Method for evaluating dynamic characteristics of acoustic emission data stream of composite adhesive test sample

Abstract

This invention relates to the field of acoustic emission detection and health monitoring of composite materials, aiming to provide a dynamic characteristic evaluation method for acoustic emission data streams of bonded composite material specimens. The method includes: acquisition and processing of acoustic emission signal data, dynamic processing of Laplace scores, dynamic processing of multi-cluster feature selection, and dynamic evaluation of conventional acoustic emission features. By analyzing the damage characterization preference of acoustic emission features, the method yields evaluation results on whether acoustic emission features are suitable for damage pattern recognition, anomalous signal detection, or damage process characterization. This invention enables dynamic evaluation of acoustic emission data streams of bonded composite material specimens, which is of great significance for uncovering damage characterization-related information contained in the features.

Description

Technical Field

[0001] This invention relates to the field of acoustic emission detection and health monitoring of composite materials, and in particular to a method for evaluating the dynamic characteristics of acoustic emission data streams of bonded composite material samples. Background Technology

[0002] Composite material engineering structures are often extremely large and complex systems, exhibiting diverse structural forms, a wide variety of materials, complex and variable load conditions, and harsh and unpredictable service environments. Furthermore, the manufacturing and assembly processes are subject to numerous uncertainties. Simultaneously, composite material damage and failure exhibit scale effects. Therefore, with the widespread application of composite materials in various engineering fields, the reliability and service safety of composite material structures have become increasingly prominent issues. To achieve the goal of long-term stable service of composite material engineering structures, avoid sudden structural failures, and reduce maintenance costs, it is essential to develop non-destructive damage detection technologies and intelligent health monitoring technologies to ensure the long-term reliability of composite material structures.

[0003] Existing composite material structural health monitoring technologies largely rely on various well-developed non-destructive testing and condition monitoring methods, with damage identification based on all acoustic emission data generated during the failure process. Since all damage generates elastic waves, these waves can be detected to monitor composite material structures—a process known as acoustic emission monitoring. As damage to composite material samples continues to occur and evolve, acoustic emission signals are constantly generated and recorded by the system; thus, the failure process is reflected to some extent as a dynamic acoustic emission data stream. Therefore, the arrival of new signals may indicate unknown hidden patterns in the data and the real-time health status of the composite material sample. However, existing composite material structural health monitoring technologies often struggle to obtain a complete picture of acoustic emission data in real-time structural health monitoring. More importantly, the degree of acoustic emission waveform distortion is positively correlated with the degree of damage accumulation. Acoustic emission data, arriving as a continuous stream, may influence damage identification and may contain information related to the damage process. The influence of this data stream on feature selection remains unclear in existing damage identification technologies. For example, it is impossible to distinguish or make judgments about the damage identification capability of the feature subset used for damage identification in the whole domain data during the damage process, the consistency between the filtered features and invalid features in the acoustic emission data stream, and the acoustic emission features in the data stream used to quantitatively describe the degree of damage.

[0004] Therefore, it is necessary to study the dynamic feature evaluation methods contained in acoustic emission data streams, interpret the information related to material damage characterization, and explore the influence of damage degree on the information contained in the features. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a dynamic characteristic evaluation method for acoustic emission data stream of composite adhesive specimens.

[0006] To solve the technical problem, the solution of the present invention is:

[0007] A dynamic feature evaluation method for acoustic emission data stream of composite bonded specimens is provided. The method includes: acquisition and processing of acoustic emission signal data, dynamic processing of Laplace score, dynamic processing of multi-cluster feature selection, and dynamic evaluation of conventional acoustic emission features. By analyzing the damage characterization preference of acoustic emission features, the evaluation results of whether acoustic emission features are suitable for damage pattern recognition, anomalous signal detection, or damage process characterization are obtained.

[0008] As a preferred embodiment of the present invention, the acquisition and processing of the acoustic emission signal data specifically includes the following steps:

[0009] (1) Prepare composite adhesive specimens according to conventional methods, perform uniaxial tensile tests on the specimens, and obtain acoustic emission signal data;

[0010] (2) The obtained acoustic emission signal data is processed as follows:

[0011] Define the features to be collected: acoustic emission signals are continuously generated in the time dimension, while the total number of features remains constant in the feature dimension; assume that the system records a total of N at time t. (t) There are 1 signal, each signal x i (i = 1, 2, ..., N) (t) ) is represented as a p-dimensional vector, i.e., (x i1 x i2 …x ip The entire acoustic emission signal data is converted into N. (t) A p-dimensional vector serves as the basic data content and form for the subsequent two dynamic processing methods.

[0012] As a preferred embodiment of the present invention, the dynamic processing of the Laplace score specifically includes the following steps:

[0013] Feature dynamic processing is performed based on the Laplace score algorithm to construct a vertex with N vertices at time t. (t) G, the nearest neighbor graph (t) Each vertex represents an acoustic emission data point; the Laplacian matrix at time t is calculated, and each feature is de-averaged; the Laplacian score of the r-th feature at time t is calculated to obtain the Laplacian scores of the features in the static dataset; at time t+1, the graph G is processed... (t) Update the graph to construct the nearest neighbor graph G at time t+1. (t+1)Calculate the adjacency matrix at time t+1; after obtaining the weight matrix, calculate the Laplacian matrix and degree matrix at time t+1; then perform de-averaging on each feature at time t+1 to obtain the Laplacian score of each feature at time t+1.

