An incremental learning method for quantitative detection of crack propagation based on stress wave signals

Abstract

This invention discloses an incremental learning method for quantitative detection of crack propagation based on stress wave signals, belonging to the field of nondestructive testing technology. The method includes: acquiring stress wave signals and extracting frequency domain features; constructing and training a BiLSTM model based on a 0-5mm crack sample dataset, extracting high-dimensional feature vectors as initial features for crack detection; mapping the high-dimensional feature vectors to a two-dimensional space to quantitatively evaluate the distribution differences of features among different crack samples; dynamically adjusting the retention ratio of historical data according to the crack sample feature distribution and model detection accuracy; introducing a chaotic sampling strategy and combining it with an entropy domain fusion index to select samples for empirical replay; introducing 6-10mm crack propagation data and updating the BiLSTM model using incremental learning; and using the final detection model to quantitatively evaluate the degree of crack propagation. Through real-time crack monitoring, high-precision detection based on multi-frequency domain features, and a phased incremental training strategy, the detection accuracy and generalization ability of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to an incremental learning method for quantitative detection of crack propagation based on stress wave signals. Background Technology

[0002] Crack propagation monitoring and assessment are crucial in aerospace, bridge construction, and industrial equipment. Traditional crack detection methods mainly rely on visual inspection, ultrasonic testing, or electromagnetic testing, but these methods are often limited by human experience, complex operating conditions, and high testing costs. In recent years, stress wave signal-based crack propagation monitoring technology has gradually become a research hotspot. This method monitors changes in stress wave signals inside the structure in real time using sensors, thereby inferring crack propagation, and has the advantages of high sensitivity and non-contact detection.

[0003] Crack propagation typically exhibits complex nonlinear characteristics, and the signal patterns differ significantly across different crack stages, making it difficult for traditional machine learning models to generalize effectively. Furthermore, deep learning models with fixed structures tend to forget prior knowledge when introducing new crack data, affecting detection stability.

[0004] Deep learning technology, especially Bidirectional Long Short-Term Memory (BiLSTM) networks, offers a new solution for crack propagation monitoring due to its advantages in time-series data modeling. In practical applications, crack propagation data accumulates over time, and traditional statically trained models struggle to adapt to the changing characteristics of new data, easily leading to performance degradation. Therefore, employing an incremental learning strategy, enabling the model to retain old knowledge while introducing new data, is crucial for improving crack detection accuracy and generalization ability. Summary of the Invention

[0005] Based on the shortcomings of the prior art described above, the purpose of this invention is to provide an incremental learning method for quantitative detection of crack propagation based on stress wave signals, so as to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an incremental learning method for quantitative detection of crack propagation based on stress wave signals, comprising:

[0007] The stress wave signal of the structure during crack propagation is acquired in real time by sensors, and the stress wave signal is preprocessed to extract frequency domain features.

[0008] A bidirectional long short-term memory (BiLSTM) model was constructed and trained based on a 0-5mm crack sample dataset. High-dimensional feature vectors were extracted through fully connected layers as the initial feature representation for crack detection.

[0009] By mapping high-dimensional feature vectors to a two-dimensional space, the distribution differences of features of different crack samples can be quantitatively evaluated.

[0010] The retention ratio of historical data is dynamically adjusted based on the characteristic distribution of crack samples and the detection accuracy of the model.

[0011] During the dynamic adjustment process, a chaotic sampling strategy is introduced, and samples with high uncertainty are selected for empirical replay in combination with the entropy domain fusion index.

[0012] We introduced 6-10mm crack propagation data and used incremental learning to update the BiLSTM model;

[0013] The final detection model after training is used to quantitatively assess the degree of crack propagation and output the crack propagation trend prediction results.

[0014] The invention is further configured to monitor stress waves during crack propagation by a sensor mounted on the structure, connect the sensor and a signal acquisition system, and use a data acquisition instrument to acquire stress wave signals; the sensor includes a piezoelectric sensor, a vibration sensor and an accelerometer.

[0015] The present invention further specifies that the frequency domain features include spectral mixing entropy, resonant transition rate, modal discrepancy factor, frequency cross-coupling index, and frequency shift skewness index;

[0016] The calculation logic for the spectral mixing entropy is as follows: Where SME is the spectral mixing entropy, X[k] is the complex spectral component after FFT transformation, k is the index of the frequency component, α and β are adjustment factors, and N is the number of sampling points;

[0017] The calculation logic for the resonant transition rate is as follows: Where RTR is the resonant transition rate, f k Let be the kth frequency component, γ be the control factor, and δ be a small constant;

[0018] The calculation logic for the modal discrepancy factor is as follows: Where MDF is the modal discretization factor, f c λ is the dominant frequency of the signal, and λ is the control factor.

