Dam crack propagation tracking method based on acoustic emission signal analysis

Abstract

The invention provides a dam crack extension tracking method based on acoustic emission signal analysis, which comprises the following steps: S1, carrying out multi-dimensional feature enhancement preprocessing and extraction on acquired acoustic emission signals, and carrying out event identification and segmentation based on a pre-trained long-short term memory network to generate high-dimensional signal feature super vectors; s2, performing fracture mode identification and classification based on a deep learning network of fracture mechanics physical constraints, and outputting a fracture mode probability distribution vector of the acoustic emission event; s3, fusing the fracture mode information and an anisotropic velocity field model related to a dam stress field, and performing sound source high-precision positioning through an iterative inversion algorithm; s4, performing cluster recognition on the positioned sound source event based on a four-dimensional density clustering algorithm of space-time correlation and physical similarity, and reconstructing a three-dimensional crack frontal surface by adopting a moving least square method; and S5, according to the reconstructed crack frontal surface trajectory and the corrected acoustic emission event energy, inverting crack propagation kinetic parameters and quantitatively evaluating a damage state.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy project safety monitoring technology, specifically, it relates to a method for tracking the propagation of dam cracks based on acoustic emission signal analysis. Background Technology

[0002] In the long-term service safety assurance system of large-scale water conservancy engineering structures, especially concrete dams, real-time and accurate monitoring of their structural integrity constitutes the core technical support for risk prevention and disaster mitigation. As heavy structures bearing complex static and dynamic loads (such as elevation water pressure, temperature gradient stress, material creep and drying shrinkage, seismic disturbances, etc.), dams inevitably experience microscopic damage to their internal concrete materials due to multi-field coupling effects, which gradually evolve into macroscopic cracks. The initiation, propagation, and penetration of these cracks are key physical processes leading to a decrease in the dam's structural bearing capacity, loss of seepage stability, and even catastrophic failure. Therefore, high-precision dynamic tracking of crack propagation behavior, accurately grasping its spatial location, geometric morphology, propagation rate, and energy evolution laws, is of irreplaceable significance for achieving scientific assessment and accurate early warning of dam safety status.

[0003] Based on this, those skilled in the art have explored and implemented various monitoring technologies. In the early stages, monitoring was mostly based on contact measurement methods, such as borehole displacement sensors or surface-mounted strain gauges. These technologies directly measure the opening and closing degree of cracks, shear displacement, or strain field changes in adjacent areas, making the measurement results intuitive and physically meaningful, and providing quantitative data on crack deformation in specific key areas. Subsequently, with the development of computer vision and image processing technologies, non-contact optical detection methods based on digital image correlation or deep learning models have gradually been applied. These methods perform periodic high-resolution imaging of the dam surface, using algorithms to automatically identify and quantify the length, width, and distribution pattern of surface cracks, significantly improving the detection range and efficiency, especially demonstrating unique advantages in the comprehensive survey of wide-area surface cracks.

[0004] However, as dam safety assessments evolve from static damage identification to dynamic process early warning, the aforementioned technical approaches have become increasingly limited. Contact sensors have extremely limited spatial coverage, only able to monitor known cracks at preset locations, completely unable to detect cracks in unmonitored areas or internal cracks. Optical visual monitoring methods are limited to monitoring only the structural surface. More importantly, both contact measurement and optical detection are essentially discrete-time observations, unable to capture cracks during the interval between observations, especially high-speed, transient propagation events caused by sudden load changes, thus making it difficult to construct a continuous and complete temporal evolution trajectory.

[0005] To overcome the aforementioned shortcomings, acoustic emission monitoring technology passively monitors the energy signals released during crack propagation by deploying acoustic emission sensor arrays on the structure. Theoretically, by analyzing the arrival time, waveform characteristics, and energy of the received acoustic emission signals, the occurrence time, three-dimensional spatial location, and relative intensity of crack events can be deduced. Existing acoustic emission analysis methods have achieved dynamic sensing and localization of internal damage events. However, current technologies treat acoustic emission signals as markers of crack event occurrences, ignoring the physical information contained within the signal waveform. The fixed threshold triggering and simple localization algorithms commonly used in existing methods do not establish a quantitative physical mapping relationship between these signal characteristic parameters and the dynamic state of crack propagation. This results in the system's inability to effectively distinguish between real crack propagation signals and environmental and operational noise (such as water flow impact, gate opening and closing, and unit vibration), leading to low signal-to-noise ratios, high false alarm and false negative rates; the final result is merely spatially discrete sound source points. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for tracking dam crack propagation based on acoustic emission signal analysis. To address the technical problems mentioned in the background section, existing acoustic emission monitoring methods only treat signals as event triggers, ignoring the physical information contained within the waveform. This results in the inability to distinguish noise from real crack signals, the inability to form continuous crack propagation trajectories, and the inability to invert the crack propagation dynamics. This invention provides a method for tracking dam crack propagation based on acoustic emission signal analysis. This method constructs a quantitative mapping model from the multidimensional characteristics of acoustic emission signals to the physical mechanism of crack propagation, realizing the reconstruction of the continuous crack front evolution process and the inversion of dynamic parameters from discrete sound source location.

[0007] To achieve the above-mentioned objectives, the technical solution adopted in this invention is a method for tracking the propagation of cracks in dams based on acoustic emission signal analysis. The method is executed through an acoustic emission monitoring system deployed inside and on the surface of the dam structure. The system includes a sensor array composed of multiple acoustic emission sensors, a multi-channel synchronous data acquisition unit, and a central data processing server. The method includes the following steps: The first step is adaptive preprocessing and feature extraction of acoustic emission signals with multidimensional feature enhancement.