[0014] As a preferred embodiment of the present invention, the dynamic processing of multi-cluster feature selection specifically includes the following steps:

[0015] Let N be the time t. (t) N consists of acoustic emission impact signals (t) The ×p matrix is ​​X (t) Singular Value Decomposition (SVD) is performed on its transpose to solve for... For the LASSO regression problem of the target variable; after obtaining the coefficient matrix Φ (t) Then, for any sound emission characteristic f r (r = 1, 2, ..., p), calculate the multi-cluster feature selection score for use in selecting feature subsets in the static dataset; at time t+1, to avoid repeated input of previous acoustic emission data, a matrix sketch method is introduced to effectively preserve X. (t) The low-rank approximation is then used; regularized regression is then used to perform feature selection, and the feature coefficient matrix at time t+1 is obtained. The scores of each acoustic emission feature at time t+1 are then calculated.

[0016] As a preferred embodiment of the present invention, when performing dynamic evaluation of conventional acoustic emission characteristics, the conventional acoustic emission characteristics include at least: rise time (RT), count (C), energy (E), duration (TD), peak amplitude (PA), average frequency (AF), root mean square (RMS), average signal level (ASL), initial frequency (IF), signal strength (SS), absolute energy (AbE), centroid frequency (FG), peak frequency (PF), rise angle (RA), and decay angle (DA).

[0017] As a preferred embodiment of the present invention, the damage characterization preference of acoustic emission characteristics is analyzed and an evaluation result is obtained. Specifically, the acoustic emission characteristics of acoustic emission impact signals are dynamically evaluated. Different acoustic emission characteristics of sample damage signals show different damage characterization preferences, and the evaluation result of whether they are suitable for singular signal detection, damage process characterization or damage pattern recognition is obtained. In subsequent acoustic emission experiments, an adaptation selection is made based on the evaluation result.

[0018] This invention further provides an application method for the evaluation results obtained from the dynamic characteristic evaluation method based on the acoustic emission data stream of the composite adhesive specimen, comprising the following steps:

[0019] (1) By wavelet packet decomposition, two sets of characteristics of the acoustic emission impact signal are obtained: wavelet packet energy ratio and wavelet packet energy.

[0020] (2) According to the steps described in claim 1, the self-extracted features based on wavelet packet decomposition are dynamically evaluated. The results show that the wavelet packet energy ratio is suitable for the classification of acoustic emission impact signals, but not for the detection of singular signals and the characterization of damage processes. Wavelet packet energy can play a damage characterization role similar to acoustic emission energy, signal intensity and absolute energy.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. The dynamic feature evaluation method for acoustic emission data streams provided by this invention enables dynamic evaluation of acoustic emission data streams from bonded composite material specimens, which is of great significance for uncovering information related to damage characterization contained in the features. This invention can measure the damage characterization role of each feature as damage persists from the perspectives of damage pattern recognition, singular signal detection, and damage process characterization, and elucidates the damage characterization preference of each feature. Therefore, the role of any acoustic emission feature can be analyzed to determine whether it is suitable for singular signal detection, damage process characterization, or damage pattern recognition, and appropriate selection can be made for different acoustic emission experimental contents.

[0023] 2. This invention is based on global dynamic analysis and evaluation, so it can distinguish the explanatory power of feature subsets in the damage persistence process on the whole domain data, consider the other roles of features that cannot be used for damage pattern recognition, and explore acoustic emission features in the data stream that can globally and quantitatively describe the degree of damage.

[0024] 3. Taking damage pattern recognition of composite adhesive samples as an example, the present invention can exclude features that are not suitable for damage pattern recognition, thus significantly reducing the dimensionality of the independent variables and eliminating the need for dimensionality reduction operations when using deep neural networks for damage pattern recognition; furthermore, since the selected features are highly correlated with damage pattern recognition, the results obtained have high accuracy and precision. Attached Figure Description

[0025] Figure 1 This is a geometric morphology diagram of a composite single-lap joint (SLJ) specimen;

[0026] Figure 2 This is a diagram showing the form of the acoustic emission data stream;

[0027] Figure 3 The real-time Laplace scores for 15 common features are: (a) SLJ1, (b) SLJ2, and (c) SLJ3.

[0028] Figure 4The real-time ranking scores of the multi-cluster feature selection algorithm for 15 common features are: (a) Sample SLJ1, (b) Sample SLJ2, (c) Sample SLJ3.

[0029] Figure 5 It is a statistical graph of the distribution matrix of seven features suitable for damage pattern recognition;

[0030] Figure 6 This is a statistical graph of the distribution matrix of eight features that are not suitable for damage pattern recognition;

[0031] Figure 7 The real-time Laplace score of wavelet packet energy ratios in eight characteristic frequency bands: (a) Sample SLJ1, (b) Sample SLJ2, (c) Sample SLJ3. Detailed Implementation

[0032] The present invention will be described in detail below with reference to the accompanying drawings and embodiments, and the objectives and effects of the present invention will become more apparent.