[0019] The calculation logic for the frequency cross-coupling index is as follows: Where SFC is the frequency cross-coupling index, ∈ is a small constant, and ζ and η are adjustment factors;

[0020] The calculation logic for the frequency shift skewness index is as follows: Where FSSI is the frequency shift skewness index and ω is the adjustment factor.

[0021] The present invention further specifies that the step of constructing the BiLSTM model includes:

[0022] The frequency domain feature data extracted from 0-5mm crack samples were standardized and then used as input.

[0023] A bidirectional long short-term memory network BiLSTM with forward and backward propagation paths was constructed, and the number of network layers, the number of hidden units, and the activation function parameters were set.

[0024] The basic BiLSTM network is iteratively trained using the preprocessed training dataset until convergence, and a fully connected layer is introduced before the network output layer to extract high-dimensional feature vectors.

[0025] The high-dimensional feature vectors output by the fully connected layer are used as inputs for subsequent dimensionality reduction analysis and feature distribution evaluation.

[0026] The present invention is further configured to perform dimensionality reduction on the high-dimensional feature vector extracted from the fully connected layer of the BiLSTM model using a difference metric function, thereby mapping the high-dimensional feature vector to a two-dimensional space. The calculation logic of the difference metric function is as follows: Where W is the difference measure, P ij For sample x i and x j Similarity in high-dimensional space, Q ij For the data points y after dimensionality reduction i and y j Similarity between P represents the ratio of two distributions. ij The calculation logic is as follows: Where, ||x i -x j ||For sample x i and x j The distance between them, σ 2 Q is the similarity scale for each pair of points in a high-dimensional space. ij The calculation logic is as follows: Among them, y i and y j This is the lower-dimensional representation after dimensionality reduction;

[0027] The distribution difference function is used to measure the distribution difference of sample features reduced to two-dimensional space. The calculation logic of the distribution difference function is as follows: Where C(P||Q) represents the difference between the two distributions, and P(x) represents the difference between the two distributions. i ) represents the probability distribution of high-dimensional data, Q(x) i ) represents the probability distribution of the dimensionality-reduced low-dimensional data, x i Let be a point in the sample space.

[0028] The present invention is further configured such that the dynamically adjusted calculation logic is: Rt =max(R) min ,min(R t-1 ·exp(-λ1H(S t D t )+λ2ΔA(t)),R max ), where R t Let H(S) be the percentage of historical data retained at time t. t D t ) represents the difference in the characteristic distribution of crack samples, ΔA(t) represents the detection accuracy, and R min and R max For adjustable parameters, R t-1 The retention ratio at time t, λ1 and λ2 are adjustment parameters, and the difference in crack sample characteristic distribution H(S) is the ratio of samples retained at time t. t D t The calculation logic for ) is as follows: Among them, P S For the current sample S t probability distribution, P D For historical data D t The probability distribution of , υ is the joint probability distribution, Π(P) S ,P D ) is a joint distribution, d(x) i ,x j (x) represents the data point. i and x j The calculation logic for the detection accuracy ΔA(t) between the distances is as follows: Among them, ∈1 is a local minimum.

[0029] The present invention is further configured to combine the entropy domain fusion index in the chaotic sampling strategy, and to select samples with high uncertainty for experience replay by optimizing the objective function to screen the entropy domain fusion index.

[0030] The formula for calculating the entropy domain fusion index is as follows: Where H(x) is the entropy domain fusion index, λ3 is the model uncertainty factor, ψ(x) is the sample importance weight, β(x) is the influence of the sample on the model decision boundary, and the calculation logic of ψ(x) is as follows: Where d(x,x) j ) represents the feature distance between samples, γ1 is a hyperparameter controlling the range of local density influence, and the calculation logic of β(x) is as follows: Where D(x) is the distance from the sample to the decision boundary, and α1 is the control influence range parameter;

[0031] The calculation logic for the optimization objective function is as follows: Where T(x) is the objective function, μ is the density adjustment coefficient, δ1 is the density influence control parameter, η1 is the boundary sample suppression coefficient, and κ is the decision boundary influence adjustment coefficient.

[0032] The present invention is further configured to introduce 6-10mm crack propagation data and use incremental learning to update the BiLSTM model. The specific implementation steps include:

[0033] Preprocessing and frequency domain feature extraction were performed on the 6-10mm crack propagation data, and the newly generated data were made consistent with the old data in terms of feature dimension, format and distribution.

[0034] A phased incremental learning strategy was adopted, and 6-10mm crack data samples were mixed with 0-5mm old data samples according to a preset ratio to construct a multi-stage training dataset.

[0035] Freeze some layers of the BiLSTM model structure, while unfreezing the upper network layers to adapt to the feature distribution of new data, and iterate and update the network parameters repeatedly after each batch of new data is introduced.

[0036] After each stage of training is completed, when the model performance reaches a preset threshold, the incremental learning process is terminated, and the final detection model is obtained.