[0008] The raw acoustic emission waveform data acquired by the sensor array is converted from analog to digital by the multi-channel synchronous data acquisition unit and then transmitted to the central data processing server in the form of a data stream. The central data processing server first performs adaptive bandpass filtering. The center frequency and bandwidth of the filter are not fixed values, but are dynamically determined based on statistical analysis of historical noise signals in the dam's operating condition database.

[0009] Specifically, the server pre-stores a noise spectrum feature library, which records the power spectral density function of the background noise signal under different operating conditions (such as flood discharge, full-load operation of generator units, and sudden seasonal temperature changes). Upon receiving a real-time acoustic emission data stream, the system first identifies the current macroscopic operating condition of the dam and retrieves the corresponding noise power spectral density function from the feature library. It then designs a Wiener filter aimed at suppressing this specific noise spectrum to filter the original signal and maximize the signal-to-noise ratio. The filtered signal is then fed into an event recognition and segmentation module. This module employs a sequence anomaly detection algorithm based on a Long Short-Term Memory (LSTM) network, replacing the traditional fixed threshold method. The LSTM network, trained with a large number of real crack signal samples and noise samples, learns the inherent patterns of crack signal evolution in time. This allows it to accurately identify the start and end points of valid acoustic emission events in weak signals below the background noise level and segment them from the continuous data stream to form independent event waveform data segments.

[0010] Furthermore, for each successfully segmented event waveform data segment, the central data processing server performs a multi-domain joint feature extraction to construct a high-dimensional signal feature supervector (SFHV). The SFHV not only includes conventional waveform parameters but also deeply mines features reflecting the physical mechanism of the crack source.

[0011] In the time domain, the feature extraction process calculates and quantifies the following parameters: signal rise time, defined as the time interval during which the signal amplitude first rises from 3 times the root mean square value of the background noise to 90% of the peak amplitude; signal duration, defined as the total time from when the signal amplitude first exceeds the aforementioned threshold to when it last falls back below that threshold; signal ringing count, defined as the number of times the signal amplitude crosses the threshold of 3 times the root mean square value of the background noise within the aforementioned time period; absolute energy of the signal, defined as the integral of the squared voltage values ​​of all sampling points within the event waveform data segment over time, in joules; peak amplitude of the signal, in microvolts; and kurtosis and skewness of the signal waveform, used to characterize the statistical features of the signal's impulsiveness and asymmetry.

[0012] In the frequency domain, the feature extraction process performs a Fast Fourier Transform (FFT) on the event waveform data segment to obtain its spectrum, and calculates the following parameters from it: peak frequency, i.e., the frequency point with the highest energy density in the spectrum; frequency centroid, i.e., the first moment of spectral energy about frequency; and the energy ratio of the low-frequency band to the high-frequency band of the spectrum, where the low-frequency band is defined as the lower half of the effective operating frequency range of the sensor (e.g., for a 50-400kHz sensor, the low-frequency band is 50-225kHz), and the high-frequency band is the upper half (225-400kHz). This ratio (LF / HF Ratio) is used to initially characterize the fracture scale. In the time-frequency joint domain, the feature extraction process employs a Continuous Wavelet Transform (CWT) and uses the Morlet mother wavelet to decompose the event waveform, generating a three-dimensional energy spectrum of time, frequency, and energy. From this spectrum, the instantaneous entropy of energy on the time axis and the spectral entropy on the frequency axis are extracted to quantify the complexity and disorder of the signal energy release process. All the extracted time-domain, frequency-domain, and time-frequency-domain parameters are combined into a unique, standardized signal feature supervector, which serves as the sole input for subsequent analysis steps.

[0013] The second step involves the identification and classification of fracture patterns using a deep learning network based on physical constraints. This invention proposes a deep neural network model based on physical constraints of fracture mechanics to accurately classify the fracture patterns corresponding to each acoustic emission event. The model is a hybrid neural network architecture, comprising a one-dimensional convolutional neural network (1D-CNN) branch for processing the original waveform sequence and a fully connected network (FCN) branch for processing the aforementioned signal feature supervectors. The input to the 1D-CNN branch is a standardized segment of event waveform data. Its structure includes three convolutional layers with kernel sizes of 7x1, 5x1, and 3x1, and alternates between max-pooling layers to automatically learn subtle local differences in the waveform morphology. These differences are directly related to the moment tensor radiation pattern of the fracture source.

[0014] The input to the FCN branch is the signal feature supervector, which contains three hidden layers and aims to learn the macroscopic distribution patterns of different fracture modes in a multidimensional feature space. The output features of the two branches are concatenated at the back end of the network and fed into a Softmax classification layer. The output of the classification layer is not a single label, but a probability distribution vector representing the event belonging to three basic fracture modes—opening, sliding, and tearing.

[0015] Crucially, the training process of the hybrid neural network is not purely data-driven, but incorporates physical constraints from fracture mechanics. Its training dataset consists of two parts: one part comprises a large number of synthetic acoustic emission waveforms with clearly defined fracture mode labels generated through finite element method (FEM) simulations, where the simulation process strictly adheres to fracture criteria for concrete materials (such as the Hillerborg cohesion model); the other part comprises acoustic emission signals generated under controlled conditions using uniaxial tension, three-point bending, and shear experiments on standard concrete specimens, simultaneously acquiring signals that reflect specific fracture modes.

[0016] By combining physical simulation with experimental calibration, the resulting classification model possesses strong generalization ability and clear physical interpretability, enabling it to accurately attribute each acoustic emission event to its underlying physical fracture mechanism.