[0033] Part One: Contents of Embodiments of the Invention

[0034] Damage pattern recognition refers to the process of processing the raw waveform signals and their derived features recorded by various sensors to uncover the acoustic emission characteristics of various damage mechanisms. It is another major goal of damage diagnosis and an important component of structural health monitoring.

[0035] Exotic signal detection refers to the detection of acoustic emission data signals with very large values ​​in certain characteristics, which are significantly different from most signals. Typically, when the damage to a sample reaches a certain level, an exotic signal will be triggered, and the sample will fail completely shortly thereafter. The relevant information from exotic signal detection can be used to prevent the sudden failure of the sample and is of great significance for damage assessment of composite materials.

[0036] Damage process characterization refers to analyzing the damage mechanism and damage evolution during the failure process of a test specimen. By observing the changes of various features in the time domain, the correlation between feature values ​​and damage degree is discovered, i.e., the sensitivity of features to the damage process. This allows for the analysis of the failure mechanism of the specimen and the extraction of new features sensitive to the damage process.

[0037] Therefore, by analyzing the damage characterization bias of acoustic emission features, we can determine whether they are suitable for damage pattern recognition, singular signal detection, or damage process characterization. The significance is that by identifying the subset of features used for damage identification across the entire dataset during the damage persistence process, we can distinguish between filtered features and invalid features in the acoustic emission data stream, and identify acoustic emission features in the data stream used to quantitatively describe the degree of damage. This allows us to analyze the role of any acoustic emission feature and determine whether it is suitable for singular signal detection, damage process characterization, or damage pattern recognition. For different acoustic emission experiments, appropriate selection can be made accordingly.

[0038] The present invention provides a feature selection method for dynamic acoustic emission data streams of composite bonded specimens, comprising: acquisition and processing of acoustic emission signal data, dynamic processing of Laplace scores, dynamic processing of multi-cluster feature selection, dynamic evaluation of conventional acoustic emission features, and finally analysis of the role of any acoustic emission feature to determine whether it is suitable for singular signal detection, damage process characterization, or damage pattern recognition. Specifically:

[0039] 1. Acquisition and processing of acoustic emission signal data

[0040] Taking the FRP single-lap joint (SLJ) specimen prepared according to the international standard ASTM D5868 as an example, it is made of two glass fiber reinforced resin matrix composite laminates (EW301F / ACTECH1203) bonded together with laminating resin MGS L285. The specimen width is 25 mm. The specimen geometry and layup angle are shown in the table below. The specimen geometry is shown in the attached figure. Figure 1 .

[0041]

[0042] Two sensors were fixed to the SLJ sample with adhesive tape, and the contact surface between them was filled with acoustic emission coupling agent to eliminate air interference. Both broadband sensors operate in the frequency range of 100-1000 kHz. The peak frequency and peak response amplitude are approximately 520 kHz and -60 dB, respectively. The lower and upper limits of the analog filters for both channels were set to 5 kHz and 450 kHz, respectively, to reduce signal aliasing. The signals detected by the sensors were first preprocessed by two 2 / 4 / 6AST preamplifiers with a gain of 40 dB, and then recorded in the acoustic emission instrument PAC-SAMOS 48. The acoustic emission signals were recorded in impact-defined form, with the relevant operational settings including threshold, peak definition time, impact definition time, impact latch-up time, and maximum duration set to 40 dB, 50 μs, 200 μs, 300 μs, and 500 ms, respectively. Before formal data acquisition, lead-breaking tests were repeated dozens of times, and the average response amplitude was measured to be greater than 95 dB, verifying the reliability of the acquisition settings and the accuracy of linear positioning. The average wave velocity was measured to be approximately 3200 m / s. Uniaxial tensile tests were performed on the SLJ sample at a sampling rate of 1 MHz to obtain acoustic emission signal data.

[0043] The obtained acoustic emission signal data is processed as follows:

[0044] As damage to the composite material specimen continues to occur and evolve, acoustic emission signals are constantly generated and recorded by the system. In other words, the failure process of the specimen is reflected to some extent as a dynamic acoustic emission data stream. In acoustic emission applications, the features to be acquired are often pre-defined in the system before acquisition. Therefore, acoustic emission signals are continuously generated in the time dimension, while the total number of features remains constant in the feature dimension.

[0045] As attached Figure 2 As shown, f r (r = 1, 2, ..., p) and x i (i = 1, 2, ...) represent the r-th dimension acoustic emission feature and the i-th acoustic emission impact signal, respectively. Each impact signal contains p features, and the set of features is used to represent the impact signal to achieve damage pattern recognition and failure process analysis. Without manual secondary feature extraction, the total number of features is a constant. In contrast, the total number of impact signals is unknown before sample failure. The impact of data flow on feature correlation and redundancy assessment remains unknown. Therefore, two basic static unsupervised feature selection algorithms are introduced and modified: Laplace scoring and multi-cluster feature selection. Assume that at time t, the system records a total of N... (t) There are 1 signal, each signal x i (i = 1, 2, ..., N) (t) ) can be represented as a p-dimensional vector, i.e., (x i1 x i2 …xip The entire acoustic emission signal data is converted into N. (t) A p-dimensional vector serves as the basic data content and form for the subsequent two dynamic processing methods.

