Dam concrete internal crack ultrasonic positioning detection method and system

Abstract

The application provides a dam concrete internal crack ultrasonic positioning detection method and system, the method comprises extracting the time domain, frequency domain and time-frequency domain features of the ultrasonic monitoring signal and constructing a feature vector, dividing the dam detection area into a three-dimensional grid, and constructing a training set. In the network training stage, adaptive feature weights are generated by combining global mutual information and similar K nearest neighbor variance, a local Gaussian kernel is constructed based on a regularized similar K nearest neighbor covariance matrix, and a boundary confidence is calculated according to the distance of dissimilar neighbors. In the detection stage, the feature vector to be measured is subjected to weighted difference transformation, substituted into the local covariance Gaussian kernel and combined with the boundary confidence to generate a class score, and a three-dimensional crack posterior probability field is formed after normalization. Finally, the high-probability connected region is subjected to principal component analysis, the crack principal axis direction and size are extracted, and the three-dimensional coordinates of the crack center are calculated through probability weighted centroid.

Description

Technical Field

[0001] This application belongs to the field of ultrasonic positioning, and in particular relates to an ultrasonic positioning and detection method and system for internal cracks in dam concrete. Background Technology

[0002] In water conservancy and hydropower projects, the structural safety of the dam is directly related to the operational stability of the project. During long-term service, dam concrete is susceptible to the formation and propagation of hidden cracks due to factors such as load, temperature stress, environmental erosion, and material aging. These cracks not only weaken the dam structure but can also create seepage channels, affecting the overall safety of the dam. Traditional surface observation is insufficient to detect internal defects in a timely manner, and destructive core drilling cannot meet the requirements for continuous, comprehensive, and non-destructive testing. Therefore, ultrasonic testing, due to its strong penetrating power and sensitivity to internal defects, is widely used for detecting cracks in dam concrete.

[0003] However, the large volume and high material inhomogeneity of dam concrete, along with coarse aggregates, micropores, and interface defects, cause ultrasonic wave scattering, attenuation, and noise superposition, resulting in complex received signal characteristics. Dividing the dam space into a three-dimensional grid and extracting the time-frequency features of the ultrasonic signal, which are then input into a probabilistic neural network for grid classification, is used to locate cracks. However, probabilistic neural networks struggle to distinguish the contributions of different features in global classification and local distribution during distance calculations; their Gaussian kernels typically use uniform smoothing parameters, making it difficult to adapt to the local anisotropic distribution of ultrasonic features; furthermore, traditional networks lack confidence constraints on the boundaries of outlier samples, easily leading to confusion at adjacent grid boundaries. In addition, existing models mostly output discrete grid labels, lacking sufficient ability to extract the spatial connectivity probability field, center coordinates, principal axis directions, and dimensions of cracks, making it difficult to meet the requirements for refined three-dimensional detection of cracks inside dams. Summary of the Invention

[0004] To address the problems in existing technologies, such as fixed ultrasonic feature weights, single kernel function smoothing parameters, easy confusion in grid boundary positioning, and difficulty in reconstructing the three-dimensional probability field and geometric morphology of cracks, this invention proposes an ultrasonic positioning and detection method and system for internal cracks in dam concrete.

[0005] In the first aspect, the present invention proposes an ultrasonic positioning and detection method for internal cracks in dam concrete, comprising the following steps:

[0006] The monitoring signals from multiple ultrasonic sensors are acquired, and the time-domain and frequency-domain features of the monitoring signals are extracted and dimensionless to form feature vectors. The dam detection area is pre-divided into a three-dimensional grid, and a training set containing feature vectors and category labels is constructed and a probabilistic neural network is established.

[0007] During the training phase, global importance weights are calculated based on the mutual information between class labels and feature vectors. For each neuron in the pattern layer, the variance of the k nearest neighbor sample set of the corresponding training sample in the same class is calculated as the local feature divergence. The global importance weights and local feature divergence are combined to generate adaptive feature weight coefficients, and the feature differences are weighted and transformed. The covariance matrix of the k nearest neighbor sample set is regularized and used as the localized covariance matrix of the Gaussian radial basis kernel function. The boundary confidence is calculated based on the average distance between the training sample corresponding to the neuron and the k nearest out-of-class samples.

[0008] For the feature vector to be tested, a quadratic form is formed based on the weighted difference vector between the feature vector to be tested and the corresponding training samples of the pattern layer neurons after adaptive feature weight coefficient weighting transformation, and the inverse matrix of the regularized local covariance matrix. The quadratic form is substituted into the Gaussian radial basis kernel function to generate the output value and multiplied by the boundary confidence weight. In the summation layer, the weighted output value is accumulated according to the category label and normalized to obtain the posterior probability of the three-dimensional mesh crack and generate a three-dimensional probability field. For connected regions with probabilities higher than the preset threshold, principal component analysis is used to extract the geometric principal axis direction and size, and the three-dimensional coordinates of the crack center are calculated by normalizing and weighting the probability values ​​of the grid points.

[0009] In another aspect, the present invention also proposes an ultrasonic positioning and detection system for internal cracks in dam concrete, comprising:

[0010] The acquisition module is used to acquire monitoring signals from multiple ultrasonic sensors, extract time-domain and frequency-domain features of the monitoring signals and perform dimensionless processing to form feature vectors, pre-divide the dam detection area into a three-dimensional grid, construct a training set containing feature vectors and category labels, and establish a probabilistic neural network.

[0011] The training module is used to calculate the global importance weights based on the mutual information of class labels and feature vectors during the training phase. For each neuron in the pattern layer, the variance of the k nearest neighbor sample set of the corresponding training sample in the same class is calculated as the local feature divergence. The global importance weights and local feature divergence are combined to generate adaptive feature weight coefficients, and the feature differences are weighted and transformed. The localized covariance matrix of the Gaussian radial basis kernel function is obtained by regularizing the covariance matrix of the k nearest neighbor sample set. The boundary confidence is calculated based on the average distance between the training sample corresponding to the neuron and the k nearest out-of-class samples.