[0017] The third step involves high-precision sound source localization using an anisotropic velocity field that integrates fracture mode information. To overcome the problem of insufficient localization accuracy caused by the reliance on the homogeneous medium assumption in traditional localization algorithms, this invention proposes an iterative inversion localization method that combines fracture mode information with an anisotropic wave velocity model related to the dam's stress field. First, a full-size three-dimensional finite element model of the dam is constructed within the central data processing server. Based on the dam's design drawings, material parameters, and real-time monitored boundary conditions such as water level and temperature, this model calculates the three-dimensional stress tensor distribution at various points within the dam. According to the acoustoelastic effect theory, the propagation speed of elastic waves in a compressed medium is higher than that in a tensile medium, and the speed is related to the principal stress direction. Accordingly, the server establishes a stress-dependent anisotropic P-wave velocity field model, where the velocity V(x,y,z) at any point in space is a tensor related to the stress state σ(x,y,z) at that point. When it is necessary to localize an acoustic emission event, the system first obtains the fracture mode probability distribution of the event from the second step. This fracture mode information is used in two ways: firstly, it determines the spatial orientation pattern of the sound source's radiated energy. For example, the energy radiated by an open fracture is weakest in the direction parallel to the fracture surface and strongest in the vertical direction. Secondly, different fracture modes correspond to different initial P-wave and S-wave energy ratios. This information is used to optimize the weights of the arrival time picking algorithm. The localization process is a nonlinear optimization iterative process. The algorithm first randomly generates a set of candidate source points in the possible region. For each candidate source point, the theoretical shortest propagation time from that point to each sensor in the sensor array is calculated by solving the equation using the anisotropic velocity field model. At the same time, considering the sound source radiation pattern, the theoretical received energy of different sensors is weighted. Then, an objective function is constructed, which is the weighted sum of squared residuals between the actual arrival time of the P-wave picked up by all sensors and the theoretical propagation time. The weights are positively correlated with the signal-to-noise ratio of the sensor received signal and the theoretical received energy based on the radiation pattern. Finally, simulated annealing or particle swarm optimization algorithms are used to search for the point in three-dimensional space that minimizes the objective function. This point is the finally determined sound source location.

[0018] By integrating fracture mode and stress field information, the positioning accuracy of this invention is significantly improved compared to traditional methods. In particular, for crack sources in stress concentration areas, the positioning error can be effectively controlled at the centimeter level.

[0019] The fourth step involves identifying sound source event clusters and reconstructing crack fronts based on spatiotemporal correlation and physical similarity. To track the continuous crack propagation process from discrete sound source points, this invention proposes a four-dimensional (3D spatial + temporal) density clustering algorithm for identifying sets of sound source events with causal correlation. The algorithm is an improved version of the DBSCAN algorithm. In this algorithm, two sound source points are considered "density reachable," depending not only on their Euclidean distance and the time interval between their occurrences, but also on a "physical mechanism similarity" metric.

[0020] Specifically, for any two sound source events i and j, the combined distance D(i,j) between them is defined as a weighted function, which has the following form: .

[0021] Where Loc represents the three-dimensional coordinates of the sound source, Time represents the occurrence time, and Mode represents the probability distribution vector of the fracture mode obtained in the second step. ||.||2 represents the Euclidean norm, and CosSim represents the cosine similarity. ws, wt, and wp are weighting coefficients, whose values ​​are pre-calibrated based on the brittle characteristics of the dam's concrete material.

[0022] This comprehensive distance metric ensures that only spatially proximate, temporally continuous, and similarly fractured sound source events are grouped together. By running this four-dimensional DBSCAN algorithm on all located sound source point datasets, the system can automatically identify and remove noise points representing independent, random micro-fractures as outliers, while aggregating a series of micro-fracture events related to the same macroscopic crack propagation process into physically meaningful sound source event clusters.

[0023] Furthermore, for each identified cluster of sound source events, this invention employs a moving least squares method to fit and reconstruct the three-dimensional crack front. Within an event cluster, the sound source points are sorted according to their occurrence time. The algorithm sets a sliding time window with time as the axis. For all sound source points falling within the current time window, the MLS method is used to fit a locally optimal surface, which represents the spatial position and morphology of the crack front at that moment. As the time window slides forward, the system generates a series of continuous local surfaces. These surfaces are pieced together to form a three-dimensional visualization trajectory of the entire process of the macroscopic crack from its initiation to its expansion. This trajectory not only shows the spatial path of the crack but also accurately depicts its geometric evolution during the expansion process, such as bifurcation and merging.

[0024] The fifth step involves inverting the dynamic parameters of crack propagation and quantitatively assessing the damage state. After reconstructing the continuous crack front evolution trajectory, this invention further inverts a series of key dynamic parameters to quantitatively assess the hazard of the crack. First, the crack propagation rate is calculated. The average propagation rate of the crack is obtained by calculating the ratio of the spatial displacement between the geometric centroids of the crack fronts at two adjacent moments to the time interval. By differentiating the displacement of a specific point on the front, the instantaneous propagation rate at that point can also be obtained, thereby identifying the acceleration or hysteresis phase in the crack propagation process. Second, the energy release rate of the crack is inverted. The absolute energy of each acoustic emission event (calculated in the first step) can be converted into the real energy released at the sound source after correction for sensor sensitivity, propagation path attenuation, and geometric diffusion effects. The corrected energy of all sound source events belonging to the same crack front propagation within a time window is accumulated and then divided by the newly added crack area within that time window (calculated from the crack front reconstruction results in the fourth step) to obtain the average energy release rate G for that stage. This parameter is a core indicator in fracture mechanics for measuring the driving force of crack propagation. Finally, this invention defines and calculates a Cumulative Damage Index (CDI). This index is defined as the sum of the corrected energy of all sound source events within a spatial region (such as a finite element mesh element) during a monitoring period, normalized according to the volume of that region. By calculating a three-dimensional CDI distribution map of the entire dam structure, the areas with the most severe damage accumulation can be visually identified, providing direct and quantitative physical evidence for dam safety assessment and maintenance decisions.

