Composite board interface defect classification and identification method and system based on echo feature fusion

Abstract

The invention belongs to the technical field of material ultrasonic detection, and relates to a composite board interface defect classification and identification method and system based on echo feature fusion, and the method comprises the steps: extracting ultrasonic echo multi-dimensional acoustic features; calculating a local acoustic feature aliasing degree; combining the acoustic impedance value, the local acoustic feature aliasing degree and the sample scarcity degree to generate a dynamic acoustic cost weight; reconstructing a random forest model, and constructing a cost-sensitive voting factor by using an information gain index and a cost weight; inputting the multi-dimensional acoustic features of the to-be-detected composite board, and executing weighted summation to obtain a classification result; according to the method, the feature aliasing degree can be evaluated more accurately, the sample scarcity and acoustic characteristics are comprehensively considered, the importance of high-risk defect features in decision making is ensured, the misjudgment problem is effectively solved, and the accuracy and reliability of composite board interface defect classification and recognition are improved.

Description

Technical Field

[0001] This invention belongs to the field of ultrasonic testing technology for materials, and relates to a method and system for classifying and identifying interface defects of composite plates based on echo feature fusion, which is used to detect impurities and quality defects present at the interface during the processing of metal composite plates. Background Technology

[0002] In the production of specialized metal composite plates such as titanium steel and zirconium steel, the explosive bonding process generates complex wavy interfaces at the material bonding points. To detect defects such as incomplete bonding, oxide inclusions, or microcracks at these interfaces, the industry commonly uses ultrasonic echo signals for non-destructive testing. However, due to the significant differences in the physical properties of interface defects, single-dimensional acoustic characteristics cannot fully characterize the multi-physics coupling properties of the defects, making it difficult for traditional methods to accurately distinguish between high-risk defects and ordinary interface fluctuations.

[0003] Chinese patent document CN114994177B describes an ultrasonic defect detection method applied to a composite material under test. The method includes: acquiring Lamb wave signals from the composite material; selecting two mutually orthogonal linear directions on the composite material as the direction of a stepping scan; performing ultrasonic detection along the stepping scan direction with a preset step size K; acquiring Lamb wave signals from the composite material; obtaining a defect index (DI) through the Lamb wave signals; defining a virtual synthetic aperture for each imaging point in the scanning area; and obtaining the defect location through the virtual synthetic aperture and the defect index DI. However, this patent document only performs shallow signal processing and fails to establish a deep correlation between physical properties and defect types in the feature space. This results in the weak features of high-risk defects being submerged by normal interface fluctuations, making accurate identification of high-risk defects impossible and leading to a high false negative rate.

[0004] Chinese patent document CN112633368B discloses a defect detection system and method for flat vibrating motors based on an improved multi-granularity cascaded forest. It combines multi-granularity cascaded forest with cascaded CatBoost. First, the multi-granularity cascaded forest structure is used to process the relationships between acquired electrical signal data points to obtain feature vectors. Then, cascaded CatBoost is used to classify the feature vectors to obtain the types of defects in the flat vibrating motor. However, this patent document does not optimize for the problem of aliasing in the ultrasonic echo characteristics of composite plates, and therefore cannot effectively distinguish between normal interface fluctuations and high-risk defect characteristics when processing composite plate defect detection.

[0005] Chinese patent document CN116642946B describes a method for detecting coatings by perpendicularly incident ultrasonic waves onto a coated sample. It collects the aliased echoes (PR) at the interfaces of the coupling medium / coating and coating / substrate. A fast Fourier transform (FFT) is performed on the PR to obtain the expanded phase ΦR. The first-order partial derivative of ΦR with respect to frequency f is used to construct the ultrasonic echo phase first derivative spectrum (UEPDS). Extreme frequencies are extracted from the UEPDS, and based on the theoretical relationship between extreme frequencies and interface stiffness Kn, the magnitude of interface stiffness is inverted, indirectly achieving a quantitative measurement of the coating / substrate interface bonding strength. However, this patent document only applies to the quantitative measurement of interface stiffness and is not used for classifying interface defects in composite panels. Furthermore, it does not consider the impact of sample scarcity on the classification results, making it unable to effectively identify high-risk defects in unbalanced industrial data.