[0012] The detection module is used to form a quadratic form based on the weighted difference vector between the feature vector to be tested and the corresponding training samples of the pattern layer neurons after adaptive feature weight coefficient weighting transformation, and the inverse matrix of the regularized local covariance matrix. The quadratic form is substituted into the Gaussian radial basis kernel function to generate the output value and multiplied by the boundary confidence weight. In the summation layer, the weighted output value is accumulated according to the category label and normalized to obtain the posterior probability of the three-dimensional mesh crack and generate a three-dimensional probability field. For connected regions with probabilities higher than a preset threshold, principal component analysis is used to extract the geometric principal axis direction and size, and the three-dimensional coordinates of the crack center are calculated by normalizing and weighting the probability values ​​of the grid points.

[0013] This invention provides a method and system for ultrasonic location detection of internal cracks in dam concrete, improving the accuracy of crack location and the completeness of geometric representation. By extracting multi-domain features from ultrasonic monitoring signals and combining mutual information and local feature divergence to generate adaptive feature weight coefficients, the contribution of different features to crack location can be distinguished, reducing the impact of redundant features and noise interference. For the probabilistic neural network, a regularized local covariance matrix is ​​introduced to optimize the Gaussian radial basis kernel function, and boundary confidence is constructed based on the heterogeneous nearest neighbor distance to adjust the output of blurred boundary samples, thereby improving the model's ability to distinguish between adjacent grids and complex class boundary regions. Furthermore, a three-dimensional crack probability field is generated by normalizing class scores, and combined with principal component analysis and probability-weighted centroid calculation, the three-dimensional coordinates, geometric principal axis directions, and dimensional parameters of the crack center can be obtained, providing reliable data support for non-destructive location of internal cracks in dam concrete and structural safety assessment. Attached Figure Description

[0014] Figure 1 A flowchart of the ultrasonic positioning and detection method for internal cracks in dam concrete;

[0015] Figure 2 A schematic diagram of the global importance weight distribution of a 96-dimensional feature vector;

[0016] Figure 3 This is a schematic diagram of the coordinate projection of the spatial probability field connected region and the crack core. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] In the first embodiment, the present invention proposes an ultrasonic positioning and detection method for internal cracks in dam concrete, such as... Figure 1 As shown, it includes:

[0019] S1. Acquire monitoring signals from multiple ultrasonic sensors, extract time-domain and frequency-domain features of the monitoring signals and perform dimensionless processing to form feature vectors, pre-divide the dam detection area into a three-dimensional grid, construct a training set containing feature vectors and category labels, and establish a probabilistic neural network.

[0020] In a piezoelectric ceramic ultrasonic sensor network deployed on the concrete surface of the dam, ultrasonic waves are excited and transmitted and reflected waves are simultaneously received as monitoring signals. For time-domain features, wavefront arrival time, first wave amplitude, and signal duration are extracted. For frequency-domain features, the peak frequency and main frequency band energy ratio are extracted using a fast Fourier transform algorithm. For time-frequency domain features, wavelet packet energy entropy is extracted using wavelet packet transform. All extracted features are merged to form an initial feature vector, and standard fractionalization is performed to achieve dimensionless processing and eliminate the influence of different dimensions on distance calculation, thus forming the feature vector. A three-dimensional Cartesian coordinate system for the dam concrete is established, and a voxelization algorithm is used to divide the detection area into three-dimensional grids with fixed side lengths at equal intervals. Each grid is assigned a unique natural number number. Crack calibration samples, artificially prefabricated crack samples, or numerically simulated crack samples covering candidate three-dimensional grids within the detection area are obtained, and the three-dimensional grid number corresponding to the crack in each sample is set as the crack category label. Simultaneously, background samples corresponding to crack-free calibration areas, healthy concrete test blocks, or crack-free numerical simulation models are obtained, and these background samples are uniformly set as crack-free category labels.

[0021] The feature vectors obtained from the above processing and their corresponding class labels are combined into a training dataset. A probabilistic neural network with a four-layer structure, including an input layer, a pattern layer, a summation layer, and an output layer, is constructed. The number of neurons in the input layer is equal to the dimension of the feature vector. The pattern layer assigns an independent neuron to each sample in the training set. The number of nodes in the summation layer is equal to the sum of the total number of pre-divided 3D grids and the number of crack-free background categories. The output layer normalizes the scores of each category obtained from the summation layer to generate the crack probability distribution corresponding to each 3D grid and the probability value corresponding to the crack-free background category. When the crack-free background category probability is the highest or the probability of each crack grid category is lower than the preset judgment threshold, it is determined that the current test signal does not correspond to a valid internal crack region.

[0022] In a preferred embodiment, the extraction of the time-domain and frequency-domain features of the monitoring signal includes:

[0023] Based on the time-domain waveform or envelope signal of the monitoring signal acquired by the ultrasonic sensor, the arrival time of the wavefront, the amplitude of the first wave, and the duration of the signal are extracted as time-domain features. Spectral analysis is performed on the monitoring signal to calculate the peak frequency and the energy proportion of the main frequency band within a set frequency range as frequency-domain features. The energy entropy of specified wavelet packet nodes is extracted using the wavelet packet decomposition algorithm as time-frequency domain features. After dimensionless processing of the extracted time-domain, frequency-domain, and time-frequency-domain features of all channels, they are spliced ​​and combined into a one-dimensional real number array according to a predetermined channel arrangement order to generate a feature vector.

