of each cycle, the system will automatically suggest and initiate a comprehensive review of the core health assessment thresholds, mainly including the benchmark cable force range and the rate of change of safety. The review will take into account all monitoring data from the past three years, the latest bridge structure calculation and analysis reports, and updates to industry standards to ensure that the system's evaluation criteria always match the actual safety status of the bridge structure and keep pace with the times.

[0109] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic identification and evaluation system for cable tension of cable-stayed bridges based on dynamic modeling, characterized in that, include: Sensor preprocessing module: used to collect and process the vibration and image data of the cable, and assign a unique quality level label to it based on the noise level and working conditions; The benchmark modeling module is used to calibrate the healthy benchmark cable force range and safe rate of change threshold of the cable based on the string vibration theory and the selected high-quality label data, and to construct a theoretical cable force calculation model that includes temperature, stiffness and real-time calibration correction coefficients. Hybrid prediction module: used to extract and fuse data-driven features and physical prior features of vibration signals, identify cable force by using a hybrid driving model that fuses physical consistency loss and rate of change constraint loss, and perform online cross-validation and dynamic calibration using the theoretical cable force calculation model to output the final identified cable force; Assessment and early warning module: used to compare the final identified cable force with the healthy baseline cable force range and the safe rate of change threshold, realize three-level status assessment, and combine trend prediction and visual impairment identification to perform abnormal root cause correlation, generate reports and execute graded early warning; Self-optimizing iterative module: Based on the three-level state evaluation results during system operation and the corresponding historical final identified cable force data, monitor the model accuracy. When performance degrades, update the hybrid driving model through incremental learning and coordinately adjust the healthy benchmark cable force range, the safe rate of change threshold, and various correction coefficients in the theoretical cable force calculation model. At the same time, iteratively optimize the system operating parameters.

2. The automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling as described in claim 1, characterized in that, The allocation of quality tags includes: judging signal quality based on the ratio of noise amplitude to peak value of vibration signal; judging operating condition based on load, wind speed, and temperature change rate information during the monitoring period; judging appearance damage status based on the results of preliminary analysis of synchronously acquired images; and automatically assigning A, B, or C grade quality tags according to a preset unique combination threshold rule for signal quality, operating condition, and appearance damage.

3. The automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling as described in claim 1, characterized in that, The healthy baseline cable force range includes: multiple initial cable force samples calculated based on the high-quality tag data, the samples being obtained by identifying the natural frequency of the vibration signal and combining it with string vibration theory; the initial cable force sample set being processed using a preset statistical outlier removal criterion; and based on the statistical distribution characteristics of the cleaned sample set, determining an upper and lower bound of a numerical range characterizing the normal fluctuation of cable force under healthy conditions.

4. The automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling as described in claim 1, characterized in that, The safety change rate threshold includes: determined based on mechanical performance test data of cable materials or design specifications of bridge engineering; used to define the maximum allowable change in cable force per unit time; and used as a quantitative standard to assess whether the dynamic change in cable force is abnormal, applied to the judgment of change rate in the three-level state assessment and the constraint on the output change rate in the training of the hybrid driving model.

5. The automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling as described in claim 1, characterized in that, The theoretical cable force calculation model includes: The calculation core, based on string vibration theory, is used to calculate the basic value of cable force according to the real-time identified natural frequency of the cable; An ambient temperature correction coefficient is applied in series to the output of the calculation core. This coefficient is dynamically adjusted according to the real-time monitored temperature to compensate for the thermal expansion and contraction effect of the material. The long-term stiffness attenuation correction coefficient, which is applied in series to the output of the calculation core, serves as a static parameter reflecting the overall performance degradation of the cable. Its update is based on the trend analysis of long-term cable force and vibration data. A real-time online calibration factor is applied in series to the output of the calculation core. This factor is a dynamic parameter, and its adjustment is based on the statistical results of the system deviation generated by the online cross-validation in the hybrid prediction module, which is used to absorb the unmodeled constant error.

6. The automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling according to claim 1, characterized in that, The hybrid driving model includes: A data-driven channel is used to learn complex nonlinear feature mapping relationships from vibration signals; The physical guidance channel's network structure design is subject to prior constraints based on string vibration theory; The feature fusion and decision layer is used to fuse the outputs of the data-driven channel and the physical guidance channel. The training framework employs a composite loss function that integrates data error, physical consistency constraints, and output rate of change constraints to optimize the model.

7. The automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling as described in claim 1, characterized in that, The final identification of cable force includes: The initial cable force prediction value is generated by forward inference of the real-time vibration signal using the hybrid driving model; Parallel theoretical reference values ​​generated based on the theoretical cable force calculation model under the same input conditions at the same time; The unique output value used to characterize the current true stress state of the cable is determined by comparing and deciding on the two values ​​in real time through an online verification mechanism.

8. The automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling according to claim 1, characterized in that, The criteria for determining the three-level status assessment include: Normal state: The final identification cable force satisfies and instantaneous rate of change ; Warning status: Final identification of cable tension exceeding limits However, the rate of change is normal, or the cable tension is within the reference range but the rate of change is normal. ; Dangerous state: Final identification of cable force satisfies or or rate of change .

9. The automatic identification and evaluation system for cable-stayed bridge cable forces based on dynamic modeling according to claim 1, characterized in that, The self-optimizing iterative module includes: Performance monitoring is used to calculate the average relative error of the hybrid driving model online. When the average relative error continuously exceeds a preset degradation threshold, it is determined that the model performance has degraded and the optimization process is triggered. Model incremental update is used to collect newly added A-level quality data during system operation after the optimization process is triggered, and update the parameters of the hybrid driven model through an incremental learning algorithm with a fixed learning rate. The parameter coordination adjustment is used to recalibrate the healthy benchmark cable force range based on the trend analysis results of long-term accumulated cable force monitoring data and vibration data, and to calibrate the stiffness attenuation correction coefficient in the theoretical cable force calculation model.

Images