[0037] The present invention is further configured such that, in the process of freezing some layers of the BiLSTM model structure, dynamic adjustments are made through layer freeze scoring, and the calculation logic for the layer freeze scoring is as follows: LFS stands for Layer Freeze Score. The loss L(X) of the old data in the l-th layer old The gradient change of ), p is the nonlinear weighting exponent of the gradient, γ2 is the adjustment parameter, and ΔS t,i The change in neuron activation value at the i-th time step is the entropy. Integrating the parameter over the range [a,b], ω j For the globally optimal parameter set, λ j and β j For weight adjustment factor, ψ j To adjust the exponential decay term, κ j For exponential decay rate, This is the optimal parameter point for the current layer.

[0038] The present invention is further configured to quantitatively assess the degree of crack propagation, analyze historical data of crack propagation using a BiLSTM model, and perform time series regression in combination with the latest crack characteristics to predict future crack propagation trends, plot crack propagation trend curves, and display the predicted results of crack propagation degree changing over time through a visualization interface.

[0039] This invention provides an incremental learning method for quantitative crack propagation detection based on stress wave signals. It involves real-time acquisition of stress wave signals from a structure during crack propagation using sensors, preprocessing the stress wave signals to extract frequency domain features, constructing and training a bidirectional long short-term memory (BiLSTM) model based on a 0-5mm crack sample dataset, and extracting high-dimensional feature vectors through fully connected layers as the initial feature representation for crack detection. The high-dimensional feature vectors are mapped to a two-dimensional space to quantitatively evaluate the distribution differences of features among different crack samples. The retention ratio of historical data is dynamically adjusted according to the crack sample feature distribution and model detection accuracy. During this dynamic adjustment, a chaotic sampling strategy is introduced, combined with an entropy fusion index to select samples with high uncertainty for empirical replay. 6-10mm crack propagation data is introduced, and incremental learning is used to update the BiLSTM model. The final trained detection model is used to quantitatively evaluate the degree of crack propagation and output crack propagation trend prediction results. The beneficial effects include:

[0040] 1. Real-time crack propagation monitoring and quantitative assessment: By acquiring stress wave signals in real time through sensors and combining the deep learning capabilities of the BiLSTM model, dynamic tracking of the crack propagation process can be achieved, ensuring high-precision quantitative assessment and trend prediction of the degree of crack propagation.

[0041] 2. Improve crack detection accuracy: By introducing multiple frequency domain features based on stress wave signals, including spectral mixing entropy, resonant transition rate, modal discrepancy factor, frequency cross-coupling index, and frequency shift skewness index, the influence of crack propagation on stress wave signals can be more comprehensively characterized, thereby improving the sensitivity and robustness of detection.

[0042] 3. Phased incremental training to improve model stability: A phased incremental learning strategy is adopted to gradually introduce new crack expansion data and dynamically adjust the layer freezing state in combination with layer freezing score. This ensures that the model maintains the stability of old knowledge while learning new knowledge, thereby improving the stability and convergence speed of model training.

[0043] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0045] Figure 1 The flowchart illustrates an incremental learning method for quantitative detection of crack propagation based on stress wave signals, as an exemplary embodiment of the present invention.

[0046] Figure 2 Images of stress waves in the time and frequency domains for cracks of different lengths;

[0047] Figure 3 This is a performance demonstration graph for different sampling algorithms in incremental learning. Detailed Implementation

[0048] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0049] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0050] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0051] An incremental learning method for quantitative detection of crack propagation based on stress wave signals, such as Figure 1 As shown, it includes:

[0052] The stress wave signal of the structure during crack propagation is acquired in real time by sensors, and the stress wave signal is preprocessed to extract frequency domain features.

[0053] A bidirectional long short-term memory (BiLSTM) model was constructed and trained based on a 0-5mm crack sample dataset. High-dimensional feature vectors were extracted through fully connected layers as the initial feature representation for crack detection.

[0054] By mapping high-dimensional feature vectors to a two-dimensional space, the distribution differences of features of different crack samples can be quantitatively evaluated.

[0055] The retention ratio of historical data is dynamically adjusted based on the characteristic distribution of crack samples and the detection accuracy of the model.

[0056] During the dynamic adjustment process, a chaotic sampling strategy is introduced, and samples with high uncertainty are selected for empirical replay in combination with the entropy domain fusion index.

[0057] We introduced 6-10mm crack propagation data and used incremental learning to update the BiLSTM model;

[0058] The final detection model after training is used to quantitatively assess the degree of crack propagation and output the crack propagation trend prediction results.

[0059] The invention is further configured to monitor stress waves during crack propagation by a sensor mounted on the structure, connect the sensor and a signal acquisition system, and use a data acquisition instrument to acquire stress wave signals; the sensor includes a piezoelectric sensor, a vibration sensor and an accelerometer.