[0006] In existing technologies, although some solutions attempt to fuse multi-dimensional features such as time domain, frequency domain, and energy, they fail to establish a deep correlation between physical properties and defect types in the feature space. This results in the weak features of high-risk defects being submerged by a large number of normal interface fluctuations. At the same time, since the proportion of high-risk defect samples is extremely small, in extremely imbalanced industrial data, existing fusion strategies do not consider the differences in the contribution of different features to the discrimination of high-risk defects.

[0007] Currently, the industry typically introduces random forest classification algorithms as core technologies. Although traditional random forest classification algorithms can accept multi-dimensional feature inputs, their core majority voting mechanism lacks adaptive evaluation of feature fusion quality, leading to high-risk defects at the physical level being ignored by the algorithm due to the scarcity of samples. In the feature space, normal fluctuations caused by the wavy structure of the interface overlap with the weak features of high-risk defects, making the fused features unable to effectively support the accurate identification of high-risk defects, thus affecting the accuracy of composite board interface defect classification and identification. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and solve the problems of shallow fusion mechanisms and the tendency of traditional random forest classification algorithms to miss judgments under extremely imbalanced data, which affect the accuracy of composite board interface defect classification and identification. The invention seeks to design and provide a composite board interface defect classification and identification method and system based on echo feature fusion.

[0009] To achieve the above-mentioned objectives, this invention provides a method for classifying and identifying interface defects in composite plates based on echo feature fusion, employing the following technical solution: A method for classifying and identifying defects at the interface of composite panels based on echo feature fusion includes: acquiring ultrasonic echo signals from various scanning points on the composite panel interface; extracting multidimensional acoustic features from the ultrasonic echo signals to form a training sample set and adding category labels; extracting the nearest neighbor relationships of each sample in the feature space within the training sample set, and calculating the local acoustic feature aliasing degree by combining the dispersion variance of the ultrasonic echo signals; acquiring the standard acoustic impedance value of the composite panel substrate and the actual inverted acoustic impedance value of the composite panel interface, and combining the local acoustic feature aliasing degree with the scarcity of the corresponding sample category in the global data. Dynamic acoustic cost weights are generated; the node splitting records of each decision tree in the random forest model are traced, and the information gain index obtained by each decision tree when splitting samples is extracted. The information gain index and dynamic acoustic cost weights are combined to construct a cost-sensitive voting factor, and the random forest model is reconstructed. Each multidimensional acoustic feature of the interface of the composite board to be tested is input into the reconstructed random forest model. The cost-sensitive voting factor is used to perform weighted summation on each multidimensional acoustic feature to obtain the final classification result of the scanning point corresponding to each multidimensional acoustic feature, so as to realize the defect classification and identification of the composite board interface.

[0010] This invention assesses the degree of aliasing between normal fluctuations and defect features in the feature space by calculating the local acoustic feature aliasing degree, accurately reflecting the impact of dispersion variance on feature discrimination and providing a reliable feature quality assessment basis for generating dynamic acoustic cost weights. By generating dynamic acoustic cost weights, it comprehensively considers sample scarcity and acoustic characteristics, accurately reflecting the discriminative importance of high-risk defect samples and providing a reasonable cost allocation basis for constructing cost-sensitive voting factors. By constructing cost-sensitive voting factors, it improves upon the traditional random forest algorithm, combining information gain indicators and dynamic acoustic cost weights to ensure that high-risk defect features receive sufficient attention in classification decisions, improving the accuracy of identifying scarce high-risk defects. Weighted summation classification based on cost-sensitive voting effectively solves the misjudgment problem caused by feature aliasing, improving the accuracy and reliability of composite board interface defect classification and identification.

[0011] The present invention describes the extraction of multidimensional acoustic features from ultrasonic echo signals to form a training sample set and the addition of category labels, comprising: performing short-time Fourier transform and wavelet packet decomposition on the ultrasonic echo signal to extract an initial acoustic feature set including echo peak value, center frequency shift, high-frequency attenuation coefficient, and frequency band energy proportion; standardizing the initial acoustic feature set to eliminate dimensional differences to obtain multidimensional acoustic features; adding defect category labels to each multidimensional acoustic feature based on the verification results of destructive testing; and mapping the multidimensional acoustic features with added category labels to the feature space to form a training sample set.