[0024] When extracting features from ultrasonic monitoring signals, the original discrete monitoring signal with a sampling frequency of 1MHz to 5MHz (preferably 2MHz) and a length of 1000 to 5000 sampling points is first acquired. The original discrete monitoring signal undergoes DC removal, bandpass filtering, and envelope extraction. The arrival time of the ultrasonic wavefront is then picked up using a set threshold method, for example, a threshold of 3 to 5 times the root mean square of the ambient noise, such as 45.2 microseconds. The corresponding first wave amplitude is extracted, such as 1.5 volts. The span from the wavefront to the envelope, decreasing to 10% of the peak value, is calculated as the signal duration, such as 120 microseconds, and this is used as the time-domain feature. Subsequently, the original discrete monitoring signal is subjected to a Hamming window followed by a Fast Fourier Transform (FFT) or Short-Time Fourier Transform (SFT) to calculate the spectrum of the monitoring signal within a set frequency band, preferably 20kHz to 200kHz. A peak search algorithm is used to determine the frequency corresponding to the maximum amplitude within this band as the peak frequency, for example, 85.5kHz. The ratio of the energy integral within a 10kHz band above and below this peak frequency to the energy integral of the entire 20kHz to 200kHz band is calculated as the main frequency band energy proportion, for example, 78.3%, thus forming the frequency domain feature. In the time-frequency domain feature extraction, the signal is decomposed into 3 to 5 layers (preferably 4 layers) of wavelet packets using the Daubechies wavelet series, such as the preferred db4 or db8 wavelet basis, generating 16 frequency band nodes. The sum of squares of the signal coefficients within each specified wavelet packet node is calculated as the node energy. Then, based on the probability distribution of each node's energy as a percentage of the total energy, the wavelet packet energy entropy is calculated using the Shannon entropy formula, for example, 2.15. The extracted wavefront arrival time, first wave amplitude, signal duration, peak frequency, main frequency band energy ratio, and wavelet packet energy entropy are dimensionless by using the Min-Max normalization method or the Z-score normalization method. The features of all ultrasonic channels, such as the 16-channel hardware layout, are strictly spliced ​​together according to the channel number order (CH1 to CH16) to generate a one-dimensional real number array feature vector with a fixed length of 96 dimensions.

[0025] S2: During the training phase, global importance weights are calculated based on the mutual information between class labels and feature vectors. For each neuron in the pattern layer, the variance of the k nearest neighbor sample set of the corresponding training sample in the same class is calculated as the local feature divergence. The global importance weights and local feature divergence are combined to generate adaptive feature weight coefficients, and the feature differences are weighted and transformed. The covariance matrix of the k nearest neighbor sample set is regularized and then used as the localized covariance matrix of the Gaussian radial basis kernel function. The boundary confidence is calculated based on the average distance between the training sample corresponding to the neuron and the k nearest out-of-class samples.

[0026] Specifically, during the network training phase, the mutual information value between each feature dimension and the 3D grid category label is calculated and divided by the sum of all feature mutual information values ​​to obtain the global importance weight of that feature dimension. For each neuron in the pattern layer, a multidimensional spatial segmentation tree algorithm is used to search for the k nearest neighbor samples of its corresponding training sample in the sample set with the same category label. The search process is preferably implemented by using the cKDTree function. The values ​​of these k nearest neighbor samples in each feature dimension are extracted and the feature variance is calculated as the local feature divergence. The obtained global importance weight is divided by the local feature divergence and a very small positive constant is added to prevent the denominator from being zero to obtain the fusion weight of each feature dimension. The fusion weight is then subjected to exponential normalization so that the sum of all dimension weights equals 1, thereby generating adaptive feature weight coefficients.

[0027] After weighting the k nearest neighbor samples of the same class according to the adaptive feature weight coefficients, the covariance matrix between each feature dimension is calculated. A regularization coefficient is added to the diagonal of the covariance matrix to obtain a regularized local covariance matrix. This regularized local covariance matrix is ​​used to replace the single and fixed smoothing parameter in the standard probabilistic neural network as the localized covariance matrix of the Gaussian radial basis kernel function to reflect the local distribution characteristics of the data in different dimensions. The cKDTree function is called again to search for the k nearest outlier samples of the training sample in different class label sample sets and calculate the average Euclidean distance between them. The average distance is substituted into the Sigmoid activation function to obtain a normalized value that is positively correlated with the distance. This value is used as the boundary confidence of the corresponding neuron. The farther away from the outlier sample, the higher the confidence.

[0028] In a preferred embodiment, the calculation of global importance weights based on the mutual information between category labels and feature vectors includes:

[0029] During the training phase, the continuous features of the feature vector in the training set are discretized by interval binning. The joint probability distribution of the j-th feature of the discretized feature vector and the corresponding category label is calculated. The marginal probability distribution of the j-th feature and the marginal probability distribution of the category label are calculated respectively. The mutual information value corresponding to the j-th feature is calculated.

[0030] The global importance weights are obtained by normalizing the mutual information values ​​of each dimension.

[0031] Before calculating the global importance weights, the continuous feature values ​​of the feature vectors in the training set are discretized using interval binning to eliminate the probability distribution calculation bias caused by the large span of continuous values. In this process, the j-th dimension continuous feature value is divided into several discrete intervals using either equal-frequency binning or equal-distance binning. The preferred number of bins is 5 to 15, for example, 10 bins are selected. The marginal probability distribution of the j-th dimension feature in each discrete interval is calculated by counting the number of samples in each discrete bin and dividing by the total number of samples in the training set (e.g., if the total number of samples is 10,000). Similarly, by statistically analyzing the sample frequencies of each type of dam grid number (e.g., crack category labels totaling 500 categories), the marginal probability distribution of the category labels is calculated. Furthermore, the joint probability distribution of features falling within a specific interval and belonging to a specific category label is calculated using a two-dimensional frequency statistics matrix. .