[0060] The present invention further specifies that the frequency domain features include spectral mixing entropy, resonant transition rate, modal discrepancy factor, frequency cross-coupling index, and frequency shift skewness index; the calculation logic for spectral mixing entropy is as follows: Where SME is the spectral mixing entropy, X[k] is the complex spectral component after FFT transformation, k is the index of the frequency component, α and β are adjustment factors, and N is the number of sampling points; the calculation logic of the resonant transition rate is as follows: Where RTR is the resonant transition rate, f k Let γ be the k-th frequency component, γ be the control factor, and δ be a small constant; the calculation logic for the modal discretization factor is as follows: Where MDF is the modal discretization factor, f c Let λ be the dominant frequency of the signal and λ be the control factor; the calculation logic for the frequency cross-coupling index is as follows: Where SFC is the frequency cross-coupling index, ∈ is a small constant, and ζ and η are adjustment factors; the calculation logic of the frequency shift skewness index is as follows: Wherein, FSSI is the frequency shift skewness index, and ω is the adjustment factor; specifically, spectral mixing entropy (SME) measures the complexity and uncertainty of the signal's spectral distribution, reflects the energy distribution characteristics of the signal spectrum during crack propagation, and can effectively characterize the energy distribution changes caused by crack propagation; α is used to adjust the weight of the spectral energy distribution, with a value range of [0.5, 2]; β is used to adjust the nonlinear adjustment of the spectral energy, with a value range of [0.5, 2]; resonant transition rate (RTR) measures the rate of energy change of the signal spectrum at different frequency components, reflects the transition behavior of the signal between different resonant points, and helps to detect the impact of crack propagation on structural vibration signals; γ is used to adjust the weight of RTR calculation in different frequency ranges, making the features more sensitive to the spectral changes caused by cracks, with a value range of [0.5, 2]; δ is used to prevent the denominator from being zero and to ensure the stability of the calculation, with a value range of

[10] . -6 10 -3 The modal dispersion factor (MDF) measures the dispersion of signal energy distribution in the spectrum, reflecting the impact of crack propagation on the vibration mode distribution; λ is used to adjust the weight of energy at different frequencies, with a value range of [0.5, 2]; the frequency cross-coupling index (SFC) measures the degree of energy coupling between different frequency components, reflecting the impact of crack propagation on the spectral structure; ∈ is used to prevent the denominator from being zero, with a value range of

[10] . -6 10 -3 ]; ζ is used to control the degree of nonlinear coupling between adjacent frequency components, with a value range of [1,5]; η is used to control the exponential decay rate of high-frequency components, with a value range of [0.01,0.5]; the frequency shift skewness index FSSI is used to measure whether the energy center of the signal spectrum has shifted, reflecting the impact of crack propagation on the frequency characteristics of the signal; ω is used to control the weight of high-frequency components, with a value range of [0.5,3]; the crack propagation characteristics are described from multiple perspectives such as energy distribution, frequency transition, modal change, nonlinear coupling, and frequency shift skewness, avoiding misjudgments that may be caused by a single feature.

[0061] The present invention further specifies that the step of constructing the BiLSTM model includes:

[0062] Frequency domain feature data extracted from 0-5mm crack samples were used as input after format standardization. Specifically, the standardization process included normalization and dimension alignment. Normalization mapped the signal amplitude and scale data of different crack samples to the same range, eliminating the impact of numerical differences on model training. Dimension alignment ensured that the number of features in the data samples was consistent, adapting to the model input. The preprocessing results of the collected 0-5mm crack samples are shown below. Figure 2 As shown;

[0063] A bidirectional long short-term memory network (BiLSTM) with forward and backward propagation paths is constructed, and the number of network layers, the number of hidden units, and the activation function parameters are set. Specifically, a BiLSTM model with forward and backward propagation paths is built to learn the temporal features of crack samples. Sigmoid and Tanh are used as activation functions inside the LSTM units, and the mean squared error is used to calculate the loss. This is a technique that will not be elaborated on further.

[0064] The basic BiLSTM network is iteratively trained using the preprocessed training dataset until convergence. A fully connected layer is introduced before the network output layer to extract high-dimensional feature vectors. Specifically, the BiLSTM network is trained using 0-5mm crack samples to enable the model to accurately learn crack features. The dataset is split into 80% training set and 20% test set, and iterative training is performed until the loss is below 0.01. The training loss is observed to ensure stable convergence of the model. The output dimension of the BiLSTM network is passed to the fully connected layer, and after matrix multiplication, the output is a new high-dimensional feature vector for each time step. The output dimension of the BiLSTM network is related to the number of hidden units and the length of the input sequence.

[0065] The high-dimensional feature vectors output by the fully connected layer are used as input for subsequent dimensionality reduction analysis and feature distribution evaluation. Specifically, the extracted high-dimensional feature vectors are used to view the clustering structure of the data and observe the feature space distribution of different zero-zero crack data.