[0012] This invention achieves comprehensive capture of multidimensional acoustic features of ultrasonic echo signals by combining short-time Fourier transform with wavelet packet decomposition, ensuring the integrity of time-frequency domain information. Standardization effectively eliminates dimensional differences between different features, and the resulting training sample set accurately reflects the multi-physics coupling characteristics of composite plate interface defects, providing a reliable feature basis for subsequent classification and recognition.

[0013] The local acoustic feature aliasing degree described in this invention satisfies: In the formula, For the first Algebraic local acoustic features of each sample According to the first The number of nearest neighbor samples selected for each sample in the feature space. and The first The sample and the first Multidimensional acoustic features of neighboring samples and These are the normalized th The sample and the first The dispersion variance of the ultrasound echo signal corresponding to the scanning point of each nearest neighbor sample. To prevent hyperparameters with a denominator of 0, It is a natural exponential function. This is the Euclidean distance symbol.

[0014] This invention achieves a scientific assessment of local acoustic feature aliasing by constructing an exponential decay model that includes the ratio of feature distance and dispersion variance. It more accurately reflects the relationship between feature similarity and dispersion variance between neighboring samples. When the feature distance is close and the dispersion variance is small, the aliasing is high, thus effectively assessing the difficulty of distinguishing between normal fluctuations and defect features in the feature space.

[0015] The dynamic acoustic cost weights described in this invention satisfy the following: In the formula, For the first Dynamic acoustic cost weights for each sample For the first Algebraic local acoustic features of each sample This is the standard acoustic impedance value of the composite board matrix. For the first The actual inverted acoustic impedance value of each sample at the interface of the composite plate. The number of samples in the training sample set. For the first The number of categories to which each sample belongs. It is the natural logarithm function.

[0016] This invention achieves a scientific assessment of dynamic acoustic cost weights by constructing a product model that includes acoustic impedance differences and category scarcity. This more accurately reflects the impact of acoustic impedance changes on defect severity, and the category scarcity term ensures that rare and high-risk defects receive higher cost weights, thereby effectively balancing the impact of physical characteristic differences and uneven sample distribution.

[0017] The cost-sensitive voting factor described in this invention satisfies: In the formula, For the first Cost-sensitive voting factors for decision trees in the forest prediction phase. For the first The training sample set of decision trees, For the training sample set, the first Dynamic acoustic cost weights for each sample For the first The decision tree in the partitioning of the first The information gain metric of the node containing the sample during the splitting process. It is the hyperbolic tangent activation function.

[0018] This invention achieves scientific evaluation of cost-sensitive voting factors by constructing a summation model that includes dynamic acoustic cost weights and hyperbolic tangent terms of information gain. This ensures the importance of high-cost-weight samples and high-information-gain nodes in the voting process. The hyperbolic tangent function performs nonlinear normalization on the information gain, thereby effectively improving the discrimination accuracy of high-risk defects in random forest classification.

[0019] The method for obtaining the dispersion variance of the ultrasonic echo signal described in this invention is as follows: perform a short-time Fourier transform on the ultrasonic echo signal at each scanning point of the composite plate interface, calculate the initial standard deviation of the frequency components within each time window; calculate the ratio of the initial standard deviation to the global dispersion variance standard deviation of the training sample set, and obtain the dispersion variance of the dimensionless ultrasonic echo signal.

[0020] The method for obtaining the actual inverted acoustic impedance value of the composite plate interface described in this invention is as follows: the reflected wave amplitude and incident wave amplitude at the corresponding scanning point of each sample are measured by ultrasonic testing equipment, and the reflectivity is calculated; the reflectivity is substituted into the reflection coefficient formula for inversion calculation to obtain the actual inverted acoustic impedance value of the composite plate interface.