[0032] After obtaining all the above probability distribution values, the Shannon mutual information formula is used to calculate the corresponding mutual information value I for each feature dimension and each category label. For example, when a certain frequency domain feature has a very high correlation with a specific crack mesh category, its mutual information value I can reach 0.85 bits; if the feature only contains background noise, the mutual information value may approach 0.01 bits. All mutual information values ​​for a total of D=96 dimensions are summed, and then the global importance weight W for each dimension is calculated using the normalization formula, such as... Figure 2 As shown, through this global normalization operation, the value range of the global importance weight W is limited to between 0 and 1, and the sum of the weights of all dimensions equals 1. The weight vector reflects the relative information contribution of features of different dimensions in the spatial three-dimensional grid localization and classification task of dam cracks. The probabilistic neural network structure adopted specifically includes an input layer, a pattern layer, a summation layer, and an output layer. The input of the probabilistic neural network is a one-dimensional real number array feature vector formed by concatenating and combining features in channel order after dimensionless processing. The output of the probabilistic neural network is the predicted three-dimensional grid point category and the corresponding posterior probability distribution.

[0033] In a preferred embodiment, the step of combining global importance weights with local feature divergence to generate adaptive feature weight coefficients and performing a weighted transformation on the feature differences includes:

[0034] The variance of the j-th dimension feature of the training sample corresponding to the neuron in the pattern layer is calculated and defined as the local feature divergence among the k nearest neighbor samples of the same type in the training set.

[0035] Adaptive feature weight coefficients are generated based on the global importance weights and local feature divergences of each feature dimension. The feature dimension with the larger global importance weight and the smaller local feature divergence corresponds to the larger adaptive feature weight.

[0036] The weights of each dimension are normalized so that the sum of the adaptive feature weights of all feature dimensions is 1.

[0037] Calculate the difference between the feature vector to be tested and the feature vector of the corresponding training sample of the pattern layer neuron in each dimension, and multiply each difference by the square root of the corresponding adaptive feature weight coefficient to obtain the weighted difference vector.

[0038] When generating adaptive feature weight coefficients, the Euclidean distance is first used to search for the training sample corresponding to the i-th neuron of the current pattern layer within the training set. The k nearest neighbor samples of the same class are denoted as , The value is preferably an integer between 5 and 20, for example, k=10. Extract the data of the k nearest neighbor samples of the same type on the j-th feature, and calculate the mean of the j-th feature. ,according to Calculate the unbiased variance of this dimension, and define the degree of numerical fluctuation in this dimension as the local feature divergence. For example, if the variance calculation result is 0.0025, then... This represents the value of the l-th nearest neighbor sample of the same type in the training sample corresponding to the i-th pattern layer neuron on the j-th dimension feature. Optionally, to reduce the impact of individual abnormal nearest neighbor samples on the local feature divergence, before calculating the local feature divergence, outlier removal, quantile truncation, or amplitude limiting can be performed on the values ​​of the k nearest neighbor samples of the same type on the j-th dimension feature, and the variance of the j-th dimension feature after processing can be used as the local feature divergence. Simultaneously, it can analyze the local feature divergence. Setting preset lower and upper limits allows features to participate in adaptive feature weight calculations within a defined range, preventing excessive weighting of globally important features due to local abnormal fluctuations.

[0039] To ensure the numerical stability of the weight calculation formula and avoid the denominator being zero, a minimal constant is used. Set as to For example, the value is In the adaptive feature weight formula, the global importance weight of the j-th dimension feature is... As a molecule, local characteristic divergence Add a minimal constant First, calculate the fusion weight factor as the denominator. Then according to An adaptive feature weight coefficient is generated for the j-th dimension feature corresponding to the i-th pattern layer neuron. This non-linearly amplifies the weights of features that contribute significantly to global classification and have high intra-class aggregation, ensuring that the weights of all dimensions sum to 1. For example, when D=96, the weight w of a certain effective feature can be enhanced to 0.12, while redundant features are attenuated to 0.001. In the feature difference transformation stage, the generated adaptive feature weight coefficients are used... The feature vector to be tested x and the training samples corresponding to the i-th pattern layer neurons We perform a weighted transformation on the dimension-wise differences between them to obtain a weighted difference vector. ,in The j-th dimension component is The weighted difference vector is used for the subsequent quadratic calculation of the local covariance Gaussian kernel function.

[0040] In a preferred embodiment, calculating the boundary confidence based on the average distance between the training sample corresponding to the neuron and the k nearest out-of-class samples includes:

[0041] Calculate the Euclidean distance between the training sample corresponding to the i-th pattern layer neuron and all out-of-class training samples, extract the k out-of-class samples with the smallest distance, calculate the arithmetic mean of the distances of these k out-of-class samples as the out-of-class neighbor distance mean, and perform normalization to obtain the boundary confidence.

[0042] The algorithm iterates through all training samples of the dam grid with different class labels, calculating the distance in the feature space between the training sample corresponding to the i-th pattern layer neuron and all out-of-class training samples using the Euclidean distance formula. The resulting set of distance values ​​is sorted in ascending order from smallest to largest, and the k out-of-class samples with the smallest distances at the top are extracted. The preferred range for the number of out-of-class neighbors k is 10 to 30, for example, k=15. The arithmetic mean of these k smallest Euclidean distance values ​​is calculated to obtain the mean out-of-class neighbor distance D, which represents the degree of out-of-class isolation. If the mean D is large, such as 2.45, it indicates that the neuron is deep within the same data cluster, with a low risk of misclassification by out-of-class samples. If the mean D is extremely small, such as 0.15, it indicates that the neuron is close to the dangerous classification decision boundary, belonging to a high-risk sample point that is highly prone to causing localization confusion. To convert the above distance-based safety margin into a multiplicative coefficient that can be called by the probabilistic network, the mean out-of-class neighbor distance of the i-th neuron is used during the operation. Input the positive correlation sigmoid mapping function to obtain the boundary confidence score. ,For example ,in, This is the slope adjustment coefficient. The benchmark threshold for the heterogeneous proximity distance can be the average or median of the heterogeneous proximity distances of neurons in each pattern layer in the training set. The larger the value, the farther the neuron is from the boundary of the out-of-class sample, and the higher the confidence level of the corresponding boundary. The smaller the value, the closer the neuron is to the class boundary, and the lower the corresponding boundary confidence. For example, when the mean distance between the neurons in the center of a cluster is significantly greater than the baseline threshold, its boundary confidence can be mapped to a value close to 1; when the mean distance between the neurons in the ambiguous boundary is significantly less than the baseline threshold, its boundary confidence can be mapped to a value close to 0. During the probabilistic forward propagation phase of the network, the boundary confidence corresponding to the i-th pattern layer neuron is... The output value of the Gaussian radial basis kernel function of this neuron Multiplication yields the boundary-weighted output value of the pattern layer. This multiplicative regulation mechanism helps suppress overactivation of neurons in inter-class overlapping regions, reducing the risk of false alarms and spatial jump bias in the ultrasonic localization task of dam concrete.