[0025] All these inversely derived dynamic parameters, including crack propagation path, rate, energy release rate, and cumulative damage index, together constitute a complete dynamic assessment report of the dam's structural integrity, thereby enabling comprehensive and in-depth tracking and analysis of crack propagation behavior.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by constructing a data analysis process, shifts from simply marking the occurrence of micro-fracture events to deeply mining the physical information in signal waveforms. This enables the identification of fracture modes, high-precision localization of sound sources, reconstruction of continuous fracture fronts, and ultimately, the inversion of relevant dynamic parameters characterizing the risk of fracture propagation. This method significantly improves the signal-to-noise ratio and reliability of monitoring results. Attached Figure Description

[0027] Figure 1 This is a system configuration block diagram of the method described in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the method described in an embodiment of the present invention; Figure 3This is a schematic diagram of the hybrid architecture neural network described in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the reconstruction of the crack front from a cluster of sound source events according to an embodiment of the present invention. Detailed Implementation

[0028] Example 1: Please refer to the appendix. Figure 1 This invention provides a method for tracking the propagation of cracks in a dam based on acoustic emission signal analysis. The method is illustrated through a specific embodiment. The embodiment is set as a typical concrete gravity dam with a height of 120 meters and a crest length of 500 meters. To achieve comprehensive monitoring of the activity of microcracks within the dam structure, a complete acoustic emission monitoring system was deployed. Figure 1 The system configuration is shown. At the physical level, the system comprises a three-dimensional sensor array 101 consisting of 64 broadband piezoelectric acoustic emission sensors. These sensors are pre-embedded in key stress areas inside the dam concrete and laid on the downstream dam face, forming a monitoring network covering the entire dam body.

[0029] Specifically, the sensor model selected is RS-2A, with an effective operating frequency range of 50 kHz to 400 kHz. The analog signals acquired by the sensor array 101 are transmitted via shielded cables to a multi-channel synchronous data acquisition unit 102 deployed within the dam's gallery. This data acquisition unit employs an 18-bit high-precision data acquisition card based on the PCI-Express bus, capable of synchronously sampling all 64 channels at up to 10 megasamples per second (MS / s), ensuring microsecond-level accuracy in signal arrival time acquisition. The digital waveform data stream, after analog-to-digital conversion, is transmitted in real-time via gigabit industrial Ethernet to the central data processing server 103 located in the dam's central control room. This server is equipped with dual Intel Xeon processors, 256GB of memory, and a RAID 5 disk array consisting of solid-state drives, providing powerful computing support for subsequent complex data processing and numerical calculations.

[0030] The complete technical process of this invention is as follows: Figure 2 As shown, it covers five core steps from raw signal acquisition to final damage assessment. The following will provide a detailed explanation of the specific engineering implementation of each step.

[0031] The first step is adaptive preprocessing and feature extraction of acoustic emission signals with multidimensional feature enhancement; this is the workflow. Figure 2Step 201. When the central data processing server 103 receives a continuous, high-bit-rate raw acoustic emission waveform data stream from the data acquisition unit 102, it does not use a traditional fixed-parameter filter for processing. Instead, the server first performs an adaptive bandpass filter based on operating condition identification. The server has pre-built and stored a dam operating condition-noise spectrum feature library. The construction of this feature library took a year. By synchronously recording the macroscopic operating conditions of the dam (e.g., upstream water level, number of floodgates opened, generator output power, temperature distribution on and inside the dam body, etc.) and the corresponding background noise signals collected by the sensor array, and performing power spectral density (PSD) analysis on the latter, a mapping relationship between the operating condition vector and the noise PSD function was established. For example, when the generator is running at full load, the noise spectrum exhibits obvious narrowband peaks at 50 Hz and its harmonics. In a specific embodiment, when the real-time monitoring system identifies the current operating condition as "generator unit No. 3 is running at full load", it calls the corresponding noise PSD model from the feature library. Based on this model, a Wiener filter is dynamically designed and applied to the system. The transfer function of this filter is optimized to minimize the mean square error between the filtered signal and the real crack signal (whose spectrum is typically broadband), thereby suppressing narrowband or broadband background noise under specific operating conditions to the greatest extent and significantly improving the signal-to-noise ratio for subsequent event identification.

[0032] The adaptively filtered data stream is fed into an event recognition and segmentation module. This module abandons the traditional fixed amplitude threshold method and instead employs a sequence anomaly detection model based on a Long Short-Term Memory (LSTM) network. This LSTM network contains two hidden layers, each with 128 neurons, and uses the hyperbolic tangent (tanh) function as the activation function. The network's training dataset contains over 100,000 samples, where positive samples are real crack acoustic emission signals collected during fracture experiments on large concrete specimens in the laboratory, and negative samples are pure noise signal segments collected under various operating conditions of the dam. Through training, the LSTM network learns the temporal evolution pattern of energy release and decay of real crack signals on a microsecond to millisecond timescale, a pattern fundamentally different from random electromagnetic interference or hydraulic noise. In actual operation, the network scans the filtered data stream using a sliding window, accurately identifying the start and end points of even weak acoustic emission events with amplitudes below the average background noise level. Once an event is identified, the system precisely segments it from the continuous data stream, forming an independent event waveform data segment containing complete preceding and following waveforms, laying the foundation for subsequent refined analysis.