[0021] The present invention describes obtaining the final classification result of each multidimensional acoustic feature corresponding to the scanning point, including: comparing the weighted sum of the random forest calculations to determine the multidimensional acoustic feature to belong to each defect category, determining the defect category corresponding to the highest weighted sum as the true defect category of the scanning point corresponding to the multidimensional acoustic feature, and obtaining the final classification result of the scanning point corresponding to the multidimensional acoustic feature; the true defect category includes at least normal binding type, micro-inclusion type and high-risk defect type.

[0022] The present invention describes a defect classification and identification method for composite panel interfaces, comprising: based on the final classification results of all scanning points on the interface of the composite panel under test, performing composite panel-level judgment according to the safety priority principle: in response to the presence of only normal bonding types, the composite panel under test is judged to be a qualified product; in response to the detection of at least one high-risk defect type, the composite panel under test is judged to be unqualified; in response to the presence of only minor inclusion types and the proportion not exceeding a threshold, the composite panel under test is judged to be a qualified product; in response to the presence of only minor inclusion types and the proportion exceeding a threshold, the composite panel under test is judged to be a downgraded product.

[0023] This invention also provides a composite plate interface defect classification and identification system based on echo feature fusion, which adopts the following technical solution: A composite plate interface defect classification and identification system based on echo feature fusion includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the aforementioned composite plate interface defect classification and identification method based on echo feature fusion.

[0024] By adopting the above technical solution, a computer program is generated from the above-mentioned method for classifying and identifying interface defects of composite plates based on echo feature fusion, and stored in the memory so that it can be loaded and executed by the processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0025] Compared with the prior art, the present invention has the following technical effects: Firstly, addressing the issues of traditional multidimensional feature fusion failing to establish a deep correlation between physical properties and defect types, and the overlapping of features between normal interface wavy structures and high-risk defects, this invention extracts the nearest neighbor relationships of training samples in the feature space and combines this with the dispersion variance of ultrasonic echo signals to calculate the local acoustic feature aliasing degree. This accurately distinguishes the feature fluctuations of normal wavy interfaces from the weak feature overlap of high-risk defects, anchoring the multi-physics coupling characteristics of defects at the physical level. This prevents the weak features of high-risk defects from being submerged by a large number of normal interface fluctuations, laying a reliable physical correlation foundation for subsequent feature fusion and classification recognition.

[0026] Secondly, addressing the issue that high-risk defect samples account for a very small proportion and that existing fusion strategies do not consider the differences in feature discrimination contributions, this invention combines the acoustic impedance difference between the composite board substrate and the interface, the degree of local acoustic feature aliasing, and the scarcity of categories to generate dynamic acoustic cost weights. This assigns higher weights to scarce high-risk defect samples and samples with high feature aliasing, allowing the random forest algorithm to prioritize high-risk defects during training and classification. This breaks the bias of traditional random forests towards majority class samples and solves the problem of high-risk defects being ignored in extremely imbalanced data.

[0027] Third, addressing the shortcomings of traditional random forest majority voting mechanisms in lacking adaptive evaluation of feature fusion quality, this invention traces the node splitting records of each decision tree, extracts information gain indicators, and constructs a cost-sensitive voting factor by combining dynamic acoustic cost weights. This completes the reconstruction of the random forest model. Instead of using simple majority voting, it uses a weighted sum based on feature contribution and sample cost weights, making the classification results more biased towards the accurate identification of high-risk defects. This avoids misjudgments caused by normal fluctuations in the interface wavy structure, thus improving the accuracy of composite board interface defect classification and identification.

[0028] Fourth, it integrates multi-dimensional information such as the multi-dimensional acoustic characteristics of ultrasonic echoes, physical properties of acoustic impedance, feature spatial nearest neighbor relationships, and sample imbalance characteristics. It not only anchors the essence of defects based on physical characteristics, but also optimizes the classification logic through data features. It can more accurately distinguish high-risk defects such as non-bonding, oxide inclusions, and microcracks at the interface of metal composite plates from ordinary wavy interface fluctuations. Attached Figure Description

[0029] Figure 1 This is a flowchart of a composite plate interface defect classification and identification method based on echo feature fusion according to an embodiment of the present invention.

[0030] Figure 2 This is a frequency distribution histogram of the local acoustic feature aliasing degree of all samples in the training sample set in a composite plate interface defect classification and identification method based on echo feature fusion according to an embodiment of the present invention.