[0043] S3. For the feature vector to be tested, a quadratic form is formed based on the weighted difference vector between the feature vector to be tested and the corresponding training samples of the pattern layer neurons after adaptive feature weight coefficient weighting transformation, and the inverse matrix of the regularized local covariance matrix. The quadratic form is substituted into the Gaussian radial basis kernel function to generate the output value and multiplied by the boundary confidence weight. In the summation layer, the weighted output value is accumulated according to the category label and normalized to obtain the posterior probability of the three-dimensional mesh crack and generate a three-dimensional probability field. For the connected regions with probabilities higher than the preset threshold, the geometric principal axis direction and size are extracted by principal component analysis and the three-dimensional coordinates of the crack center are calculated by normalizing and weighting the probability values ​​of the grid points.

[0044] When the ultrasonic signal to be tested is input, the dimension-wise differences between the feature vector to be tested and the training sample vectors corresponding to each neuron in the pattern layer are calculated. The differences in each dimension are scaled according to the square root of the corresponding adaptive feature weight coefficient to obtain a weighted difference vector. The transpose of the weighted difference vector is multiplied by the inverse of the regularized local covariance matrix, and then multiplied by the weighted difference vector to obtain the quadratic operation result. The quadratic result is multiplied by negative half and used as the input variable of the natural exponential function to generate the initial output value of the Gaussian radial basis kernel function. The initial output value is multiplied by the boundary confidence corresponding to the neuron to obtain the weighted output value. In the summation layer, the neurons in the pattern layer are grouped according to their category labels. The weighted output values ​​belonging to the same three-dimensional grid number are accumulated or averaged to obtain the category score of the corresponding three-dimensional grid. The category score of each crack grid is divided by the sum of all category scores, which includes the category score of each crack grid and the category score of the no-crack background, to obtain the normalized posterior probability corresponding to each crack grid.

[0045] The posterior probability values ​​corresponding to the crack mesh categories are mapped to a three-dimensional probability field according to their spatial coordinates. The posterior probability of the crack-free background category is only used to determine the existence of valid cracks and does not participate in the three-dimensional probability field mapping, principal component analysis of crack connected regions, or probability-weighted centroid calculation. A probability threshold is set, and the probability of crack meshes below the threshold is set to zero, thereby segmenting high-probability meshes and finding and marking connected regions in the three-dimensional probability field. The spatial coordinates of the meshes in each connected region are extracted, and the three-dimensional point cloud feature values ​​and feature vectors of the region are extracted using PCA principal component analysis. The direction of the feature vector corresponding to the first principal component is taken as the geometric principal axis direction of the crack, and the coordinate point matrix is ​​projected onto the three principal component directions respectively. The difference between the maximum and minimum values ​​of the projection in each direction is calculated as the size of the crack in the principal axis direction and the two orthogonal directions. Finally, the three-dimensional coordinate center position of all mesh points in the connected region is extracted. The three-dimensional coordinates of each mesh are multiplied by its corresponding posterior probability, summed, and then divided by the sum of the posterior probabilities of all meshes in the connected region to obtain the three-dimensional coordinates of the crack center inside the dam concrete. Figure 3 As shown.

[0046] In a preferred embodiment, the step of forming a quadratic form based on the weighted difference vector obtained by adaptive feature weighting transformation between the feature vector to be tested and the corresponding training samples of the pattern layer neurons, and the inverse matrix of the regularized local covariance matrix, and substituting the quadratic form into the Gaussian radial basis function to generate the output value, includes:

[0047] Obtain the k nearest neighbor samples with the smallest distance among similar training samples for the corresponding neurons in the pattern layer. Perform a weighted transformation on the feature vectors of these k nearest neighbor samples according to adaptive feature weight coefficients. Calculate the arithmetic center coordinate vectors of the k nearest neighbor samples after the weighted transformation and obtain the covariance matrix. Add a regularization coefficient to the diagonal of the covariance matrix to obtain a regularized local covariance matrix. Calculate the inverse of the regularized local covariance matrix as the matrix weight parameter C. Calculate the output value of the radial Gaussian kernel of the network layer according to the following formula. ,in, Let C be the difference vector between the feature vector to be tested and the feature vector of the corresponding training sample of the pattern layer neuron after weighting by adaptive feature weight coefficients, and let C be the inverse of the regularized local covariance matrix.