[0046] Due to the high feature dimensionality, dynamic evaluation of the features only begins when the accumulated impact signals exceed a certain number, which is set to 1000. Thereafter, the evaluation results are continuously updated as the number of signals increases, with an update frequency of once every 100 new impact signals. Here, 100 is merely a semi-empirical number used to present the dynamic evaluation results. If this number is too large, dynamic information about the features may be missed; conversely, if it is too small, computational resources may be wasted.

[0047] 2. Dynamic processing of Laplace scores

[0048] The Laplace score is based on the assumption that two points that are close to each other are more likely to be classified into the same class. For each acoustic emission feature, it measures its ability to preserve the local geometry of the data space. The larger the LS value of a feature, the stronger its clustering ability. The specific method for dynamic feature processing based on the Laplace score algorithm is as follows:

[0049] The number of vertices constructed at time t is N. (t) G, the nearest neighbor graph (t) Each vertex represents an acoustic emission data point. If x i It is x j One of the k nearest neighbors, or x j It is x i If a vertex is one of the k nearest neighbors, then connect the i-th vertex and the j-th vertex with an edge. Let G be a nearest neighbor graph. (t) Construct an adjacency matrix W (t) If x i and x j If the matrix is ​​connected, then the corresponding adjacent elements W in the matrix are connected. ij Set as Where e and η represent the natural logarithm and a suitable constant, respectively; otherwise, the adjacency elements are set to zero. The resulting adjacency matrix W is calculated in this way. (t) It is also considered to be graph G (t) The adjacency matrix. Define the value of the original acoustic emission data on the r-th feature as... The Laplace matrix at time t is calculated as shown in equation (1).

[0050] L (t) =D (t) -W (t) (1)

[0051] Where: degree matrix D (t) It is a diagonal matrix, which can be represented as D(t) =diag(W (t) 1), 1 = (1,1,...,1) T To avoid the graph construction being affected by the singular data in certain features, each feature is de-averaged, as shown in Equation (2).

[0052]

[0053] The Laplace score of the r-th feature at time t can be calculated according to equation (3).

[0054]

[0055] In the formula: Let w represent the estimated variance of the r-th feature. For good features with high local information retention capabilities, w ij The larger (x) ir -x jr ) 2 The smaller the variance, the better. Furthermore, features with larger variances are usually preferred because they are more representative. Therefore, in practical applications, the score of the r-th feature is usually calculated according to equation (4).

[0056]

[0057] This allows us to obtain the Laplace scores of features in a static dataset, thus revealing the clustering ability of those features within the dataset. As the signals in the dataset increase, the Laplace scores of the features should also be updated.

[0058] Now assume that at time t+1, the total number of acoustic emission signals increases to N. (t+1) Then it is necessary to modify graph G. (t) Update the graph to construct the nearest neighbor graph G at time t+1. (t+1) The key is to obtain the adjacency matrix W at time t+1. (t+1) Therefore, it is necessary to consider adding new signals. The connectivity between and with the original signal. If x l (l=N (t) +1,N (t) +2,...,N (t +1) ) is x m (m=1,2,...,N (t+1) If a given element is one of the k nearest neighbors of a given element, then the newly added adjacent element W is calculated. lm It can be assembled into N (t+1) Line N (t+1) sparse adjacency matrix of columns At the same time, the original adjacent elements to be replaced are assembled into a sparse adjacency matrix. If we consider the adjacency matrix W at time t...(t) Pad with zeros to N (t+1) Line N (t+1) Columns The adjacency matrix at time t+1 can be calculated according to equation (5).

[0059]

[0060] After obtaining the weight matrix, the Laplace matrix L at time t+1 can be calculated according to equation (1). (t+1) Sum degree matrix D (t+1) Then, the features at time t+1 are de-averaged, which is easy to verify. Therefore, the averaging process at time t+1 can be transformed into equation (6).

[0061]

[0062] The Laplace score of the r-th feature at time t+1 can be calculated according to equation (7).

[0063]

[0064] As can be seen from equations (6) and (7), the Laplace score of each acoustic emission feature at time t+1 can be updated based on the score calculation at time t.

[0065] The above formulas can be used to plot the Laplace score of each feature over time. Features with high Laplace scores across the entire time domain are potential clusters in the signal and are suitable for damage pattern recognition of composite materials based on acoustic emission response characteristics, but are rarely used for singular signal detection or damage degree estimation. Features with significantly decreased Laplace scores can provide relevant information for singular signal detection and can be used to prevent sudden failure of the sample. The larger the change in value, the more sensitive the feature is to the damage process and the more suitable it is as an indicator for detecting singular signals, which can be used to qualitatively describe the degree of damage. Features with insignificant decreases in Laplace scores contain information on damage pattern recognition, singular signal detection, and damage process characterization. They can be used to roughly characterize the degree of damage or to identify the damage category.

[0066] 3. Dynamic processing for multi-cluster feature selection

[0067] Multi-cluster feature selection methods achieve feature selection by choosing the most relevant subset of features. The dynamic feature processing method for acoustic emission data streams based on the multi-cluster feature selection algorithm is introduced below.

