Crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm

Abstract

The present application relates to the technical field of crack calculation, in particular to a crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm, original crack data obtained by using a multi-source crack instrument array to carry out synchronous data acquisition at a set sampling frequency is cleaned and optimized, the fractal dimension of the cracks in the obtained standard crack data set is calculated by using a multi-scale box counting method, meanwhile, the multi-fractal spectrum parameters and dynamic characteristic sequences of the crack profile are extracted, the obtained fractal feature vector is converted into a feature sequence vector and input into a dynamic trend prediction model for calculation to obtain a predicted propagation path, propagation rate and critical point, the dynamic trend prediction model is updated and optimized based on a dynamic update trigger standard, and the problems that the prior art lacks real-time dynamic description of the crack propagation path, is difficult to adapt to the random changes of cracks in actual engineering and limits the reliability and early warning capability of the monitoring system are solved.

Description

A Crack Propagation Dynamic Trend Prediction Method Based on Crack Evolution and Fractal Algorithms Technical Field

[0001] This invention relates to the field of crack calculation technology, and in particular to a method for predicting the dynamic trend of crack propagation based on crack evolution and fractal algorithms. Background Technology

[0002] In the field of structural health monitoring, predicting the dynamic trend of crack propagation is a crucial step in ensuring structural safety. Current technologies often employ crack analyzers (such as acoustic emission sensors, strain gauges, or optical measurement devices) to monitor cracks on structural surfaces and obtain their geometric parameters (such as length, width, and depth). However, these methods largely rely on simple linear models or statistical methods for trend prediction, neglecting the complexity and irregularity of crack morphology. Cracks often exhibit fractal characteristics during propagation (such as self-similarity and scale invariance), but traditional prediction models cannot effectively capture these characteristics, leading to insufficient prediction accuracy, especially in long-term dynamic predictions. Furthermore, existing technologies lack real-time dynamic descriptions of crack propagation paths, making it difficult to adapt to the random changes in cracks in actual engineering projects, thus limiting the reliability and early warning capabilities of monitoring systems. Summary of the Invention

[0003] The purpose of this invention is to provide a dynamic trend prediction method for crack propagation based on crack evolution and fractal algorithms. This method solves the problem that existing technologies lack real-time dynamic description of crack propagation paths, making it difficult to adapt to the random changes of cracks in actual engineering and limiting the reliability and early warning capabilities of monitoring systems.

[0004] To achieve the above objectives, this invention provides a method for predicting the dynamic trend of crack propagation using a crack analyzer and fractal algorithm, comprising the following steps:

[0005] Synchronous data acquisition is performed using a multi-source crack analyzer array at a set sampling frequency. The obtained raw crack data is then cleaned and optimized to obtain a standard crack dataset. The multi-source crack analyzer array is composed of various types of crack analyzer sensors working together, including:

[0006] Digital image crack gauge: used for non-contact measurement to acquire two-dimensional or three-dimensional full-field displacement and strain data of the crack region;

[0007] Acoustic emission crack analyzer: used to capture transient elastic waves released inside the material during crack propagation, and to locate the initiation and active points of microcracks;

[0008] Distributed fiber Bragg grating sensors: These sensors are deployed at preset intervals on the surface of a structure or embedded inside the structure to achieve continuous, distributed measurement of crack width changes.

[0009] The fractal dimension of the cracks in the standard crack dataset is calculated using the multi-scale box counting method. At the same time, the multifractal spectrum parameters and dynamic feature sequences of the crack profile are extracted to form a fractal feature vector.

[0010] The fractal feature vector is converted into a feature sequence vector, and the feature sequence vector is input into the dynamic trend prediction model for calculation to obtain the predicted expansion path, expansion rate and critical point.

[0011] The process involves synchronous data acquisition using a multi-source crack analyzer array at a set sampling frequency, followed by cleaning and optimization of the raw crack data to obtain a standard crack dataset, including:

[0012] The target crack area is scanned for the first time using a digital image crack analyzer to obtain a reference digital image containing the crack. At the same time, a distributed fiber optic grating sensor records the center wavelength value of all measuring points under zero stress or initial state as a reference for the width change.

[0013] The distributed fiber optic grating sensor stimulates wavelength changes at all points along the deployment path using a set sampling frequency. Simultaneously, the acoustic emission crack instrument acquires acoustic emission signals within the structure and issues a trigger signal when the amplitude of the acoustic emission event signal exceeds a preset amplitude threshold.

[0014] Based on the trigger signal, the digital image crack instrument and the distributed fiber optic grating sensor are controlled to perform synchronous data acquisition;

[0015] The raw crack data is cleaned and optimized to obtain a standard crack dataset.