[0066] The present invention is further configured to perform dimensionality reduction on the high-dimensional feature vector extracted from the fully connected layer of the BiLSTM model using a difference metric function, thereby mapping the high-dimensional feature vector to a two-dimensional space. The calculation logic of the difference metric function is as follows: Where W is the difference measure, P ij For sample x i and x j Similarity in high-dimensional space, Q ij For the data points y after dimensionality reduction i and y j Similarity between P represents the ratio of two distributions. ij The calculation logic is as follows: Where, ||x i -x j ||For sample x i and x j The distance between them, σ 2 Q is the similarity scale for each pair of points in a high-dimensional space. ij The calculation logic is as follows: Among them, y i and y jThis represents the low-dimensional representation after dimensionality reduction. The distribution difference function is used to measure the distribution difference of the sample features in the reduced two-dimensional space. The calculation logic of the distribution difference function is as follows: Where C(P||Q) represents the difference between the two distributions, and P(x) represents the difference between the two distributions. i ) represents the probability distribution of high-dimensional data, Q(x) i ) represents the probability distribution of the dimensionality-reduced low-dimensional data, x i Let σ be a point in the sample space. Specifically, the difference metric W measures the difference between the high-dimensional and low-dimensional data distributions. The smaller the W value, the closer the dimensionality-reduced data distribution is to the original data distribution. The distribution difference function C(P||Q) measures the difference between the sample feature distributions in the high-dimensional space and the dimensionality-reduced low-dimensional space. It quantifies the difference between the distributions in the high-dimensional space and the dimensionality-reduced low-dimensional space by comparing the differences between the two probability distributions. 2 The local scale parameter in the high-dimensional space controls the scale size during similarity calculation, with a value range of [0,10]. The low-dimensional features obtained after dimensionality reduction are used as input to improve the accuracy of crack propagation stage identification. By calculating the difference in probability distribution between high-dimensional and low-dimensional data in the sample space, important feedback on the dimensionality reduction effect can be provided, which helps to optimize the performance of the dimensionality reduction algorithm and ensure that key information of the data is retained after dimensionality reduction.

[0067] The present invention is further configured such that the dynamically adjusted calculation logic is: R t =max(R) min ,min(R t-1 ·exp(-λ1H(S t D t )+λ2ΔA(t)),R max ), where R t Let H(S) be the percentage of historical data retained at time t. t D t ) represents the difference in the characteristic distribution of crack samples, ΔA(t) represents the detection accuracy, and R min and R max For adjustable parameters, R t-1 The retention ratio at time t, λ1 and λ2 are adjustment parameters, and the difference in crack sample characteristic distribution H(S) is the ratio of samples retained at time t. t D t The calculation logic for ) is as follows: Among them, P S For the current sample S t probability distribution, P D For historical data D t The probability distribution of , υ is the joint probability distribution, Π(P) S ,P D ) is a joint distribution, d(x) i,x j (x) represents the data point. i and x j The calculation logic for the detection accuracy ΔA(t) between the distances is as follows: Where ∈1 is a local minimum; specifically, the main purpose of the dynamically adjusted calculation logic is to dynamically adjust the data retention ratio R based on the real-time changes of the crack detection model, such as differences in crack sample feature distribution and detection accuracy. t Retention ratio R t This is achieved by maximizing the retention of historical data while avoiding exceeding preset upper and lower limits R. min and R max To ensure the effectiveness of data storage and computation; crack sample feature distribution differences H(S) t D t The difference between the current crack sample characteristics and historical data is reflected in the following values: A larger difference indicates a greater discrepancy between the current sample and historical data, leading to the retention of more historical data; conversely, a smaller difference indicates a reduced retention rate. The detection accuracy ΔA(t) measures the change in detection accuracy from the previous time point. A higher ΔA(t) value indicates a significant improvement in the accuracy of the current model in crack detection, requiring a reduction in the retention rate of historical data to avoid data redundancy. R min The minimum value used to control the retention ratio, with a range of [0,1]; R max The maximum value of the retention ratio is used to control the retention ratio, with a value range of [0,1]; λ1 is used to control the influence of feature distribution differences on the retention ratio, with a value range of [0,1]; λ2 is used to control the influence of detection accuracy on the retention ratio, with a value range of [0,1]; ∈1 is used to prevent division by zero errors, with a value range of

[10] . -7 10 -5 By considering differences in feature distribution and detection accuracy, the model can dynamically adjust the retention ratio of historical data based on real-time data, thereby avoiding data redundancy and improving computational efficiency.