[0068] Let N be the time t. (t) N consists of acoustic emission impact signals (t) The ×p matrix is ​​X (t)Singular Value Decomposition (SVD) is performed on its transpose, as shown in Equation (8).

[0069] X (t)T =U (t) Σ (t) V (t)T (8)

[0070] In the formula: U (t) Let U be an orthogonal matrix of size p × p, satisfying U (t)T U (t) =I p , where I p Let Σ represent the p-order identity matrix. (t) For p×N (t) In a diagonal matrix, each element on the diagonal is a singular value. Singular values ​​are usually arranged in non-increasing order. All elements except the singular values ​​are 0. (t) For N (t) ×N (t) An orthogonal matrix that satisfies Given the large differences between singular values, the sum of the first 10% or even 1% of singular values ​​usually accounts for more than 90% of the sum of all singular values. Therefore, the first K largest singular values ​​can be used. To obtain X (t)T The low-rank approximation is shown in equation (9).

[0071]

[0072] In the formula: They are p×K, K×K, and N respectively. (t) A matrix of size ×K, a diagonal matrix It can be represented as

[0073] Multi-cluster feature selection methods are essentially about solving the problem of finding the most common feature among clusters. The LASSO regression problem with the target variable is expressed as Equation (10).

[0074]

[0075] Where: Φ (t) It can be represented as Its column vectors each contain the approximation of X. (t) X (t)T The combination coefficients of different features of the first K eigenvectors, any column vector can be represented as λ is a regularization parameter that controls the balance between the loss function and the 1-norm. This is achieved by obtaining the coefficient matrix Φ. (t) Then, for any sound emission characteristic f r(r=1,2,...,p), whose multi-cluster feature selection score can be calculated according to Equation (11).

[0076]

[0077] Thus, the multi-cluster feature selection method can be used to select feature subsets in static datasets. As the signals in the dataset increase, the feature subsets will also change accordingly.

[0078] Now assume that at time t+1, the total number of acoustic emission signals increases to N. (t+1) The number of new signals is N. (t+1) -N (t) And the matrix formed by the newly added impact signals is denoted as To avoid repeated input of acoustic emission data in the early stages, a matrix sketching method is introduced to effectively preserve X. (t) The low-rank approximation is then used, followed by regularized regression for feature selection. Ridge regression can replace LASSO regression when certain orthogonality conditions are met. Let X... (t) The sketch is of type B, which is l×p. (t) where l is less than N (t) And p, but greater than K. Considering that the singular value decomposition of the sketch matrix can be expressed as That is, its right singular vectors form the identity matrix, and thus the eigencoefficient matrix Φ (t+1) This can be obtained by minimizing the fitting error, as shown in equation (12).

[0079]

[0080] In the formula: e i Let B be a column vector with l elements, where the i-th element is 1 and the rest are 0. (0) The initial state is a zero matrix with all elements equal to 0. Therefore, we can assume that the sketch matrix at time t is known, and the sketch matrix B at time t+1 is... (t+1) and the eigenvalue matrix Φ (t+1) This can be obtained through an iterative process. First, let X... (t) The sketch matrix and the newly added signal matrix are assembled into a new matrix A. (t+1) As shown in equation (13).

[0081]

[0082] The singular value decomposition of the assembly matrix can be expressed as equation (14).

[0083]

[0084] Where: diagonal matrix It can be represented as By processing the diagonal elements of the diagonal matrix in equation (14), a new diagonal matrix is ​​obtained. As shown in equation (15).

[0085]

[0086] in: It is easy to see that the last diagonal element is 0. The sketch matrix B at time t+1. (t+1) It can be calculated according to formula (16).

[0087]

[0088] Finally, the diagonal matrix in equation (15) The first K singular values ​​are processed to obtain the diagonal matrix as shown in equation (17).

[0089]

[0090] In the formula: This leads to the eigencoefficient matrix at time t+1. The scores of each acoustic emission characteristic at time t+1 can be obtained by using formula (11).

[0091] The Laplace score selects important features containing rich information related to damage identification. Further filtering using multi-cluster feature selection can refine these features for optimal damage identification. Similarly, a multi-cluster feature selection score graph showing the changes of each feature over time is plotted, with the highest ranking score corresponding to the top-ranking feature in the relevant feature subset; more important features receive higher scores. The selection of damage preference representations is similar to that of the dynamic Laplace score: higher scores in the early stages are suitable for damage identification; significant changes before and after failure are suitable for anomaly signal detection; and minimal changes before and after failure are suitable for damage process characterization. If the early Laplace score is high, it can also be used for coarse damage identification.

[0092] 4. Dynamic evaluation of conventional acoustic emission characteristics

[0093] During acoustic emission (AE) testing, the initially recorded AE signal is a series of waveform envelope signals, which can be divided into one or more AE impact signals using the impact definition of the acquisition system. The impact signal can typically be represented as a combination of numerous conventional AE features extracted from the original waveform. The feature itself or its cumulative amount can be defined as a function of the load time or other test parameters (such as pressure and temperature) to describe the material damage. The damage characterization role of most features in real-time data streams is still unclear. Dynamic evaluation using conventional AE features serves two purposes: firstly, to verify the correctness of this invention based on existing research results; and secondly, to further clarify the damage characterization role of each conventional AE feature.