[0016] The raw crack data is cleaned and optimized to obtain a standard crack dataset, including:

[0017] All the raw crack data collected were spatiotemporally aligned, and abnormal data were identified and repaired.

[0018] The repaired original data is subjected to data denoising and signal enhancement, and feature extraction and standardization are performed to obtain a standard crack dataset.

[0019] The repaired original data undergoes data denoising and signal enhancement, followed by feature extraction and standardization to obtain a standard crack dataset, including:

[0020] From the noise-reduced crack contour sequence, primary geometric features are extracted at fixed time intervals or event-driven moments. The primary geometric features include contour length, bounding box area, and perimeter-area ratio.

[0021] Statistical features were extracted from the noise-reduced crack width variation sequence;

[0022] Activity features are extracted from acoustic emission activity data, including event count rate, cumulative energy, and cluster center location;

[0023] The primary geometric features, statistical features, and activity features, along with their corresponding timestamps and spatial location labels, are combined into a multimodal spatiotemporal feature vector.

[0024] The multimodal spatiotemporal feature vectors are standardized to obtain a standard crack dataset.

[0025] Specifically, the fractal dimension of the cracks in the standard crack dataset is calculated using the multi-scale box counting method. Simultaneously, the multifractal spectrum parameters and dynamic feature sequences of the crack profiles are extracted to form a fractal feature vector, including:

[0026] Based on the single-pixel width contour of the crack in the standard crack dataset, the fractal dimension is calculated using the multi-scale box counting method.

[0027] Based on the single-pixel width contour of the crack, partition functions of different orders are calculated, the corresponding quality index is calculated, and feature extraction is performed based on the obtained multifractal spectrum curve to obtain the multifractal spectrum parameters of the crack contour.

[0028] Based on the fractal dimension, the instantaneous rate of change, acceleration of change, and trend fitting are calculated to obtain the dynamic feature sequence;

[0029] A fractal feature vector is constructed based on the fractal dimension, the multifractal spectrum parameters, and the dynamic feature sequence.

[0030] The method further includes, after obtaining the fractal feature vector, the following:

[0031] The fractal feature vectors are cached in the data buffer pool according to their generation time and in chronological order.

[0032] Specifically, the fractal feature vector is converted into a feature sequence vector, and the feature sequence vector is input into a dynamic trend prediction model for calculation to obtain the predicted expansion path, expansion rate, and critical point, including:

[0033] Multiple fractal feature vectors are extracted from the data buffer pool according to the set compensation sliding time window to construct a feature sequence vector;

[0034] The feature sequence vector is input into the dynamic trend prediction model for calculation to obtain the predicted expansion path, expansion rate, and critical point.

[0035] The method further includes:

[0036] The prediction results are compared with the data in the standard crack dataset, and the dynamic trend prediction model is updated and optimized using an online incremental learning algorithm based on the dynamic update triggering standard.

[0037] The dynamic update triggering criteria include timed triggering, performance drift triggering, and critical time triggering;

[0038] The performance drift trigger is triggered when the prediction result error exceeds the preset tolerance multiple times in a row, thus triggering the dynamic trend prediction model to update and optimize.

[0039] The key time trigger is to trigger the dynamic trend prediction model to update and optimize when the data in the standard crack dataset reaches a set event threshold.

[0040] The method further includes:

[0041] The predicted propagation path, propagation rate, and critical point are visualized and output, and crack status reports and propagation trend graphs are generated.

[0042] This invention discloses a dynamic trend prediction method for crack propagation based on crack evolution and fractal algorithms. The method involves synchronously acquiring data using a multi-source crack analyzer array at a set sampling frequency, cleaning and optimizing the obtained raw crack data to obtain a standard crack dataset. The fractal dimension of the cracks in the standard crack dataset is calculated using multi-scale box counting, and simultaneously, multifractal spectrum parameters and dynamic feature sequences of the crack profile are extracted to form a fractal feature vector. This fractal feature vector is converted into a feature sequence vector and input into a dynamic trend prediction model for calculation, yielding the predicted propagation path, propagation rate, and critical point. The prediction results are compared with the data in the standard crack dataset. Based on a dynamic update trigger standard, an online incremental learning algorithm is used to update and optimize the dynamic trend prediction model. By introducing fractal algorithms, the fractal characteristics of crack morphology are fully explored, making the prediction model closer to actual crack propagation behavior. Combined with real-time data acquisition and model updates, long-term dynamic trend prediction of crack propagation is achieved, adapting to random crack changes and improving the reliability and early warning capability of the monitoring system. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0044] Figure 1 is a schematic diagram of the steps of a crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm according to the first embodiment of the present invention.