[0033] For each successfully segmented event waveform data segment, the central data processing server executes a complex multi-domain joint feature extraction process to construct a high-dimensional signal feature supervector (SFHV) capable of comprehensively characterizing the physical essence of the event. In the time domain, six key parameters are calculated and quantified. The signal rise time is precisely defined as the time it takes for the signal amplitude to rise from three times the root mean square (RMS) value of the background noise (denoted as Vth) to 90% of the peak signal amplitude (Vp), typically ranging from 5 to 80 microseconds. The signal duration is defined as the total duration from when the signal amplitude first exceeds Vth to when it last falls back below Vth, usually between 100 and 2000 microseconds. The signal ringing count is the number of times the signal waveform crosses Vth upwards within the duration, reflecting the signal's oscillation decay characteristics. The absolute energy of the signal is obtained by integrating the squared voltage values ​​(in volts) across all sampling points within the event waveform data segment over time (in seconds), and then dividing by the sensor termination resistance (typically 50 ohms). Its physical unit is joules, and its magnitude typically ranges from attojoules (1aJ = 10⁻¹⁸ J) to femtojoules (1fJ = 10⁻¹⁵ J). The peak amplitude of the signal, Vp, is measured in microvolts. Furthermore, the kurtosis and skewness of the signal waveform are calculated; these two dimensionless statistical moments characterize the sharpness and asymmetry of the signal waveform, respectively.

[0034] In the frequency domain, a Hanning window function is first applied to the event waveform data segment to suppress spectral leakage, followed by a Fast Fourier Transform (FFT) to obtain its energy spectral density map. Peak frequencies, i.e., the frequencies with the highest energy density, are extracted from this spectrum; for example, the peak frequencies of tensile cracks are typically concentrated between 80-150 kHz. Simultaneously, the frequency centroid is calculated, which is the first moment of the spectral energy about frequency divided by the total energy, reflecting the overall distribution of the spectral energy. The low-frequency to high-frequency energy ratio (LF / HF ratio) is also calculated.

[0035] Specifically, for the RS-2A sensor, the low-frequency band is defined as 50-225 kHz and the high-frequency band as 225-400 kHz. This ratio has been found to be somewhat correlated with the size of the crack source, with larger-scale rupture events tending to produce richer low-frequency components.

[0036] In the time-frequency joint domain, to capture the dynamic evolution of signal energy over time and frequency, continuous wavelet transform (CWT) is employed, and the Morlet mother wavelet, with its good time-frequency localization characteristics, is selected to decompose the event waveform. This generates a three-dimensional energy spectrum (time-frequency map or spectral rogram) of energy. From this spectrum, the instantaneous entropy of the energy distribution is calculated along the time axis to quantify the complexity and suddenness of the signal energy release process; simultaneously, the spectral entropy of the energy distribution is calculated along the frequency axis to quantify the richness and disorder of the signal frequency components. All the extracted time-domain, frequency-domain, and time-frequency-domain parameters, totaling more than ten items, are combined into a unique, standardized signal feature hypervector (SFHV) after Z-score normalization to eliminate the influence of dimensions, and used as the input to the second-step classification model.

[0037] The second step involves intelligent identification and classification of fracture patterns based on physically constrained deep learning networks, and the corresponding process is as follows: Figure 2 Step 202. To fundamentally answer the question of "how" the crack expands, this invention designs and trains a physically interpretable hybrid architecture deep neural network, the specific structure of which is as follows: Figure 3As shown, this network model precisely classifies each acoustic emission event into one of three basic fracture modes: opening (Mode I), sliding (Mode II), or tearing (Mode III). This hybrid architecture network comprises two parallel information processing branches. The first branch is a one-dimensional convolutional neural network (1D-CNN) 301, whose input is a normalized segment of the original event waveform data (e.g., 4096 sampling points in length). This branch consists of three convolutional layers with kernel sizes of 7x1, 5x1, and 3x1, and the number of kernels per layer is 32, 64, and 128, respectively. Each convolutional layer is followed by a rectified linear unit (ReLU) activation function and a max-pooling layer (pooling size 2x1), automatically learning and extracting subtle, abstract features related to the sound source radiation pattern in the local waveform morphology. The second branch is a fully connected network (FCN) 302, whose input is the signal feature hypervector (SFHV) constructed in the first step. This branch contains three hidden layers with 256, 128, and 64 neurons respectively, and also uses the ReLU activation function to learn the distribution patterns of different fracture modes in the high-dimensional feature space from macroscopic statistical features. The output feature vectors of the two branches are merged in the concatenation layer 303 at the back end of the network to form a comprehensive feature representation that integrates waveform morphology and macroscopic features. This comprehensive feature is then fed into a Softmax classification layer 304 containing three neurons, and finally outputs a three-dimensional probability distribution vector [P(Mode I), P(Mode II), P(Mode III)], where the sum of each component is 1, representing the probability that the event belongs to one of the three fracture modes.