[0094] Fifteen conventional characteristics of acoustic emission impact signals were dynamically evaluated, including rise time (RT), count (C), energy (E), duration (TD), peak amplitude (PA), average frequency (AF), root mean square (RMS), average signal level (ASL), initial frequency (IF), signal strength (SS), absolute energy (AbE), centroid frequency (FG), peak frequency (PF), rise angle (RA), and decay angle (DA).

[0095] 5. Overview based on the aforementioned three processing or evaluation steps:

[0096] Real-time Laplace scores for 15 common features, as shown below Figure 3 As shown, PA, AF, RMS, ASL, IF, FG, and PF have relatively high Laplace scores. These seven conventional features show a gradual upward trend in scores as the number of signals increases, reaching almost 1 when considering all signals. Therefore, they are suitable for identifying potential clusters in signals. The remaining eight features exhibit high Laplace scores in the early stages, but their scores drop sharply before sample failure. Therefore, these eight features have low scores when considering all data and are unsuitable for identifying damage modes. On the other hand, these eight features indicating newly added signals are significantly different from previously accumulated signals, suggesting that the damage has reached a certain level and can be used for failure prevention.

[0097] The multi-cluster feature selection algorithm considers the correlation between different features ignored by the Laplace score. The real-time ranking scores of 15 acoustic emission features under the multi-cluster feature selection algorithm are as follows: Figure 4 As shown, most of the feature subsets selected in real time in the early stage basically contain 7 features: AF, RMS, ASL, IF, FG, PF, and PA. That is, the set of these 7 features contains rich information related to damage identification. Considering the set of 7 features containing rich information related to damage identification selected by the Laplace score, it is consistent with the Laplace score results.

[0098] For the eight features suitable for damage pattern recognition, their distribution matrix statistics are as follows: Figure 5As shown, the subplots on the diagonal are statistical histograms representing the distribution of the corresponding features, while the subplots off-diagonal represent the correlation between two corresponding features. The focus is on the concentrated distribution of signals within a certain interval of a feature, rather than the specific number of signals distributed within that interval. The diagonal subplots show that the signals exhibit regional clustering distributions across seven features: PF, AF, RMS, ASL, and IF. For example, the specific ranges for AF are 0-500kHz, 500-1000kHz, and above 2000kHz; the specific ranges for PF are 0-50kHz, 75-125kHz, and 125-200kHz; and the specific range for RMS varies slightly depending on the sample. Among these five features, the regional concentration of signals on PA and FG is not very obvious, therefore, PA and FG have relatively weak damage identification capabilities. This can also be seen from the diagonal portion, where the PA and FG of the impact signal approximate a semi-normal and a normal distribution, respectively. The boundaries of the categories in the space formed can be defined but are not very clear, thus the clustering ability of these two features is relatively weak. PF exhibits the most pronounced regionally concentrated distribution characteristic, thus it is often considered a reliable feature for damage pattern recognition. However, there are contradictions between the regional distribution of PF and other features, such as AF, RMS, and ASL. Therefore, damage pattern recognition is often based on a subset of features composed of multiple reliable features, rather than relying solely on a single feature.

[0099] For the eight features that are not suitable for damage pattern recognition, their distribution matrix statistics are as follows: Figure 6As shown in the diagram, the diagonal plots indicate that the values ​​of most impact signals on these eight features are fixed within a small range; no regional aggregation of signals was observed in the off-diagonal subplots either. Therefore, it is difficult to identify various damage mechanisms in the SLJ specimen based on the signal response values ​​on these eight features. Each off-diagonal subplot clearly contains several singular signals, which have very large values ​​on these features, clearly distinguishing them from the vast majority of signals. The results indicate that damage reaching a certain level will trigger singular signals. The threshold for abrupt changes in dynamic evaluation scores was set at 30%. Taking the Laplace score as an example, when the addition of a new signal causes a change in the overall score exceeding 30%, the new signal is considered a singular signal, and all found singular signals are compared with previous singular signals. Singular signals appear after the damage reaches a certain level, after which the specimen will completely fail shortly. Although RT, C, E, TD, SS, AbE, RA, and DA lack information related to damage pattern recognition, they can provide relevant information for singular signal detection and can be used to prevent sudden specimen failure. The Laplace scores for E, SS, and AbE decreased more significantly than those for RT, C, TD, RA, and DA, indicating that E, SS, and AbE are more sensitive to the damage process and are therefore more suitable as indicators for detecting singular signals. Unlike E, SS, and AbE, the Laplace scores for RT, C, TD, RA, and DA are higher before the boundary, suggesting that these five features can be used to roughly characterize the degree of damage as well as to identify the damage category, containing information on damage pattern recognition, singular signal detection, and damage process characterization.