[0031] Figure 3 This is a line graph showing the value of the cost-sensitive voting factor of each decision tree in the reconstructed random forest model of a composite plate interface defect classification and identification method based on echo feature fusion in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram comparing the distribution of predicted defect categories on the test set of a composite plate interface defect classification and identification method based on echo feature fusion according to an embodiment of the present invention. The traditional random forest model and the reconstructed random forest model of the present invention are shown in the diagram. Detailed Implementation

[0033] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings.

[0034] Example 1: This invention discloses a method for classifying and identifying interface defects in composite plates based on echo feature fusion, referring to... Figure 1 This includes steps S001-S005: S001: Obtain the training sample set of the composite board interface.

[0035] Acquire ultrasonic echo signals from various scanning points on the composite plate interface, extract multidimensional acoustic features of the ultrasonic echo signals to form a training sample set, and add category labels.

[0036] Specifically, a digital ultrasonic flaw detector with a broadband ultrasonic probe is used to scan the composite plate, acquiring ultrasonic echo signals from various scanning points on the composite plate interface. Short-time Fourier transform and wavelet packet decomposition are performed on the ultrasonic echo signals to extract an initial acoustic feature set, including echo peak value, center frequency shift, high-frequency attenuation coefficient, and frequency band energy proportion. The initial acoustic feature set is then standardized, such as by Z-score normalization, to eliminate dimensional differences, resulting in multidimensional acoustic features. Based on the verification results of destructive testing, defect category labels are added to each multidimensional acoustic feature, including labels for normal bonding types, minor inclusion types, and high-risk defect types, such as non-bonding and through-cracks. The multidimensional acoustic features with added category labels are then mapped into a feature space to form a training sample set.

[0037] S002: Calculate the aliasing degree of local acoustic features.

[0038] It should be noted that the explosive bonding process of composite panels creates complex wavy structures at the interface. The ultrasonic echo characteristics of normal interfaces easily overlap with those of minute inclusions and high-risk defects in the feature space. The dispersion variance of the ultrasonic echoes effectively characterizes the phase distortion of ultrasound waves as they penetrate the composite panel interface. This parameter can assess the waveform differences of echo signals at different scanning points, providing a basis for distinguishing between normal fluctuations and defect features. Therefore, this step, by combining the nearest neighbor relationship in the feature space with the dispersion variance to calculate the local acoustic feature aliasing, can accurately assess the degree of feature confusion in the region where each sample is located, providing a precise feature quality reference for subsequent weight allocation.

[0039] Extract the nearest neighbor relationships of each sample in the training sample set in the feature space, and calculate the local acoustic feature aliasing degree by combining the dispersion variance of the ultrasonic echo signal.

[0040] Specifically, the local acoustic feature aliasing degree satisfies: ; In the formula, For the first Algebraic local acoustic features of each sample According to the first The number of nearest neighbor samples selected for each sample in the feature space is determined by grid search based on the data distribution. and The first The sample and the first Multidimensional acoustic features of neighboring samples and These are the normalized th The sample and the first The dispersion variance of the ultrasonic echo signal corresponding to the scanning point of each nearest neighbor sample is obtained by performing a short-time Fourier transform on the ultrasonic echo signal of each scanning point on the composite plate interface, calculating the initial standard deviation of the frequency component in each time window, and calculating the ratio of the initial standard deviation to the standard deviation of the global dispersion variance of the training sample set to obtain the dispersion variance of the ultrasonic echo signal after dimensionless processing. To prevent hyperparameters with a denominator of 0, in this embodiment... ; It is a natural exponential function. This is the Euclidean distance symbol.

[0041] Since dispersion variance reflects the degree of phase distortion of ultrasound waves when penetrating the interface between different media, the calculated local acoustic feature aliasing degree increases when two samples are extremely close in the feature space and their own waveform dispersion variance is large. As it rises, The increase in indicates that the high-risk defect features at this location have been deeply integrated and enveloped by the acoustic features of most normal samples, making it highly likely that subsequent decision trees will misjudge this region.