[0038] The training process of this neural network is a key aspect of this invention. Instead of relying solely on limited field data, the training incorporates physical constraints from fracture mechanics, ensuring the model's generalization ability and reliability. The training dataset consists of two parts. The first part comprises synthetic acoustic emission waveforms generated through large-scale numerical simulations using commercial finite element software (such as ABAQUS). In the simulation, a micromechanical model of concrete with random aggregate distribution was constructed, and the Hillerborg cohesion model was used to describe crack initiation and propagation. By applying different boundary conditions (such as uniaxial tension and pure shear), specific crack propagation modes can be accurately induced, and the elastic wave field radiated in space is calculated based on the moment tensor of the crack source, thereby generating synthetic acoustic emission waveforms with clear fracture mode labels. The second part of the data comes from laboratory calibration experiments. By loading standard-sized concrete specimens (such as beam specimens for three-point bending tests to generate Mode I cracks, or special geometric specimens for shear tests to generate Mode II cracks), and simultaneously acquiring acoustic emission signals during the loading process, a large number of real signal samples with clear physical backgrounds were obtained. By combining these two sets of data and using classification cross-entropy as the loss function and Adam optimizer for network training, the resulting classification model can attribute any input acoustic emission event to its underlying physical fracture mechanism with over 95% accuracy.

[0039] The third step involves high-precision sound source localization of the anisotropic velocity field by fusing fracture mode information, corresponding to the process. Figure 2 Step 203. Traditional positioning methods typically assume that sound waves propagate at a constant speed in concrete, which is seriously inconsistent with the actual situation of complex stress distribution inside dams. To solve this problem, this invention proposes an iterative inversion positioning method that takes into account the influence of stress. First, inside the central data processing server 103, a full-size three-dimensional finite element model is constructed based on the dam design drawings, the elastic modulus of concrete (e.g., 30 GPa), Poisson's ratio (e.g., 0.2), and boundary conditions such as upstream water level and ambient temperature monitored in real time by sensors. Through static structural analysis, this model can calculate the three-dimensional stress tensor distribution σ(x,y,z) at any point inside the dam. According to the theory of acoustoelasticity, the propagation speed of P-waves in the medium is closely related to the stress state. Based on this, this invention establishes a stress-dependent anisotropic P-wave velocity field model V(x,y,z). In this model, the wave velocity at a point is no longer a scalar, but a tensor related to the principal stress direction and magnitude at that point. Under normal circumstances, the wave velocity along the principal direction of compressive stress will increase, while the wave velocity along the principal direction of tensile stress will decrease, with the change being approximately 1% for every 10 MPa of stress.

[0040] When locating an acoustic emission event, the system first obtains the fracture mode probability distribution vector of the event from the second step. This vector information is used in two key aspects. First, it is used to construct the radiation pattern of the sound source. For example, in a pure Mode I (opening) crack, the P-wave radiation energy is strongest in the direction perpendicular to the crack surface and zero in the direction parallel to the crack surface. The system synthesizes a weighted radiation pattern model based on the fracture mode probability distribution. Second, different fracture modes correspond to different initial P-wave and S-wave energy ratios. This information is used to optimize the parameters of the automatic acquisition algorithm (such as the Akaike information criterion method) for the first arrival time of the P-wave on each sensor channel.

[0041] The localization process itself is a nonlinear optimization problem. The algorithm first randomly generates a set (e.g., 1000) of candidate source points within a region where sound sources might exist (typically roughly defined by the first few triggered sensors). For each candidate source point, the system uses the aforementioned anisotropic velocity field model and the efficient Fast Marching Method to solve the equations, accurately calculating the theoretical shortest propagation time from that point to each sensor in the sensor array. Subsequently, an objective function is constructed, defined as the weighted sum of squares of the residuals between the actual arrival time of the P-wave picked up by all triggered sensors and the theoretical propagation time. The weighting factor considers two factors: the signal-to-noise ratio (SNR) of the received signal from the sensor, and the theoretical received energy in the sensor's direction calculated based on the sound source radiation pattern model. Sensors with high SNR and high theoretical received energy are assigned higher weights.

[0042] Finally, a global optimization algorithm, such as simulated annealing, is used to search for the point in three-dimensional space that minimizes the objective function. This point is then determined as the final sound source location (x, y, z). In this way, the physical information of the fracture mode is deeply integrated with the stress state of the dam, resulting in an order-of-magnitude improvement in positioning accuracy compared to traditional uniform velocity model methods. In areas with well-placed sensors, the positioning error can be effectively controlled within 0.2 meters.

[0043] The fourth step involves identifying sound source event clusters and reconstructing crack fronts based on spatiotemporal correlation and physical similarity. The corresponding process is as follows: Figure 2 Step 204. Individual sound source localization points are discrete. To track the continuous crack propagation process, sound source points associated with the same physical process must be aggregated. This invention proposes an improved four-dimensional density clustering algorithm (DBSCAN) for this purpose. The core of this algorithm lies in its definition of a new comprehensive distance metric D(i,j) that reflects physical correlation. For any two sound source events i and j, their comprehensive distance is defined as: Where Loc is the three-dimensional coordinate vector of the sound source, Time is the timestamp of the event, and Mode is the probability distribution vector of the fracture mode obtained in the second step. ||.||2 represents the Euclidean distance, and CosSim represents the cosine similarity between two probability vectors. The weight coefficients ws, wt, and wp are pre-calibrated according to the brittle fracture characteristics of the dam concrete material. In one embodiment, their values ​​are ws=1.0, wt=500.0 (m / s), and w_p=2.0 (m), respectively. Such values ​​mean that temporal proximity (reflecting the crack propagation rate) is given a higher weight. This comprehensive distance metric ensures that only sound source events that are spatially close, temporally consecutive, and have highly similar fracture mechanisms (opening, slip, etc.) are considered "density reachable" by the algorithm and thus classified into the same event cluster.