[0100] In summary, different acoustic emission characteristics of FRP single-lap joint specimen damage signals exhibit different damage characterization preferences. Features such as PF, AF, RMS, ASL, IF, PA, and FG in the FRP acoustic emission data stream show relatively obvious regional distributions, making them more suitable for damage pattern recognition. Features such as E, SS, AbE, C, RT, TD, DA, and RA show relatively obvious clustered and outlier distributions. Among them, E, SS, and AbE are highly sensitive to critical damage and are suitable for singular signal detection, but not for damage process characterization. C, RT, TD, DA, and RA, on the other hand, demonstrate versatility: they can be used for singular signal detection, damage process characterization, and damage pattern recognition. These conclusions are consistent with existing research results and further clarify the potential roles of some acoustic emission characteristics.

[0101] For different acoustic emission characteristics, regardless of the combination of conventional parameters or through a series of process transformations, this invention can analyze their damage characterization preferences and determine whether they are suitable for singular signal detection, damage process characterization, or damage pattern recognition. Therefore, this invention can be used to discover new acoustic emission characteristics and explore their characterization preferences. Similarly, for different material samples, some acoustic emission characteristics will exhibit different damage characterization preferences. The characterization preferences of existing acoustic emission characteristics in different material samples can also be explored one by one using this invention.

[0102] Part Two: Examples of Application Methods of the Invention

[0103] After obtaining the dynamic evaluation results of conventional acoustic emission features, the dynamic evaluation of self-extracted features based on wavelet packet decomposition is taken as an example as a practical application of the present invention. Specifically, it is as follows:

[0104] 1. Feature extraction from wavelet packet decomposition

[0105] Unlike conventional acoustic emission features, wavelet packet decomposition features are novel features obtained by performing a series of transformations on the original acoustic emission impact signal, rather than simply combining the basic parameters of the original signal. However, due to their correlation with the original signal, the features obtained through wavelet packet decomposition have great potential for signal analysis and can be used for dynamic evaluation of features.

[0106] Taking the SLJ sample as an example, while extracting conventional acoustic emission features, the acoustic emission impact signal was also decomposed into wavelet packets, and the wavelet packet energy ratio and wavelet packet energy of each decomposed frequency band were extracted for dynamic evaluation.

[0107] Wavelet packets can be defined as a set of functions that satisfy the recursive relation of equation (18).

[0108]

[0109] In the formula: φ(t) and ψ(t) represent the scaling function and wavelet function, respectively; n is the frequency parameter; k is the time-domain factor; h(k) and g(k) are the coefficients of the low-pass filter and high-pass filter in the multi-resolution analysis, respectively. These two coefficients satisfy an orthogonal relationship, i.e., g(k) = (-1). k h(1-k).

[0110] If n is the subdivision octave parameter, i.e., n = 2 l +m, then the wavelet packet can be simplified as a function as shown in equation (19).

[0111] ψ j,k,n (t)=2 -j / 2 ψ n (t·2 -j -k) (19)

[0112] In the formula: The integers m and l take values ​​from 0 to 2. l -1 and 1 to j.

[0113] The wavelet packet coefficients of the input waveform x(t) can be calculated according to equation (20).

[0114]

[0115] Therefore, each component of the signal can be reconstructed according to equation (21).

[0116]

[0117] Using the minimum Shannon entropy as the wavelet packet basis, the energy ratio of the reconstructed signal in the nth frequency range of the jth level can be calculated according to equation (22).

[0118]

[0119] In the formula: E j,n (t) represents the wavelet packet energy of the signal in the nth frequency range of the j-th level, x j,n (t) represents the reconstructed signal in the nth frequency range of the j-th level.

[0120] Considering the sampling frequency is 1MHz, the Daubechies wavelet basis function (db5) with a vanishing moment of 5 is used to decompose each acoustic emission signal to the third layer, that is, to decompose it into 8 components, each with a frequency range of approximately 50kHz. Therefore, each impact signal can be regarded as a vector composed of the energy proportions of 8 frequency bands (denoted as F1-F8).

[0121] 2. Dynamic evaluation of self-extracted features

[0122] The wavelet packet energy ratio in frequency bands F1-F8 is dynamically evaluated following the steps described above. Evaluation begins when the accumulated signal reaches 1000, and the result is updated every additional 100 signals thereafter. (See attached image.) Figure 7As shown, for any given sample, the Laplace score of the wavelet packet energy ratio in each frequency band is generally greater than 0.9, and the scores of these eight energy ratios show a continuous upward trend as the number of signals increases. Even after the load-displacement curve shows a second zigzag bend, the wavelet packet energy ratio score does not exhibit a sharp drop. This indicates that the wavelet packet energy ratios of the newly added signal and the original signal in frequency bands F1-F8 remain similar, thus maintaining good clustering ability unaffected by the occurrence and evolution of sample damage. Therefore, the energy ratio combination of the signal after wavelet packet decomposition in each frequency band is highly suitable for classifying acoustic emission impact signals. Simultaneously, the wavelet packet energy ratios do not show sensitivity to the damage process, indicating that they are not suitable for detecting singular signals and characterizing the damage process. Since all eight energy ratios are important features and their definitions already avoid correlation, a subset containing all eight features is not selected based on the multi-cluster feature selection method.