[0042] The above-mentioned relational structure is based on the Gaussian kernel function in machine learning and spatial statistics. In acoustics and wave physics, wave dispersion and phase distortion objectively reflect the inhomogeneity of the medium interface. The above-mentioned relational structure aims to rely on this objective acoustic law to achieve a scientific assessment of the degree of physical aliasing between normal wave and defect features in the feature space. Traditional RBF kernel functions usually use a constant variance as the denominator, which is detached from the physical signal itself and cannot characterize the true signal distortion at the composite plate interface. This invention uses target samples... Construct a local feature evaluation space centered on the molecule, and preserve the squared Euclidean distance between the feature vectors. The core change lies in the denominator: the product of the dispersion variance of the ultrasonic echo signals of the target sample and its nearest neighbors is introduced. This modification strictly follows the attenuation and distortion rules of physical signals. When two samples have extremely similar features and the physical signal phase distortion is extremely small, it indicates that the weak features of high-risk defects have been deeply wrapped and assimilated by the acoustic features of normal wave-like interfaces. The above-mentioned modification of adaptive physical signal quality more accurately assesses the degree of physical aliasing in the feature space where normal fluctuations and defect features are difficult to distinguish.

[0043] Eigenvectors in molecules Z-score normalization was performed during the feature extraction stage to eliminate the dimensions, therefore Dimensionless; dispersion variance in the denominator It is calculated by the ratio of the initial standard deviation to the standard deviation of the global dispersion variance, and has also undergone dimensionless processing; parameters To prevent the denominator from being zero for dimensionless hyperparameters, the natural exponent term is dimensionless overall, resulting in the final local acoustic aliasing degree. It is a dimensionless, purely numerical ratio.

[0044] like Figure 2 As shown in the figure, the horizontal axis represents the local acoustic feature aliasing value, and the vertical axis represents the sample frequency. It can be seen from the figure that the local acoustic feature aliasing of most samples is concentrated between 0.2 and 0.4, and the sample frequency is the highest in this range. As the aliasing value increases, the sample frequency gradually decreases, and only a small number of samples have an aliasing value exceeding 0.7. This distribution reflects that the feature confusion of most samples in the training samples is low, and only a few samples have features that are highly aliased with the features of other samples. These samples with high aliasing are the samples that are prone to misclassification during the classification process, and are also the samples that need to be focused on in the dynamic acoustic cost weight of this invention.

[0045] S003: Generate dynamic acoustic cost weights.

[0046] It should be noted that the type of defect at the composite panel interface directly affects the interface's acoustic impedance characteristics. The acoustic impedance of a normally bonded interface is close to that of the matrix, while high-risk defects, such as non-bonding or through-cracks, cause a significant drop in interface acoustic impedance. Furthermore, in industrial settings, the number of high-risk defect samples is far fewer than normal samples and samples with minor inclusions. This data imbalance causes traditional classification algorithms to favor the majority class, failing to effectively identify high-risk defects. Therefore, this step generates dynamic acoustic cost weights by combining acoustic impedance differences, local feature aliasing, and sample scarcity. This allows different samples to be assigned weights that match their defect risk level and data scarcity, enabling the classification algorithm to prioritize high-risk and easily confused samples, thus compensating for the shortcomings of traditional algorithms in defect identification under imbalanced data.

[0047] The standard acoustic impedance value of the composite plate matrix and the actual inverted acoustic impedance value of the composite plate interface are obtained. Combined with the local acoustic feature aliasing degree and the scarcity of the corresponding sample category in the global data, dynamic acoustic cost weights are generated.

[0048] Specifically, the dynamic acoustic cost weights satisfy: ; In the formula, For the first Dynamic acoustic cost weights for each sample For the first Algebraic local acoustic features of each sample This refers to the standard acoustic impedance value of the composite panel matrix. Taking titanium-steel composite panels as an example, steel is typically used as the matrix, and the acoustic impedance of steel is approximately... ; For the first The actual inverted acoustic impedance value of each sample at the interface of the composite plate is obtained by measuring the reflected wave amplitude and incident wave amplitude at the corresponding scanning point of each sample using an ultrasonic testing device, calculating the reflectivity, and substituting the reflectivity into the reflection coefficient formula for inversion calculation to obtain the actual inverted acoustic impedance value of the interface of the composite plate. The number of samples in the training sample set. For the first The number of categories to which each sample belongs. It is the natural logarithm function.