[0044] After performing this four-dimensional DBSCAN algorithm on all located sound source point datasets (with algorithm parameters set for example, neighborhood radius Eps = 1.5 meters and minimum number of neighbors MinPts = 5 for the core object), the system can automatically aggregate a series of micro-rupture events representing the same macroscopic crack propagation process into one or more sound source event clusters, while identifying and removing those spatiotemporally isolated or physically different random noise sources as outliers.

[0045] For each identified event cluster containing a series of time-ordered sound source points, this invention further employs the moving least squares (MLS) method to dynamically reconstruct the three-dimensional crack front, the process of which is as follows: Figure 4 As shown, the algorithm sets a sliding time window with a certain width (e.g., 2.0 seconds) along the time axis. For all sound source points 401 falling within the current time window, a locally optimal surface 402 is fitted using the MLS method. The MLS method generates a smooth and continuous surface even when the data points are sparse or unevenly distributed by performing a weighted least squares fitting on the neighboring points around each point. As the time window slides forward along the time axis (e.g., with a step size of 0.5 seconds), the system generates a series of continuous local surfaces. By stitching together these temporally continuous surfaces, a three-dimensional visualization trajectory 403 of the entire process of a macroscopic crack from its initiation to its expansion is formed. This method can not only clearly show the spatial expansion path of the crack, but also accurately depict its geometric evolution during the expansion process, such as the expansion from an initial small crack, branching into multiple branches, or merging with other cracks.

[0046] The fifth step involves inverting crack propagation dynamic parameters and quantitatively assessing the damage state, with the corresponding workflow. Figure 2Step 205. After reconstructing the continuous fracture front evolution trajectory, this invention further inverts a series of key dynamic parameters to achieve a quantitative assessment of fracture hazard. First, the fracture propagation rate is calculated. By calculating the spatial displacement between the geometric centroids of the fracture front fitted by MLS at two adjacent time points (e.g., tk and t{k+1}), and then dividing by the time interval (t{k+1}-tk), the average propagation rate of the fracture within that time period can be obtained. By performing time differentiation on the displacement trajectory of a specific point on the front, the instantaneous propagation rate at that point can also be obtained, thereby identifying the acceleration, stabilization, or stagnation phases in the fracture propagation process.

[0047] Secondly, the core parameter in fracture mechanics inversion—the energy release rate G—is crucial. For each acoustic emission event, the absolute energy (in joules) calculated in the first step needs to be corrected to convert it into the true elastic wave energy released at the sound source. The correction process considers three main factors: geometric diffusion attenuation during signal propagation (energy is inversely proportional to the square of the propagation distance), frequency-dependent attenuation due to the inelastic absorption of the medium (compensated by the pre-determined concrete material quality factor Q), and the sensor's own frequency response and sensitivity (corrected using the sensor's calibration certificate). The corrected energies of all sound source events belonging to the same crack front advancement within a time window (e.g., tk to t{k+1}) are summed to obtain the total energy release ΔE. Simultaneously, using the crack front reconstruction results from the fourth step, the newly added crack area ΔA within this time window is calculated. The average energy release rate G for this stage is then calculated as G = ΔE / ΔA, with units of joules per square meter (J / m²). 2 This parameter directly reflects the magnitude of the driving force for crack propagation and is a key indicator for assessing crack stability.

[0048] Finally, this invention defines and calculates a Cumulative Damage Index (CDI) to achieve a spatial quantitative assessment of the overall damage state of the dam. This index is defined on a three-dimensional mesh corresponding to the dam's finite element model. For each mesh cell (or voxel) in the model, the corrected energy of all sound source events located within that cell during the monitoring period is summed and then divided by the cell's volume. By calculating the three-dimensional CDI distribution of the entire dam structure, an intuitive damage "heatmap" can be generated, highlighting the areas with the most severe damage accumulation and the most significant decline in structural integrity. All these inverted dynamic parameters, including the three-dimensional propagation path of cracks, propagation rate, energy release rate G, and the global CDI distribution map, together constitute a complete, dynamic, and quantitative assessment report of the dam's structural integrity, providing unprecedented, physically based, direct evidence for dam safety early warning and maintenance decisions.

[0049] To verify the superiority of the method described in this invention over the prior art, a comparative experiment was conducted.

[0050] Example The complete method disclosed in this invention is employed. Within a prefabricated large concrete block (2m x 2m x 2m), the slow propagation of a crack is manually controlled via a hydraulic device. Sixteen RS-2A sensors are embedded within the concrete block. The monitoring system records all acoustic emission signals as the crack propagates from 10cm to 80cm and processes them using the five-step method of this invention.

[0051] Comparative Example Traditional acoustic emission monitoring and analysis methods were employed. Signals acquired in the same experiment were filtered using a fixed 40-400kHz bandpass filter, and event detection was performed using a fixed amplitude threshold of 45dB. Sound source localization employed an iterative algorithm based on the least squares method, assuming a constant P-wave velocity of 4000 m / s in the concrete. The analysis results are only discrete three-dimensional distribution maps of sound source localization points; damage assessment is based on the number of events per unit volume (i.e., event density).

[0052] The comparison of the data in the table above clearly shows that the method proposed in this invention is significantly superior to traditional acoustic emission technology in all key performance indicators, such as the sensitivity and accuracy of signal detection, the precision of sound source localization, the ability to reconstruct crack morphology, and the quantitative assessment of damage state.