[0068] The present invention is further configured to incorporate an entropy fusion index into the chaotic sampling strategy, and to prioritize the selection of samples with high uncertainty for empirical replay by optimizing the objective function to achieve the entropy fusion index; the formula for calculating the entropy fusion index is as follows: Where H(x) is the entropy domain fusion index, λ3 is the model uncertainty factor, ψ(x) is the sample importance weight, β(x) is the influence of the sample on the model decision boundary, and the calculation logic of ψ(x) is as follows: Where d(x,x) j ) represents the feature distance between samples, γ1 is a hyperparameter controlling the range of local density influence, and the calculation logic of β(x) is as follows: Where D(x) is the distance from the sample to the decision boundary, and α1 is the control influence range parameter; the calculation logic of the optimization objective function is as follows: Where T(x) is the objective function, μ is the density adjustment coefficient, δ1 is the density influence control parameter, η1 is the boundary sample suppression coefficient, and κ is the decision boundary influence adjustment coefficient; specifically, the entropy domain fusion index H(x) is used to measure the measurement method of sample uncertainty, local importance, and their influence on the model decision boundary; the sample importance weight term ψ(x) is used to measure the relative importance of the sample in the entire data distribution; the sample influence term β(x) is used to measure the degree of influence of the sample on the model decision boundary; λ3 is used to adjust the influence of sample uncertainty on entropy, with a value range of [0.5, 2]; γ1 is used to control the local density influence range, with a value range of [0.5, 2]. The range is [0.1, 10]; α1 is used to control the range of the influence of the sample on the decision boundary, with a value range of [0.1, 5]; the value range is [0, 1]; the optimization objective function T(x) is used to measure the sample, with a value range of [0.1, 10]; η1 is used to determine the overall influence of the boundary sample, with a value range of [0.1, 5]; κ is used to control the influence of the boundary sample, with a value range of [0.1, 10]; samples with high uncertainty are screened by entropy domain fusion index, and combined with the optimization objective function, samples with larger information content are selected first for experience replay, so that the model focuses on more challenging and representative samples, avoids learning redundant information, and improves data utilization.

[0069] The present invention is further configured to introduce 6-10mm crack propagation data and use incremental learning to update the BiLSTM model. The specific implementation steps include:

[0070] Preprocessing and frequency domain feature extraction were performed on the 6-10mm crack propagation data, and the newly generated data were made consistent with the old data in terms of feature dimension, format and distribution.

[0071] A phased incremental learning strategy is adopted, in which 6-10mm crack data samples are mixed with 0-5mm old data samples at a preset ratio to construct a multi-stage training dataset. Specifically, the model is trained with 80% old data and 20% new data at the preset ratio, and then the proportion of new data is gradually increased, such as 60% old data and 40% new data, until it is fully adapted to 6-10mm crack data. The 6-10mm crack data is added in batches, such as 10% in the first stage, 30% in the second stage, until all of them are introduced, to reduce the pressure of sudden adaptation of the model. Phased learning can ensure that the model gradually adapts to the new data, while training directly with new data may cause the model to forget the features of the old data.

[0072] The BiLSTM model structure is partially frozen while the upper network layers are unfrozen to adapt to the feature distribution of the new data. The network parameters are iteratively updated after each batch of new data is introduced. Specifically, the BiLSTM model is first trained with existing 0-5mm data to obtain the basic model. New data of 6-10mm is introduced and trained in a phased learning manner. After each phase of training, the model parameters are updated to make the model gradually adapt to the feature distribution of the new data. By dynamically adjusting the layer freezing strategy, more layers can be unfrozen when the model is stable, so that the entire model can better adapt to the new data.

[0073] After each training stage is completed, the incremental learning process is terminated when the model performance reaches a preset threshold, and the final detection model is obtained. Specifically, after each training stage, the mean squared error, accuracy, and other indicators of the validation set are calculated, and the model convergence is observed. When the model performance reaches a preset threshold, such as when the detection accuracy is less than 95% or the error between the model's predicted value and the true value is less than 0.01, incremental training is stopped, the final detection model is obtained, and finally, the model is validated on an independent test set to improve the model's generalization ability.

[0074] The present invention is further configured such that, in the process of freezing some layers of the BiLSTM model structure, dynamic adjustments are made through layer freeze scoring, and the calculation logic for the layer freeze scoring is as follows: LFS stands for Layer Freeze Score. The loss L(X) of the old data in the l-th layer old The gradient change of ), p is the nonlinear weighting exponent of the gradient, γ2 is the adjustment parameter, and ΔS t,i The change in neuron activation value at the i-th time step is the entropy. Integrating the parameter over the range [a,b], ω j For the globally optimal parameter set, λ j and β j For weight adjustment factor, ψ j To adjust the exponential decay term, κ j For exponential decay rate, This represents the optimal parameter point for the current layer. Specifically, the Layer Freeze Score (LFS) is used to measure whether a certain layer should be frozen. A threshold is determined based on the LFS score. When the LFS score is greater than the preset LFS threshold, the corresponding BiLSTM model structure layer calculated by LFS is frozen. p controls the contribution of the gradient, with a value range of [1,3]. γ2 controls the impact of entropy changes on the score, with a value range of [0.01,1]. ω j Used to evaluate the optimization level of the current parameters, with values ​​ranging from [-1, 1] to λ. j and β jUsed to control the influence weights of different layers, with values ​​ranging from [0.1, 10]; ψ j The value range is [0,1]; κ j The range of values ​​used to control the influence of parameter changes is [0.1, 5]. Used to evaluate the importance of layer freezing, with a value range of [0,1]; LFS dynamically determines which layers need to be frozen, and different freezing strategies are used at different stages, making training more intelligent, while enhancing model robustness, coping with changes in data distribution, and improving the model's performance on non-stationary data.