[0045] Figure 2 is a schematic diagram of the complete steps of a crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm provided by the present invention.

[0046] Figure 3 is a flowchart illustrating a crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm provided by the present invention. Detailed Implementation

[0047] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0048] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms "a," "say," and "this" used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms used herein refer to and / or include any or all possible combinations of one or more associated listed items.

[0049] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, words used herein may be interpreted as meaning "when..." or "when..." or "in response to determination".

[0050] Please refer to Figures 1-3. This invention provides a method for predicting the dynamic trend of crack propagation based on crack evolution and fractal algorithms, including the following steps:

[0051] S101. Synchronous data acquisition is performed using a multi-source crack instrument array at a set sampling frequency, and the obtained raw crack data is cleaned and optimized to obtain a standard crack dataset.

[0052] Specifically, a multi-source crack meter array is used as the core data acquisition unit. This array is composed of various types of crack meter sensors, mainly including:

[0053] Digital image crack gauge: used for non-contact measurement to acquire two-dimensional or three-dimensional full-field displacement and strain data of the crack area.

[0054] Acoustic emission crack analyzer: used to capture transient elastic waves released inside the material during crack propagation, and to accurately locate the initiation and activation points of microcracks.

[0055] Distributed fiber Bragg grating sensors: These sensors are deployed at preset intervals on the surface of a structure or embedded inside the structure to achieve continuous, distributed, and precise measurement of changes in crack width.

[0056] This array-based deployment enables multi-dimensional, cross-scale synchronous sensing of the same crack, from macroscopic geometry to microscopic activity signals, providing a data foundation for subsequent fractal analysis.

[0057] At the start of structural monitoring, a digital imaging crack analyzer is activated to perform an initial scan of the target crack area, acquiring a baseline digital image containing the crack. This image clearly records the initial outline of the crack, its branching network, and its relative position to the surrounding material.

[0058] Meanwhile, the distributed fiber optic grating sensor records the center wavelength value of all measuring points under zero stress or initial conditions, which serves as a reference for the width variation.

[0059] Distributed fiber Bragg grating (FBG) sensors continuously monitor wavelength changes at a fixed, high sampling frequency (e.g., several times per second) along their deployment path, transmitting the data in real-time to a cloud server for caching. This mode is used to capture slow, steady-state changes in crack width and deformation caused by factors such as ambient temperature. Acoustic emission (AE) crack analyzers continuously listen for acoustic emission signals within the structure. Once the amplitude of a detected AE event exceeds a preset threshold (indicating the formation of a new microcrack or significant expansion of an existing crack, typically set to 45-55 dB), the crack analyzer immediately sends a trigger signal. Upon receiving the trigger signal, the cloud server immediately sends a synchronous acquisition command to the digital image crack analyzer and the distributed FBG sensors. The digital image correlation crack analyzer quickly captures one (or a set of) current crack images, while the FBG crack analyzer records the precise wavelength values ​​at all measurement points at that moment. This mode ensures that all types of sensors are activated and record the most valuable data at the instant of critical dynamic changes in the crack.

[0060] Through the above process, the multi-source crack instrument array collects and outputs the following standardized data packets with unified timestamps:

[0061] Crack spatiotemporal geometry dataset: primarily provided by digital image crack gauges and distributed fiber Bragg grating sensors.

[0062] Crack profile sequence: a series of high-precision pixel coordinates or vector paths of crack edges arranged in chronological order. These profiles clearly show the changes in complex morphology such as crack length, direction, branching, and closure over time.

[0063] Crack width variation sequence: provided by distributed fiber Bragg grating sensors, it is a set of time series data corresponding to spatial location, which accurately records the width opening and closing changes at multiple key points along the crack path.

[0064] The full-field displacement and strain fields are calculated using digital image correlation techniques and reflect the deformation of the area surrounding the crack.

[0065] Crack activity acoustic dataset: provided by acoustic emission crack analyzer.

[0066] Acoustic emission event parameters include the occurrence time, spatial coordinates (located by multiple sensors), energy, amplitude, count, and rise time of each acoustic emission event. These parameters directly reflect the intensity and location of the microscopic activity of crack propagation.

[0067] By using acoustic emission signals as sentinels, the most detailed images and quantitative data can be acquired synchronously at the moments when the cracks are most active and most likely to undergo morphological changes. This mode not only improves the efficiency and targeting of data acquisition, but more importantly, it provides the subsequent fractal feature extraction module with spatiotemporally synchronized multimodal data that accurately reflects the key moments of the crack's dynamic evolution. This allows the calculation of features such as fractal dimension to be directly correlated with the microscopic activity events of the cracks, greatly improving the sensitivity and accuracy of dynamic trend prediction.