[0123] Besides the wavelet packet energy ratio, the wavelet packet decomposition features also include the wavelet packet energy in frequency bands F1-F8. The cumulative wavelet packet energy in each frequency band, i.e., the integral of the square of the signal voltage in each decomposed frequency band over time from 0 to the present divided by the resistance, shows a phased upward trend, with the cumulative wavelet packet energy in frequency bands F1, F2, and F4 being the most significant. The large jump in the cumulative wavelet packet energy in frequency bands F1, F2, and F4 obscures the trend of energy accumulation in frequency bands F3, F5, F6, F7, and F8, resulting in a sharp change in both scores. Therefore, when the damage reaches a certain level, singular signals will be generated, but not all wavelet packet decomposition features of singular signals are numerically much greater than the decomposition results of previous ordinary signals, thus making it suitable for roughly characterizing the degree of damage and detecting singular signals. Since the wavelet packet energy in the main frequency bands of each damage mode is numerically higher than that in other frequency bands, it can be inferred that frequency bands F1, F2, and F4 are the main frequency bands of the damage modes generated in the SLJ sample.

[0124] In summary, the wavelet packet energy of each damage signal in the main frequency band can play a damage characterization role similar to acoustic emission energy, signal strength, and absolute energy.

Claims

1. A method for evaluating dynamic characteristics of an acoustic emission data stream of a composite adhesive specimen, characterized by, The method comprises: acquisition and processing of acoustic emission signal data, dynamic processing of Laplace score, dynamic processing of multi-cluster feature selection, and dynamic evaluation of conventional acoustic emission features; analysis of the damage characterization role preference of acoustic emission features to obtain an evaluation result of whether the acoustic emission features are applicable to damage mode identification, singular signal detection, or damage process characterization; The acquisition and processing of acoustic emission signal data specifically comprises the following steps: (1) A composite material bonding sample is prepared according to a conventional method, and a uniaxial tensile test is performed on the sample to obtain acoustic emission signal data; (2) The obtained acoustic emission signal data is processed as follows: Set the features to be collected: acoustic emission signal is constantly generated in time dimension, while the total number of features remains unchanged in feature dimension; suppose that the system records signals at time t, each signal is represented as a p-dimensional vector, i.e. , ; the entire acoustic emission signal data is converted into p-dimensional vectors as the basic data content and form for subsequent two dynamic processing The dynamic processing of Laplace score specifically comprises the following steps: Based on the Laplace score algorithm for feature dynamic processing, a nearest neighbor graph with a vertex number of at time t is constructed , wherein each vertex represents an acoustic emission data point; the Laplace matrix at time t is calculated, and each feature is processed by deaveraging; the Laplace score of the rth feature at time t is calculated to obtain the Laplace score of the feature in the static data set; at time t+1, the graph is updated to construct a nearest neighbor graph at time t+1, and the adjacency matrix at time t+1 is calculated; after obtaining the weight matrix, the Laplace matrix and the degree matrix at time t+1 are calculated; then, the deaveraging processing is performed on each feature at time t+1 to obtain the Laplace score of each feature at time t+1; The dynamic processing of multi-cluster feature selection specifically comprises the following steps: Record time t Composed of acoustic emission impact signals The matrix is Singular value decomposition is performed on its transpose to solve for... For the LASSO regression problem of the target variable; after obtaining the coefficient matrix Subsequently, for any sound emission characteristic Calculate multi-cluster feature selection scores for use in selecting feature subsets from a static dataset. At time t+1, to avoid repeated input of previous acoustic emission data, a matrix sketching method is introduced to effectively preserve... The low-rank approximation is then used; regularized regression is then used to perform feature selection, and the feature coefficient matrix at time t+1 is obtained. The scores of each acoustic emission feature at time t+1 are then calculated.

2. The method of claim 1, wherein, In the dynamic evaluation of conventional acoustic emission features, the acoustic emission conventional features at least include: rise time RT, count C, energy E, duration TD, peak amplitude PA, average frequency AF, root mean square RMS, average signal level ASL, initial frequency IF, signal strength SS, absolute energy AbE, center of gravity frequency FG, peak frequency PF, rise angle RA, and decay angle DA.

3. The method of claim 1, wherein, The analysis of the damage characterization role preference of acoustic emission features and the evaluation result are specifically as follows: The acoustic emission features of acoustic emission impact signals are dynamically evaluated, different acoustic emission features of sample damage signals show different damage characterization preferences, and an evaluation result of whether they are applicable to singular signal detection, damage process characterization, or damage mode identification is obtained; in subsequent acoustic emission experiments, adaptive selection is made according to the evaluation result.

4. Application method of the evaluation result obtained by the dynamic feature evaluation method of the acoustic emission data stream of the composite material bonding sample according to claim 1, comprising the following steps: (1) Wavelet packet energy proportion and wavelet packet energy of acoustic emission impact signals are obtained by wavelet packet decomposition; (2) The dynamic evaluation of self-extraction features based on wavelet packet decomposition according to the steps of claim 1 is performed to obtain a result showing that the wavelet packet energy proportion is applicable to the classification of acoustic emission impact signals, but not applicable to the detection of singular signals and the characterization of damage processes; the wavelet packet energy can play a damage characterization role similar to acoustic emission energy, signal strength, and absolute energy.

Images