[0075] The present invention is further configured to quantitatively assess the degree of crack propagation, analyze historical crack propagation data using a BiLSTM model, and combine the latest crack features for time series regression to predict future crack propagation trends, plot crack propagation trend curves, and display the predicted results of crack propagation degree changing over time through a visualization interface. Specifically, stress wave signals during crack propagation are collected using sensors, and the collected signals are preprocessed, including denoising, normalization, and window segmentation, to facilitate input into the model for analysis. Denoising is used to improve the purity of stress wave signals, which may be affected by environmental noise and sensor error factors during the acquisition process. Processing methods include low-pass filtering or Kalman filtering. Normalization is used to map the signal amplitude and scale data of different crack samples to the same range, eliminating the influence of numerical differences on model training. Processing methods include mean-variance normalization or minimum-maximum normalization. Window segmentation is used to extract local temporal features, making it easier for the model to capture crack propagation trends. The analysis methods include sliding window or fixed window segmentation; extraction of crack propagation feature vectors, including spectral mixing entropy, resonant transition rate, modal discrepancy factor, frequency cross-coupling index, and frequency shift skewness index, to describe the dynamic characteristics of crack propagation; extraction of deep feature vectors through the fully connected layer of a trained BiLSTM model as a feature representation of crack propagation; analysis of crack evolution patterns based on time-series information using the BiLSTM model, outputting quantitative assessments of crack length or propagation extent; analysis of historical crack propagation data using the BiLSTM model, combined with the latest crack features for time-series regression to predict future crack propagation trends; data is segmented into sliding windows according to time windows, such as 1 hour, 1 day, or 1 week, generating input-output pairs, where the input represents past crack data and the output represents the predicted crack length or propagation rate; plotting crack propagation trend curves to show the predicted crack length changes over time, and marking critical crack lengths or crack propagation rates to provide early warning of potential structural failure risks.

[0076] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0077] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0078] In this application, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or multiple items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0079] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0080] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0081] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0082] In the several embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0083] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0084] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0085] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An incremental learning method for quantitative detection of crack propagation based on stress wave signals, characterized in that, include: The stress wave signal of the structure during crack propagation is acquired in real time by sensors, and the stress wave signal is preprocessed to extract frequency domain features. A bidirectional long short-term memory (BiLSTM) model was constructed and trained based on a 0-5mm crack sample dataset. High-dimensional feature vectors were extracted through fully connected layers as the initial feature representation for crack detection. By mapping high-dimensional feature vectors to a two-dimensional space, the distribution differences of features of different crack samples can be quantitatively evaluated. The retention ratio of historical data is dynamically adjusted based on the characteristic distribution of crack samples and the detection accuracy of the model; the calculation logic for this dynamic adjustment is as follows: ,in, In order to be in At any given time, the percentage of historical data retained, Differences in the distribution of crack sample characteristics For detection accuracy, and For adjustable parameters, The retention ratio before time t and To adjust parameters, the distribution differences of crack sample characteristics The calculation logic is as follows: ,in, For the current sample probability distribution For historical data probability distribution For joint probability distribution, For joint distribution, For data points and Distance between them, detection accuracy The calculation logic is as follows: ,in, It is a local minimum; During the dynamic adjustment process, a chaotic sampling strategy is introduced, and samples with high uncertainty are selected for empirical replay in combination with the entropy domain fusion index. We introduced 6-10mm crack propagation data and used incremental learning to update the BiLSTM model; The final detection model after training is used to quantitatively assess the degree of crack propagation and output the crack propagation trend prediction results.

2. The incremental learning method for quantitative detection of crack propagation based on stress wave signals according to claim 1, characterized in that, Stress waves during crack propagation are monitored by sensors installed on the structure. The sensors and signal acquisition system are connected, and stress wave signals are collected using data acquisition instruments. The sensors include piezoelectric sensors, vibration sensors, and accelerometers.

3. The incremental learning method for quantitative detection of crack propagation based on stress wave signals according to claim 1, characterized in that, Frequency domain characteristics include spectral mixing entropy, resonant transition rate, modal discrepancy factor, frequency cross-coupling index, and frequency shift skewness index; The calculation logic for the spectral mixing entropy is as follows: ,in, For the spectral mixing entropy, For the complex spectral components after FFT transformation, For the index of frequency components, and For regulatory factors, This represents the number of sampling points; The calculation logic for the resonant transition rate is as follows: ,in, For resonant transition rate, For the first Each frequency component For control factors, It is a tiny constant; The calculation logic for the modal discrepancy factor is as follows: ,in, For modal discretization factor, For the main frequency of the signal, As a control factor; The calculation logic for the frequency cross-coupling index is as follows: ,in, For frequency cross-coupling index, For small constants, and As a regulating factor; The calculation logic for the frequency shift skewness index is as follows: ,in, For frequency shift skewness index, It is a regulating factor.