[0068] Because digital image crack meters, acoustic emission crack meters, and distributed fiber optic grating sensors differ in sampling frequency, data format, and physical reference frame during the acquisition process, they need to be unified under the same spatiotemporal framework.

[0069] First, time alignment is performed, assigning high-precision synchronization timestamps to all data streams. Using a unified timeline (such as the absolute clock of the data acquisition system) as a reference, interpolation or resampling is applied to all continuous data streams and event-driven data packets to ensure that every data point (such as image frames, fiber optic wavelength readings, and acoustic emission events) accurately corresponds to a unified timeline. Next, spatial registration is performed. For the mapping between image data and physical locations, a spatial transformation matrix is ​​calculated using reference points pre-set in the monitoring area, whose coordinates are known in both the digital image and physical space. This matrix accurately converts the coordinates of each pixel in the digital image into the actual surface coordinates of the structure. Simultaneously, the coordinates of each measurement point of the distributed fiber optic grating sensor and the coordinates located by acoustic emission events are also unified into this common world coordinate system. This achieves precise spatiotemporal fusion of cross-modal data. It not only solves the problem of time alignment between different sensor data but, more importantly, constructs a unified digital map of the structural surface. This allows the morphology of a crack in the image, the width reading on the fiber optic sensor, and the location of the acoustic emission event to be accurately superimposed on the same map, laying a solid foundation for subsequent joint analysis.

[0070] The raw data inevitably contains outliers caused by environmental interference (such as instantaneous strong light, vibration), instantaneous sensor failure, or transmission errors. A two-level strategy is used for data cleaning.

[0071] Gross error identification based on statistical models: For continuous width data sequences transmitted by fiber Bragg gratings, the system uses a dynamic threshold method (such as the 3σ criterion based on a sliding window) for preliminary screening, marking isolated data points that significantly deviate from the normal fluctuation range as suspected anomalies.

[0072] Collaborative verification and repair based on correlation analysis: Instead of judging whether a data point is abnormal in isolation, this approach utilizes the correlation between multiple data sources for cross-verification. For example, if a reading from a fiber optic sensor experiences a drastic jump and is flagged as potentially abnormal, the system will retrieve other data from the same time and spatial location for verification: It will examine the image from a digital image crack analyzer at that location to confirm whether the crack width has indeed changed abruptly. It will also check whether an acoustic emission crack analyzer has captured a high-energy acoustic emission event in the vicinity of that location.

[0073] If neither the image nor the acoustic emission data indicates abnormal activity, the fiber optic reading is considered an anomaly and corrected using linear interpolation of adjacent normal data or a predicted value based on historical data. If multiple data sources confirm the change, the data is retained and identified as a critical active event.

[0074] Even after removing gross errors, random noise still exists in the data. Adaptive filtering algorithms are applied to improve the signal-to-noise ratio for different data characteristics.

[0075] For the crack contour sequence (image data): a morphology-based fractal-preserving denoising method is employed. Traditional smoothing filters may blur the crack edges and impair their fractal characteristics. This method first removes isolated noise pixels through morphological opening and closing operations, and then utilizes the principle that fractal objects have self-similarity at different scales to perform multi-scale analysis on the preserved crack contours, enhancing their inherent fractal structure while suppressing random noise.

[0076] For crack width variation sequences (fiber optic data): an adaptive Kalman filter is employed. This algorithm can dynamically adjust the filtering parameters based on the real-time statistical characteristics of the data, effectively smoothing high-frequency random noise while rapidly and accurately tracking the actual, slow creep or rapid propagation of cracks.

[0077] For acoustic emission event parameters: This stage mainly involves event clustering and filtering. Multiple acoustic emission events that are close in time, spatially adjacent, and have similar characteristics are clustered into a more representative activity cluster to avoid over-interpreting microscopic and repetitive activities and to focus on meaningful crack propagation events.

[0078] From the purified crack contour sequence, a series of primary geometric features, such as contour length, bounding box area, and perimeter-area ratio, are extracted at fixed time intervals or event-driven moments.

[0079] Statistical features, such as mean, variance, and rate of change at specific points, are extracted from the denoised crack width variation sequence.

[0080] Activity features are extracted from the clustered acoustic emission activity data, including features such as event count rate, cumulative energy, and cluster center location.