[0053] This invention elevates acoustic emission monitoring technology from a simple event alarm to a dynamic mechanism capable of accurately tracking and analyzing the internal damage evolution process of structures by constructing a complete analytical chain from in-depth mining of signal physical characteristics to inversion of crack propagation dynamics. This provides strong technical support for ensuring the safety of major engineering structures.

Claims

1. A method for tracking crack propagation of a dam based on acoustic emission signal analysis, characterized by, The method comprises the following steps: S1: multi-dimensional feature enhancement preprocessing and extraction are performed on the collected acoustic emission signals, the preprocessing comprises dynamically determining the center frequency and bandwidth of the Wiener filter according to the dam operation condition to maximize the signal-to-noise ratio, and performing event identification and segmentation based on a pre-trained long short-term memory network, and the feature extraction generates a high-dimensional signal feature hyper vector; S2: a deep learning network based on the physical constraints of fracture mechanics is used to identify and classify the fracture mode of the acoustic emission event, and a fracture mode probability distribution vector of the acoustic emission event is output; S3: the fracture mode information is fused with an anisotropic velocity field model related to the dam stress field, and high-precision acoustic source positioning is performed through an iterative inversion algorithm; S4: a four-dimensional density clustering algorithm based on space-time correlation and physical similarity is used to identify clusters of the positioned acoustic source events, and a moving least squares method is used to reconstruct a three-dimensional crack front; S5: the crack propagation dynamics parameters are inverted and the damage state is quantitatively evaluated according to the reconstructed crack front trajectory and the corrected acoustic emission event energy.

2. The method of claim 1, wherein, The multi-dimensional surveying and characterization of the dam cracks comprises: a pre-stored operating condition-noise spectrum feature library, which records the power spectral density functions of background noise under different operating conditions; when receiving a real-time acoustic emission data stream, identifying the current macro operating condition of the dam, and calling the corresponding noise power spectral density function from the feature library; designing a Wiener filter based on the noise power spectral density function to filter the original acoustic emission data stream; inputting the filtered signal into a sequence anomaly detection module based on a pre-trained long short-term memory network to identify and segment effective acoustic emission events and form independent event waveform data segments.

3. The method of claim 2, wherein, The multi-dimensional feature extraction comprises: extracting time domain features from the event waveform data segments, including rise time, duration, ring count, absolute energy, peak amplitude, kurtosis and skewness; extracting frequency domain features, including peak frequency, frequency centroid and low-to-high frequency band energy ratio; extracting time-frequency joint domain features, including instantaneous entropy of energy on the time axis and spectral entropy on the frequency axis; and combining the extracted domain parameters into the signal feature hyper vector.

4. The method of claim 1, wherein, The deep learning network based on the physical constraints of fracture mechanics is a hybrid architecture neural network, which includes a one-dimensional convolutional neural network branch for processing the event waveform data segments and a fully connected network branch for processing the signal feature hyper vector; the training data set of the neural network is composed of physically constrained finite element simulation waveforms and controlled experimental acquisition waveforms; the neural network outputs a probability distribution vector of the acoustic emission event belonging to three basic fracture modes of opening, sliding and tearing.

5. The method of claim 1, wherein, The three-dimensional finite element model of the dam is built, the three-dimensional stress tensor distribution of each point inside the dam is calculated, and a stress-dependent anisotropic P-wave velocity field model is established based on the theory of acoustic-elastic effect; the fracture mode probability distribution is used to obtain fracture mode information, which is used to determine the spatial directivity pattern of the acoustic source radiation energy and optimize the weight of the time-picking algorithm; a nonlinear optimization iterative algorithm is used to search for the acoustic source position in the anisotropic velocity field model that minimizes the weighted residual sum of squares between the actual P-wave arrival time picked up by all sensors and the theoretical propagation time.

6. The method of claim 1, wherein, The four-dimensional density clustering algorithm based on space-time correlation and physical similarity is an improved DBSCAN algorithm; in the algorithm, two acoustic source points are considered to be "density reachable" depending on their Euclidean distance, the interval of occurrence time, and the similarity of physical mechanisms determined by the fracture mode probability distribution vector; by executing the four-dimensional DBSCAN algorithm, acoustic source events that are spatially adjacent, temporally continuous, and similar in fracture mechanism are aggregated into acoustic source event clusters.

7. The method of claim 6, wherein, The reconstruction of the three-dimensional crack front trajectory of the identified acoustic source event cluster includes: In an event cluster, the acoustic source points are sorted by their occurrence time; A sliding time window is set; For the acoustic source points falling within the time window, a local optimal surface is fitted using the moving least squares method, and the surface represents the spatial position and shape of the crack front at that time; As the time window slides forward, a series of continuous local surfaces are generated, which are spliced to form a three-dimensional extension trajectory of the macroscopic crack from initiation to extension.

8. The method of claim 1, wherein, The inversion of crack propagation dynamics parameters includes: The ratio of the spatial displacement between the geometric centroids of two adjacent time instants and the time interval is calculated to obtain the average propagation rate of the crack; The displacement of a specific point on the front surface is differentiated to obtain the instantaneous propagation rate of the point; The corrected energy of all acoustic source events belonging to the same crack front propagation within a time window is accumulated and divided by the newly added crack area in the time window to obtain the average energy release rate at that stage.

9. The method of claim 1, wherein, The quantitative evaluation of the damage state includes: A cumulative damage index (CDI) is defined and calculated, which is the volume-normalized value of the total corrected energy of all acoustic source events in a certain spatial region within the monitoring period; By calculating the three-dimensional CDI distribution map of the entire dam structure, the region with the most serious damage accumulation is identified.

Images