4. The incremental learning method for quantitative detection of crack propagation based on stress wave signals according to claim 1, characterized in that, The steps to build a BiLSTM model include: The frequency domain feature data extracted from 0-5mm crack samples were standardized and then used as input. A bidirectional long short-term memory network BiLSTM with forward and backward propagation paths was constructed, and the number of network layers, the number of hidden units, and the activation function parameters were set. The basic BiLSTM network is iteratively trained using the preprocessed training dataset until convergence, and a fully connected layer is introduced before the network output layer to extract high-dimensional feature vectors. The high-dimensional feature vectors output by the fully connected layer are used as inputs for subsequent dimensionality reduction analysis and feature distribution evaluation.

5. The incremental learning method for quantitative detection of crack propagation based on stress wave signals according to claim 1, characterized in that, The high-dimensional feature vectors extracted from the fully connected layers of the BiLSTM model are reduced in dimensionality using a difference metric function, mapping the high-dimensional feature vectors to a two-dimensional space. The calculation logic of the difference metric function is as follows: ,in, For difference measurement, For the sample and Similarity in high-dimensional space Data points after dimensionality reduction and Similarity between It represents the ratio of the two distributions. The calculation logic is as follows: ,in, For the sample and Distance between For the similarity scale of each pair of points in high-dimensional space, The calculation logic is as follows: ,in, and This is the lower-dimensional representation after dimensionality reduction; The distribution difference function is used to measure the distribution difference of sample features reduced to two-dimensional space. The calculation logic of the distribution difference function is as follows: ,in, For the difference between the two distributions, For the probability distribution of high-dimensional data, The probability distribution of the dimensionality-reduced low-dimensional data, Let be a point in the sample space.

6. The incremental learning method for quantitative detection of crack propagation based on stress wave signals according to claim 1, characterized in that, In the chaotic sampling strategy, the entropy domain fusion index is combined. By optimizing the objective function, the entropy domain fusion index is selected to prioritize samples with high uncertainty for empirical replay. The formula for calculating the entropy domain fusion index is as follows: ,in, For entropy domain fusion index, For model uncertainty factors, For sample importance weights, The impact of samples on the model's decision boundary, The calculation logic is as follows: ,in, For the feature distance between samples, To control the hyperparameters of the local density influence range, The calculation logic is as follows: ,in, The distance from the sample to the decision boundary, To control the parameters of the scope of influence; The calculation logic for the optimization objective function is as follows: ,in, To optimize the objective function, For density adjustment coefficient, To control parameters for the influence of density, Boundary sample suppression coefficient, This is the adjustment coefficient for the impact of the decision boundary.

7. The incremental learning method for quantitative detection of crack propagation based on stress wave signals according to claim 1, characterized in that, The specific implementation steps for updating the BiLSTM model using incremental learning, based on 6-10mm crack propagation data, include: Preprocessing and frequency domain feature extraction were performed on the 6-10mm crack propagation data, and the newly generated data were made consistent with the old data in terms of feature dimension, format and distribution. A phased incremental learning strategy was adopted, and 6-10mm crack data samples were mixed with 0-5mm old data samples according to a preset ratio to construct a multi-stage training dataset. Freeze some layers of the BiLSTM model structure, while unfreezing the upper network layers to adapt to the feature distribution of new data, and iterate and update the network parameters repeatedly after each batch of new data is introduced. After each stage of training is completed, when the model performance reaches a preset threshold, the incremental learning process is terminated, and the final detection model is obtained.

8. The incremental learning method for quantitative detection of crack propagation based on stress wave signals according to claim 7, characterized in that, The process involves freezing certain layers of the BiLSTM model structure and dynamically adjusting the results using layer freeze scores. The calculation logic for the layer freeze scores is as follows: ,in, For layer freezing scoring, For the loss of old data in layer l gradient change, For the nonlinear weighted exponent of the gradient, To adjust parameters, The change in neuron activation value at the i-th time step is the entropy. Integrating over the parameter range [a, b] For the globally optimal parameter set, and For weight adjustment factor, To adjust the exponential decay term, For exponential decay rate, This is the optimal parameter point for the current layer.

9. The incremental learning method for quantitative detection of crack propagation based on stress wave signals according to claim 1, which quantitatively evaluates the degree of crack propagation, uses a BiLSTM model to analyze historical data of crack propagation, and combines the latest crack features to perform time series regression to predict future crack propagation trends, plots crack propagation trend curves, and displays the prediction results of crack propagation degree changing over time through a visualization interface.

Images