[0081] All the above features, along with their unified timestamps and spatial location labels, are combined into a fixed-dimensional multimodal spatiotemporal feature vector. Finally, each dimension of this vector is Z-score standardized (i.e., subtracting the mean and dividing by the standard deviation) to unify all features to the same dimension, eliminating the bias caused by differences in numerical ranges on the model, resulting in a standard crack dataset. This condenses multi-dimensional information such as the macroscopic geometry, microscopic physical activity (acoustic emission), and precise local deformation (optical fiber) of the crack into a structured data object. This feature vector not only contains the original contour information used to calculate the fractal dimension but also incorporates auxiliary features that reflect crack dynamics, providing the subsequent dynamic trend prediction module with an extremely rich input far exceeding that of a single data source.

[0082] S102. The fractal dimension of the cracks in the standard crack dataset is calculated using the multi-scale box counting method. At the same time, the multifractal spectrum parameters and dynamic feature sequences of the crack profile are extracted to form a fractal feature vector.

[0083] Specifically, based on the resolution of the input image and the actual physical size of the crack, a denser and physically meaningful scale sequence is dynamically generated, covering the entire range from sub-pixel level fine structure to macroscopic contours. This is primarily based on the cleaned and spatially registered single-pixel width contour of the crack at a specific time point t. This contour is represented as a binary image (background 0, crack contour 1). The specific calculation process of the multi-scale box counting method is as follows:

[0084] 1. Mesh Covering: Cover the entire binary image containing cracks with a square mesh of side length ε.

[0085] 2. Box Counting: Count the number of boxes N(ε) that contain at least one crack pixel.

[0086] 3. Scale iteration: Continuously increase the value of ε (according to the adaptive scaling sequence) and repeat steps 1 and 2 to obtain a series of data pairs (ε, N(ε)).

[0087] 4. Double log-linear regression: Perform linear fitting on the data points on a double log-coordinate graph (log(ε) is the horizontal axis and log(N(ε)) is the vertical axis).

[0088] 5. Fractal Dimension Calculation: The absolute value of the slope of the fitted straight line is the fractal dimension D of the crack profile at time t. A simple line segment has a dimension of 1, a fully filled plane has a dimension of 2, and a complex fractal curve has a dimension between 1 and 2. The closer it is to 2, the more tortuous, complex, and detailed the profile is.

[0089] Output a time-varying fractal dimension sequence {D(t)} that quantifies the evolution of the overall morphological complexity of the crack.

[0090] A single fractal dimension D describes the average complexity of the crack as a whole, but may mask the heterogeneity of its different parts. For example, the complexity of the crack trunk and its minute branches may differ. Therefore, multifractal analysis is needed to reveal these local characteristics.

[0091] This is also a binary contour image of a crack. However, instead of treating the contour as a simple presence or absence, it is transformed into a probability measure distribution. An effective method is to convert the contour image into a distance field, where the value of each pixel represents its shortest distance to the crack contour, and then construct a measure representing the influence distribution based on this. The specific calculation process is as follows:

[0092] 1. Partition function calculation: Calculate the partition function of different orders q at multiple scales ε. This process is essentially using different magnifying glasses (controlled by q) to examine regions with different crack densities. When q >> 1, the magnifying glass focuses on regions with dense measure (such as clustered branches and tortuous contours); when q << 1, it focuses on regions with sparse measure (such as smooth main trunks).

[0093] 2. Solving for the quality index: For each q, the partition function and the scale ε have a power-law relationship in double logarithmic coordinates, and its slope is the quality index τ(q).

[0094] 3. Legendre Transform: By performing a Legendre transformation on τ(q), the multifractal spectrum f(α) is obtained. Here, α is the singularity index, describing the local irregularity (Hölder index), and f(α) represents the dimension of the fractal subset with that α value.

[0095] Three key features were extracted from the obtained multifractal spectrum f(α) curve:

[0096] Spectral width Δα = α_max - α_min: describes the range of singularities in different regions of the crack profile. The larger Δα is, the more heterogeneous the crack is, indicating the simultaneous existence of very smooth and extremely tortuous regions.

[0097] Spectral asymmetry B: describes whether the shape of the f(α) spectrum is biased towards high α values ​​(right-biased, representing the dominance of small branches) or low α values ​​(left-biased, representing the dominance of profile roughness).

[0098] Maximum dimension f_max: the value of f(α) at ​​the spectral vertex, which is usually close to the overall fractal dimension D.

[0099] For the crack profile at each time point t, output a feature set containing Δα(t), B(t), and f_max(t), which is a feature set containing the multifractal spectrum parameters of the crack profile.