[0048] When correcting the localized covariance matrix of the Gaussian radial basis function kernel, to capture the multidimensional geometric distribution of training samples in its feature space, the k nearest neighbor samples with the smallest distance to the neurons in the pattern layer are first selected. The number of nearest neighbors k can be set according to the feature dimension and the number of samples of the same class, and should not exceed the number of training samples in the corresponding class minus one, so as to improve the stability of the local covariance matrix estimation. For example, it can be 10 to 30 depending on the number of samples of the same class. When the number of nearest neighbors is less than the feature dimension, the stability of matrix inversion and local distribution estimation is improved by regularized diagonal loading or shrinking the covariance estimation. The features of each dimension of these k nearest neighbors are scaled according to the square root of the adaptive feature weight coefficient to obtain the weighted transformed feature vector of the nearest neighbors. The arithmetic mean of the values ​​of each dimension of the weighted transformed k nearest neighbors is obtained to obtain the coordinate vector representing the local cluster center. Next, the covariance matrix of the k nearest neighbor samples around the cluster center is calculated. This matrix is ​​a D×D symmetric matrix, such as 96×96, where the diagonal elements represent the variance of each dimension and the off-diagonal elements represent the covariance between dimensions. Before inversion, a regularized diagonal loading process is performed, for example, by uniformly adding a size of 10 to the diagonal. -4 The positive real numbers are used to obtain the regularized local covariance matrix, and the inverse of the regularized local covariance matrix is ​​used to obtain the matrix weight parameter C. This parameter C is used to characterize the precision scaling of the local multidimensional distribution features.

[0049] When calculating the final output value O of the radial Gaussian kernel of the network layer, a weighted difference vector is generated between the feature vector to be tested and the corresponding feature vector of the training sample. The weighted difference vector has dimensions D rows and 1 column, for example, 96×1. The transpose of the weighted difference vector... Matrix weight parameters C and weighted difference vector Perform matrix multiplication to form a quadratic scalar value representing the anisotropy of the local spatial distribution of the data; for example, the result is 3.45. Multiply this quadratic scalar value by -1 / 2 and then input it into the natural exponential function exp to complete the nonlinear mapping, generating the output value O of the Gaussian radial basis function kernel. This kernel function output value is used to characterize the similarity between the current test sample and the local sample cluster represented by the neurons in this pattern layer.

[0050] In a preferred embodiment, the step of extracting the geometric principal axis direction and size from the connected regions with probabilities higher than a preset threshold using principal component analysis and then calculating the three-dimensional coordinates of the crack center using normalized weighted calculations based on the grid point probability values ​​includes:

[0051] A set of three-dimensional meshes with posterior probabilities greater than a preset threshold is selected. The set of three-dimensional meshes is divided into at least one high-probability connected region according to spatial adjacency. For the r-th high-probability connected region, the three-dimensional coordinate matrix of the meshes within the high-probability connected region is extracted. The covariance matrix of the coordinate matrix is ​​decomposed into eigenvalues. The direction of the eigenvector corresponding to the largest eigenvalue is taken as the geometric principal axis direction of the crack in the high-probability connected region. The extreme difference of the projection of the coordinate matrix in the principal axis direction is taken as the size of the crack. The three-dimensional coordinates of each mesh point in the r-th high-probability connected region are multiplied by the corresponding posterior probability value, summed, and divided by the sum of the posterior probability values ​​of all mesh points in the high-probability connected region to obtain the core three-dimensional coordinates of the crack.

[0052] When extracting crack geometry and core location coordinates based on the posterior three-dimensional probability field, a preset threshold for the posterior probability is first set according to the signal-to-noise ratio requirements of the actual dam engineering detection accuracy. This preset threshold can be adaptively determined based on the maximum posterior probability or the quantile of the probability distribution in the three-dimensional probability field. For example, it can be a preset proportion of the maximum posterior probability. As a threshold, where The value can be between 0.5 and 0.8, or the grids with the highest posterior probability ranking within a preset proportion can be selected as high-probability grids. The posterior probabilities of all crack grid nodes within the spatial field are traversed and filtered, removing grids with probabilities below the adaptive threshold, retaining the internal connected regions composed of high-probability grids, and extracting the three-dimensional physical space center coordinates of these effective grids. , , A coordinate matrix containing Q discrete points, for example, Q = 350 spatial grid points, is constructed. Next, the 3×3 covariance matrix of these 350 three-dimensional coordinate points is calculated, and eigenvalue decomposition is performed on the matrix using Singular Value Decomposition (SVD) or the Jacobi iteration algorithm to obtain three non-negative eigenvalues ​​and their corresponding orthogonal eigenvectors.

[0053] The unit eigenvector corresponding to the largest principal eigenvalue is selected and calibrated as the geometric principal axis direction of the concrete crack. For example, the direction vector is calculated as [0.866, 0.500, 0.000], representing that the crack mainly extends along a specific angle in the XY plane. To determine the physical size distribution range of the internal crack in three-dimensional space, the Q selected and retained three-dimensional coordinate lattices are projected onto the three principal component directions respectively. The difference between the maximum and minimum values ​​of the projected scalar values ​​in each direction is calculated as the size parameters of the crack in the principal axis direction and the two orthogonal directions. For example, the extreme span of the projection in the principal axis direction is calculated to be 2.35 meters. In the step of calculating the three-dimensional center point of the crack core, the probability-weighted centroid estimation theory is introduced, and the formula is used to solve independently for the three-dimensional coordinate components of space X, Y, and Z: using the posterior probability value of the q-th grid. If the probability of a specific grid cell is 0.82, and this is used as a weight, multiplied by its corresponding coordinate position component (e.g., if the X-coordinate value is 15.5 meters), the summation is applied to all Q grid nodes, and then divided by the posterior probability value. The overall summation term. The weighted centroid coordinates X, Y, and Z obtained through this calculation mode, for example, the final output is located at X=15.82m, Y=8.45m, and Z=12.10m. It can reduce the grid positioning scattering error caused by ultrasonic diffraction and material inhomogeneity by utilizing the intensity distribution of the three-dimensional probability field, which is beneficial to improving the positioning stability and positioning accuracy of the center coordinates of deep cracks in dam concrete.