[0100] The first-order difference of the fractal dimension at adjacent time points is calculated, i.e., ΔD / Δt. This instantaneous rate of change reflects the instantaneous speed of the change in crack complexity. A positive value indicates that the crack is branching and becoming more complex; a negative value indicates that branch closure or contour smoothing may have occurred. The difference of the rate of change of the fractal dimension is calculated, i.e., Δ²D / Δt². This acceleration reflects whether the change in complexity is accelerating or decelerating, which is crucial for predicting the critical instability point. The {D(t)} sequence is linearly or polynomially fitted within a certain time window, and its slope is used as a feature of the long-term evolution trend. Finally, a dynamic feature sequence containing the dynamic rate of change and acceleration of the fractal dimension is obtained.

[0101] The fractal dimension D(t), multifractal spectral parameters (spectral width Δα(t), asymmetry B(t), maximum dimension f_max(t)), dynamic evolution parameters (rate of change ΔD / Δt, acceleration Δ²D / Δt²), and auxiliary geometric and physical parameters are aggregated. These auxiliary geometric and physical parameters are validated physical quantities directly related to the cracks introduced during data preprocessing, such as the maximum crack width from fiber optic data and the cumulative energy from acoustic emission data. These parameters provide physical scale and activity evidence for the fractal features. At each analysis time point t, all the above features are arranged in a predefined, fixed order to form a one-dimensional array. This array is then subjected to a final Z-score normalization to ensure that all features are numerically of the same order of magnitude, preventing certain large-valued features from unreasonably dominating subsequent machine learning models. The final output is a unified, normalized, high-dimensional fractal feature vector. This vector is generated at each sampling time, forming a feature trajectory that evolves over time.

[0102] S103. Convert the fractal feature vector into a feature sequence vector, and input the feature sequence vector into the dynamic trend prediction model for calculation to obtain the predicted expansion path, expansion rate and critical point.

[0103] Specifically, the fractal feature vectors are cached in a data buffer pool according to their generation time and in chronological order, using a sliding time window of fixed length. This window extracts the feature vectors from the most recent N time steps from the data buffer pool, arranges them in chronological order, and constructs a multi-dimensional time series sample. For example, a sample may contain one fractal feature vector generated every hour within the past 24 hours (N=24). This sample encapsulates complete information about the recent dynamic evolution of the cracks, and the final output is a regular, equally timed feature sequence sample, which will serve as the direct input to the dynamic trend prediction model.

[0104] The dynamic trend prediction model employs an attention-enhanced spatiotemporal recurrent neural network. Its architecture consists of: a recurrent neural network layer, typically based on long short-term memory networks or gated recurrent units. This layer effectively learns long-term dependencies in time series, understanding dynamic patterns such as the accelerated expansion often accompanying a sustained increase in fractal dimension; and an attention mechanism layer that dynamically and automatically assigns different weights to feature vectors at different times in the input sequence. For example, the model can learn to focus more on critical moments when the acceleration of fractal dimension change suddenly increases, or when there are sharp changes in the cumulative acoustic emission energy, rather than treating all historical data equally. This allows the model to focus on the most critical historical segments for predicting the future. The model's output layer is designed with a multi-task learning architecture, simultaneously predicting multiple key future trend parameters.

[0105] Future Expansion Path Probability Map: The model outputs a probability distribution map overlaid on the structural coordinate graph, predicting the areas where the crack is most likely to extend forward and generate new branches within a future period (e.g., the next 72 hours). The higher the probability, the greater the risk of cracking in that area.

[0106] Key trend indicator prediction: The model outputs specific numerical predictions for a future time series, including:

[0107] Fractal dimension prediction values ​​{D(t+1), D(t+2), ...};

[0108] Predicted maximum crack width;

[0109] Predicted crack propagation rate;

[0110] Critical risk probability: The model outputs a scalar between 0 and 1, representing the combined probability of the crack rapidly expanding instably (i.e., the critical state) within a specific timeframe in the future.

[0111] The system integrates the model's output of the future expansion path probability map, key trend indicator prediction sequence, and critical risk probability. By presetting a series of safety thresholds (such as fractal dimension growth rate threshold, crack width threshold, and critical risk probability threshold), the predicted values ​​are compared with these thresholds in real time. Once any predicted value exceeds the threshold, the system will automatically generate different levels of early warning information (such as attention, warning, and severity).

[0112] The final result is a structured report predicting the dynamic trend of cracks. This report is presented in the form of data and charts, and includes:

[0113] Current Status Summary: Crack Overview Based on the Latest Data.

[0114] Future expansion heatmap: Descriptive text describing the expansion areas with the highest probability.

[0115] Key parameter prediction curves: describe the changing trends of fractal dimension, width, and rate over a future period.