[0054] The experiment was conducted using a dataset collected by the dam's ultrasonic detection system. The dataset included at least 6000 training samples and 1200 test samples, covering 500 categories of crack 3D mesh labels and crack-free background categories. The nearest neighbor number k for each category was adaptively determined based on the number of training samples in that category, and did not exceed the number of training samples in that category minus 1. The experiment was run on a unified processor with 64GB of memory. A 96-dimensional real number array generated by the feature extraction module was used as the model input, and the 3D mesh positioning accuracy, average spatial positioning error, and false alarm rate in hazardous areas were used as the main evaluation metrics. Comparative models included a basic probabilistic neural network model, an ablation model one that removed adaptive feature weights and local covariance matrix correction, an ablation model two that removed the multiplicative weighting of boundary confidence, and a complete scheme integrating the above mechanisms.

[0055] Validated by forward propagation inference on the test set, the basic probabilistic neural network model achieved a localization accuracy of 81.5%, an average spatial localization error of 0.42 meters, and a false alarm rate of 7.8%. Ablation Model 1 improved the localization accuracy to 87.2%, reduced the average spatial localization error to 0.28 meters, and lowered the false alarm rate to 5.3%. Ablation Model 2 achieved a localization accuracy of 91.4%, an average spatial localization error of 0.19 meters, and a false alarm rate of 3.6%. The complete solution, under the same test conditions, achieved a localization accuracy of 98.2%, an average spatial localization error of 0.08 meters, and a false alarm rate of 0.5%. Compared to the basic probabilistic neural network model, the complete solution improved the localization accuracy by 16.7 percentage points and significantly reduced the average spatial localization error. The results of Ablation Model 1 indicate that the adaptive feature weights formed by combining global importance weights and local feature divergence help to weaken the interference of redundant features and local fluctuation features on the localization results. Combined with regularized local covariance matrices, the model can better adapt to the local anisotropic distribution of ultrasound feature data. The results of ablation model 2 show that in regions with overlapping features or similar class boundaries, the boundary confidence weighting mechanism can suppress the overactivation of samples with blurred boundaries, reduce the risk of false alarms and spatial jump bias, and thus improve the stability and reliability of three-dimensional probability field construction, principal component analysis and probability weighted centroid localization.

[0056] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0057] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0058] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for ultrasonic location detection of internal cracks in dam concrete, characterized in that, include: The monitoring signals from multiple ultrasonic sensors are acquired, and the time-domain and frequency-domain features of the monitoring signals are extracted and dimensionless to form feature vectors. The dam detection area is pre-divided into a three-dimensional grid, and a training set containing feature vectors and category labels is constructed and a probabilistic neural network is established. During the training phase, global importance weights are calculated based on the mutual information between class labels and feature vectors. For each neuron in the pattern layer, the variance of the k nearest neighbor sample set of the corresponding training sample in the same class is calculated as the local feature divergence. The global importance weights and local feature divergence are combined to generate adaptive feature weight coefficients, and the feature differences are weighted and transformed. The covariance matrix of the k nearest neighbor sample set is regularized and used as the localized covariance matrix of the Gaussian radial basis kernel function. The boundary confidence is calculated based on the average distance between the training sample corresponding to the neuron and the k nearest out-of-class samples. For the feature vector to be tested, a quadratic form is formed based on the weighted difference vector between the feature vector to be tested and the corresponding training samples of the pattern layer neurons after adaptive feature weight coefficient weighting transformation, and the inverse matrix of the regularized local covariance matrix. The quadratic form is substituted into the Gaussian radial basis kernel function to generate the output value and multiplied by the boundary confidence weight. In the summation layer, the weighted output value is accumulated according to the category label and normalized to obtain the posterior probability of the three-dimensional mesh crack and generate a three-dimensional probability field. For connected regions with probabilities higher than the preset threshold, principal component analysis is used to extract the geometric principal axis direction and size, and the three-dimensional coordinates of the crack center are calculated by normalizing and weighting the probability values ​​of the grid points.

2. The method according to claim 1, characterized in that, The extraction of time-domain and time-frequency domain features of the monitoring signal includes: Based on the time-domain waveform or envelope signal of the monitoring signal acquired by the ultrasonic sensor, the arrival time of the wavefront, the amplitude of the first wave, and the duration of the signal are extracted as time-domain features. Spectral analysis is performed on the monitoring signal to calculate the peak frequency and the energy proportion of the main frequency band within a set frequency range as frequency-domain features. The energy entropy of specified wavelet packet nodes is extracted using the wavelet packet decomposition algorithm as time-frequency domain features. After dimensionless processing of the extracted time-domain, frequency-domain, and time-frequency-domain features of all channels, they are spliced ​​and combined into a one-dimensional real number array according to a predetermined channel arrangement order to generate a feature vector.

3. The method according to claim 1, characterized in that, The calculation of global importance weights based on the mutual information between category labels and feature vectors includes: During the training phase, the continuous features of the feature vector in the training set are discretized by interval binning. The joint probability distribution of the j-th feature of the discretized feature vector and the corresponding category label is calculated. The marginal probability distribution of the j-th feature and the marginal probability distribution of the category label are calculated respectively. The mutual information value corresponding to the j-th feature is calculated. The global importance weights are obtained by normalizing the mutual information values ​​of each dimension.

4. The method according to claim 2 or 3, characterized in that, The step of combining global importance weights with local feature divergence to generate adaptive feature weight coefficients and performing a weighted transformation on the feature differences includes: The variance of the j-th dimension feature of the training sample corresponding to the neuron in the pattern layer is calculated and defined as the local feature divergence among the k nearest neighbor samples of the same type in the training set. Adaptive feature weight coefficients are generated based on the global importance weights and local feature divergences of each feature dimension. The feature dimension with the larger global importance weight and the smaller local feature divergence corresponds to the larger adaptive feature weight. The weights of each dimension are normalized so that the sum of the adaptive feature weights of all feature dimensions is 1. Calculate the difference between the feature vector to be tested and the feature vector of the corresponding training sample of the pattern layer neuron in each dimension, and multiply each difference by the square root of the corresponding adaptive feature weight coefficient to obtain the weighted difference vector.

5. The method according to claim 1, characterized in that, The step of forming a quadratic form based on the weighted difference vector (after adaptive feature weight coefficient weighting transformation) between the feature vector to be tested and the corresponding training samples of the pattern layer neurons, and the inverse of the regularized local covariance matrix, and substituting the quadratic form into the Gaussian radial basis function to generate the output value includes: Obtain the k nearest neighbor samples with the smallest distance among similar training samples for the corresponding neurons in the pattern layer. Perform a weighted transformation on the feature vectors of these k nearest neighbor samples according to adaptive feature weight coefficients. Calculate the arithmetic center coordinate vectors of the k nearest neighbor samples after the weighted transformation and obtain the covariance matrix. Add a regularization coefficient to the diagonal of the covariance matrix to obtain a regularized local covariance matrix. Calculate the inverse of the regularized local covariance matrix as the matrix weight parameter C. Calculate the output value of the radial Gaussian kernel of the network layer according to the following formula. ,in, Let C be the difference vector between the feature vector to be tested and the feature vector of the corresponding training sample of the pattern layer neuron after weighting by adaptive feature weight coefficients, and let C be the inverse of the regularized local covariance matrix.

6. The method according to claim 1, characterized in that, The calculation of boundary confidence based on the average distance between the training sample corresponding to the neuron and the k nearest out-of-class samples includes: Calculate the Euclidean distance between the training sample corresponding to the i-th pattern layer neuron and all out-of-class training samples, extract the k out-of-class samples with the smallest distance, calculate the arithmetic mean of the distances of these k out-of-class samples as the out-of-class neighbor distance mean, and perform normalization to obtain the boundary confidence.

7. The method according to claim 1, characterized in that, The step of extracting the geometric principal axis direction and size from the principal component analysis of connected regions with probabilities higher than a preset threshold, and then using the grid point probability values ​​for normalized weighted calculation to obtain the three-dimensional coordinates of the crack center includes: A set of three-dimensional meshes with posterior probabilities greater than a preset threshold is selected. The set of three-dimensional meshes is divided into at least one high-probability connected region according to spatial adjacency. For the r-th high-probability connected region, the three-dimensional coordinate matrix of the meshes within the high-probability connected region is extracted. The covariance matrix of the coordinate matrix is ​​decomposed into eigenvalues. The direction of the eigenvector corresponding to the largest eigenvalue is taken as the geometric principal axis direction of the crack in the high-probability connected region. The extreme difference of the projection of the coordinate matrix in the principal axis direction is taken as the size of the crack. The three-dimensional coordinates of each mesh point in the r-th high-probability connected region are multiplied by the corresponding posterior probability value, summed, and divided by the sum of the posterior probability values ​​of all mesh points in the high-probability connected region to obtain the core three-dimensional coordinates of the crack.

8. An ultrasonic positioning and detection system for internal cracks in dam concrete, characterized in that, include: The acquisition module is used to acquire monitoring signals from multiple ultrasonic sensors, extract time-domain and frequency-domain features of the monitoring signals and perform dimensionless processing to form feature vectors, pre-divide the dam detection area into a three-dimensional grid, construct a training set containing feature vectors and category labels, and establish a probabilistic neural network. The training module is used to calculate the global importance weights based on the mutual information of class labels and feature vectors during the training phase. For each neuron in the pattern layer, the variance of the k nearest neighbor sample set of the corresponding training sample in the same class is calculated as the local feature divergence. The global importance weights and local feature divergence are combined to generate adaptive feature weight coefficients, and the feature differences are weighted and transformed. The localized covariance matrix of the Gaussian radial basis kernel function is obtained by regularizing the covariance matrix of the k nearest neighbor sample set. The boundary confidence is calculated based on the average distance between the training sample corresponding to the neuron and the k nearest out-of-class samples. The detection module is used to form a quadratic form based on the weighted difference vector between the feature vector to be tested and the corresponding training samples of the pattern layer neurons after adaptive feature weight coefficient weighting transformation, and the inverse matrix of the regularized local covariance matrix. The quadratic form is substituted into the Gaussian radial basis kernel function to generate the output value and multiplied by the boundary confidence weight. In the summation layer, the weighted output value is accumulated according to the category label and normalized to obtain the posterior probability of the three-dimensional mesh crack and generate a three-dimensional probability field. For connected regions with probabilities higher than a preset threshold, principal component analysis is used to extract the geometric principal axis direction and size, and the three-dimensional coordinates of the crack center are calculated by normalizing and weighting the probability values ​​of the grid points.

9. The system according to claim 8, characterized in that, The extraction of time-domain and frequency-domain features of the monitoring signal includes extracting the wavefront arrival time, wavefront amplitude, and signal duration as time-domain features based on the time-domain waveform or envelope signal of the monitoring signal obtained by the ultrasonic sensor. Spectral analysis is performed on the monitoring signal to calculate the peak frequency and the energy ratio of the main frequency band within the set frequency band as frequency domain features. The energy entropy of the specified wavelet packet nodes is extracted using the wavelet packet decomposition algorithm as time-frequency domain features. The time domain, frequency domain, and time-frequency domain features of all extracted channels are dimensionless and then spliced ​​and combined into a one-dimensional real number array according to the predetermined channel arrangement order to generate feature vectors.

10. The system according to claim 8, characterized in that, The calculation of global importance weights based on the mutual information between category labels and feature vectors includes: During the training phase, the continuous features of the feature vector in the training set are discretized by interval binning. The joint probability distribution of the j-th feature of the discretized feature vector and the corresponding category label is calculated. The marginal probability distribution of the j-th feature and the marginal probability distribution of the category label are calculated respectively. The mutual information value corresponding to the j-th feature is calculated. The global importance weights are obtained by normalizing the mutual information values ​​of each dimension.

Images