[0116] Comprehensive risk level and early warning information: clear warning tips and recommendations.

[0117] To facilitate timely viewing, all forecast results and reports are also visualized.

[0118] To ensure that the prediction model can adapt to the possible new characteristics of crack evolution over a long period of time, an innovative and automated feedback learning loop was integrated.

[0119] S104. Compare the prediction results with the data in the standard crack dataset, and update and optimize the dynamic trend prediction model using an online incremental learning algorithm based on the dynamic update triggering standard.

[0120] Specifically, the model's prediction results are compared with the actual data collected subsequently and obtained after preprocessing and feature extraction, and a series of performance indicators are calculated, such as the mean squared error of prediction, the accuracy and recall of critical state prediction, etc.

[0121] Dynamic update trigger mechanism:

[0122] Scheduled Trigger: Regardless of performance, the system will perform a routine update of the model parameters at a low frequency (e.g., monthly).

[0123] Performance drift trigger: When the prediction error is detected to continuously exceed the preset tolerance, it indicates that the current state of the model can no longer accurately describe the dynamics of the crack, that is, model performance drift has occurred. At this time, the update process is immediately triggered.

[0124] Critical event triggering: When the system actually detects a significant crack propagation event (such as a fractal dimension jump or a surge in acoustic emission energy), an update will be forcibly triggered regardless of whether the model successfully predicted the event, so that the model can learn from this important event.

[0125] Once an update is triggered, the system will not retrain the model using all historical data (which would be too computationally intensive). Instead, it will employ an online incremental learning algorithm. This algorithm uses only feature sequence sample-to-real value data pairs from the most recent period (e.g., the past three months) to fine-tune the existing model parameters, enabling it to quickly adapt to new dynamic patterns.

[0126] To prevent the model from forgetting long-term patterns learned in the past during incremental learning (i.e., catastrophic forgetting), the system introduces a knowledge distillation technique. During updates, the system allows the old model to make predictions on the new batch of data, and this prediction is used as part of the soft objective, along with the true hard objective of the new data, to train the new model. This is equivalent to allowing the new model to learn from new data while also receiving guidance from the old model, thus achieving a balance between adapting to new changes and retaining old knowledge.

[0127] Workflow: The constructed feature sequence samples are input into the AST-RNN model. Inside the model, temporal features are extracted sequentially through RNN layers, important information is focused through attention layers, and finally, the three comprehensive prediction results are generated through a multi-task output layer.

[0128] Beneficial effects:

[0129] 1. Improve prediction accuracy: By introducing fractal algorithms, the fractal characteristics of crack morphology (such as fractal dimension and complexity) are fully explored, making the prediction model closer to the actual crack propagation behavior. For Paris' Law and its derivative models or linear / nonlinear regression models based on simple geometric parameters (such as crack length and width), this invention significantly improves prediction accuracy.

[0130] 2. Enhanced dynamic adaptability: By combining real-time data acquisition and model updates, long-term dynamic trend prediction of crack propagation is realized, which can adapt to random changes in cracks and provide timely early warning for the safety of engineering structures.

[0131] 3. Optimize data utilization: By closely integrating crack instrument data with fractal algorithms, multi-parameter integrated analysis is achieved, which improves data utilization efficiency and reduces prediction uncertainty.

[0132] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0133] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for predicting the dynamic trend of crack propagation based on crack evolution and fractal algorithms, characterized in that, Includes the following steps: Synchronous data acquisition is performed using a multi-source crack analyzer array at a set sampling frequency. The obtained raw crack data is then cleaned and optimized to obtain a standard crack dataset. The multi-source crack analyzer array is composed of various types of crack analyzer sensors, including: a digital image crack analyzer for non-contact measurement, acquiring two-dimensional or three-dimensional full-field displacement and strain data of the crack region; an acoustic emission crack analyzer to capture transient elastic waves released within the material during crack propagation, locating the initiation and active points of micro-cracks; and distributed fiber optic grating sensors, arranged at preset intervals on the structural surface or embedded within the structure, enabling continuous, distributed measurement of crack width changes. The fractal dimension of the cracks in the standard crack dataset is calculated using a multi-scale box counting method. Multifractal spectrum parameters and dynamic feature sequences of the crack contour are extracted to form fractal feature vectors. These fractal feature vectors are then converted into feature sequence vectors and input into a dynamic trend prediction model for calculation to obtain the predicted expansion path, expansion rate, and critical point. The dynamic trend prediction model employs a spatiotemporal recurrent neural network enhanced with an attention mechanism. The model architecture of the dynamic trend prediction model is as follows: a recurrent neural network layer, using a long short-term memory network or gated recurrent units as a foundation, learns long-term dependencies in the time series and understands the dynamic pattern of accelerated expansion accompanying the continuous increase of fractal dimension; and an attention mechanism layer, which dynamically and automatically assigns different weights to the feature vectors at different times in the input sequence.

2. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 1, characterized in that, Synchronous data acquisition is performed using a multi-source crack analyzer array at a set sampling frequency. The obtained raw crack data is then cleaned and optimized to obtain a standard crack dataset. This includes: using a digital image crack analyzer to perform an initial scan of the target crack area to obtain a reference digital image containing the crack; simultaneously, a distributed fiber optic grating sensor records the center wavelength values ​​of all measuring points under zero stress or initial conditions as a reference for width variation; using the distributed fiber optic grating sensor at a set sampling frequency to record the wavelength variation of all points along the deployment path; simultaneously, an acoustic emission crack analyzer acquires acoustic emission signals within the structure, and when the amplitude of the acoustic emission signal exceeds a preset amplitude threshold, a trigger signal is emitted; based on the trigger signal, the digital image crack analyzer and the distributed fiber optic grating sensor are controlled to perform synchronous data acquisition; and the obtained raw crack data is then cleaned and optimized to obtain the standard crack dataset.

3. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 2, characterized in that, The obtained raw crack data is cleaned and optimized to obtain a standard crack dataset, including: spatiotemporal alignment of all the collected raw crack data, and identification and repair of abnormal data; data denoising and signal enhancement of the repaired raw crack data, and feature extraction and standardization processing to obtain a standard crack dataset.

4. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 3, characterized in that, The repaired original crack data undergoes data denoising and signal enhancement, followed by feature extraction and standardization to obtain a standard crack dataset. This includes: extracting primary geometric features from the denoised crack contour sequence at fixed time intervals or event-driven moments, the primary geometric features including contour length, bounding box area, and perimeter-to-area ratio; extracting statistical features from the denoised crack width variation sequence; extracting activity features from acoustic emission activity data, the activity features including event count rate, cumulative energy, and cluster center location; combining the primary geometric features, statistical features, and activity features, along with corresponding timestamps and spatial location labels, into a multimodal spatiotemporal feature vector; and standardizing the multimodal spatiotemporal feature vector to obtain the standard crack dataset.

5. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 1, characterized in that, The fractal dimension of the cracks in the standard crack dataset is calculated using multi-scale box counting. Simultaneously, multifractal spectrum parameters and dynamic feature sequences of the crack contours are extracted to form a fractal feature vector. This process includes: calculating the fractal dimension based on the single-pixel width contour line of the cracks in the standard crack dataset using multi-scale box counting; calculating partition functions of different orders based on the single-pixel width contour line, calculating the corresponding quality index, and extracting features based on the obtained multifractal spectrum curves to obtain the multifractal spectrum parameters of the crack contours; calculating the instantaneous rate of change, acceleration, and trend fitting based on the fractal dimension to obtain the dynamic feature sequence; and constructing a fractal feature vector based on the fractal dimension, the multifractal spectrum parameters, and the dynamic feature sequence.

6. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 1, characterized in that, After obtaining the fractal feature vector, the method further includes: caching the fractal feature vector in a data buffer pool according to the generation time and in chronological order.

7. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 6, characterized in that, The process involves converting the fractal feature vectors into feature sequence vectors and inputting these feature sequence vectors into a dynamic trend prediction model for calculation to obtain the predicted expansion path, expansion rate, and critical point. This includes: extracting multiple fractal feature vectors from the data buffer pool according to a set compensated sliding time window to construct a feature sequence vector; and inputting the feature sequence vectors into the dynamic trend prediction model for calculation to obtain the predicted expansion path, expansion rate, and critical point.

8. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 1, characterized in that, The method further includes: comparing the prediction results with the data in the standard crack dataset, and updating and optimizing the dynamic trend prediction model using an online incremental learning algorithm based on the dynamic update triggering standard.

9. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 8, characterized in that, The dynamic update triggering criteria include timed triggering, performance drift triggering, and critical time triggering; the performance drift triggering is triggered when the prediction result error exceeds the preset tolerance multiple times consecutively, thus triggering the dynamic trend prediction model to update and optimize; the critical time triggering is triggered when the data in the standard crack dataset reaches a set event threshold, thus triggering the dynamic trend prediction model to update and optimize.

10. The crack propagation dynamic trend prediction method based on crack evolution and fractal algorithm as described in claim 1, characterized in that, The method further includes: visualizing the predicted propagation path, propagation rate, and critical point, and generating a crack status report and propagation trend graph.

Images