[0049] Among them, because high-risk defects such as gas interlayers or cracks are filled with air, their actual inverted acoustic impedance values ​​are extremely low, resulting in an extremely high acoustic impedance mismatch rate. Furthermore, the local acoustic aliasing of the sample itself... The larger the value, the higher the impedance mismatch rate reflected at the physical level, and the scarcer this type of sample is in the global statistics, the higher the calculated dynamic acoustic cost weight. The larger, By directly linking the level of physical danger with the scarcity of data, weight is added to the few samples that are being masked from the bottom up.

[0050] The above relationship originates from the cost-sensitive learning mechanism in machine learning and the physical laws of acoustic reflection and transmission in nondestructive testing of materials. According to these physical laws, the difference in acoustic impedance is the core physical quantity that determines the energy of the reflected echo when ultrasound passes through different media interfaces. The above relationship aims to directly map this objectively existing physical change into algorithm weights. Traditional classification weights often rely solely on the statistical proportion of data labels, i.e., category scarcity. Detached from the physical flaw detection properties of materials, this invention constructs a composite product model with both physical and data dimensions. Due to high-risk defects such as non-adhesion or vacuum microcracks, the internal medium causes a decrease in actual local acoustic impedance, resulting in a high impedance mismatch rate. This solution utilizes this physical law... The term is directly introduced into the formula, combined with the degree of aliasing of local features. This approach ensures that objects with more drastic drops in physical impedance, are more easily confused with normal features, and have scarcer samples receive higher penalty costs. Simultaneously, to prevent extreme drops in physical impedance or extreme scarcity from causing abnormal surges in the calculated absolute value weights, a natural logarithm function is nested in the outer layer. Nonlinear smooth scaling ensures the stability of the model's convergence to physical facts.

[0051] It is a dimensionless constant; the standard acoustic impedance value of the composite board matrix. Compared with the actual inverted acoustic impedance value The dimensions are consistent, and the absolute value of the difference between the two is divided by . This makes the acoustic impedance mismatch rate term strictly dimensionless; sample size and The ratio is also dimensionless; the product inside the logarithmic function is a dimensionless pure number, and the resulting dynamic acoustic cost weights are... It is a dimensionless weighting coefficient.

[0052] S004: Construct a cost-sensitive voting factor to complete the reconstruction of the random forest model.

[0053] It's important to note that the majority voting mechanism of traditional random forests doesn't consider the differences in discriminative importance among different samples. In industrial scenarios like composite panel interface defects, where samples are imbalanced and features are easily aliased, traditional mechanisms ignore the discriminative contribution of high-risk defect samples, failing to effectively identify high-risk defects masked by normal sample features. In contrast, the information gain of a decision tree node reflects its ability to distinguish sample categories; higher information gain indicates a more effective division of sample categories and a stronger ability to identify high-risk defects. Therefore, this step, by combining dynamic acoustic cost weights and information gain to construct a cost-sensitive voting factor, reconstructs the voting logic of the random forest. This allows decision trees with strong discriminative abilities against high-risk defects to receive higher voting weights, improving the classification algorithm's ability to identify high-risk defects and preventing them from being missed.

[0054] By tracing the node splitting records of each decision tree in the random forest model, extracting the information gain index obtained by each decision tree when splitting samples, and combining the information gain index with the dynamic acoustic cost weight, a cost-sensitive voting factor is constructed to complete the reconstruction of the random forest model.

[0055] Specifically, the cost-sensitive voting factor satisfies: ; In the formula, For the first Cost-sensitive voting factors for decision trees in the forest prediction phase. For the first The training sample set for each decision tree is different from the training sample set. For use in the first The training sample set of each decision tree; For the training sample set, the first Dynamic acoustic cost weights for each sample For the first The decision tree in the partitioning of the first The information gain index of the node containing the sample during the splitting process is obtained directly by tracing the node splitting records of the existing decision tree model. The sample at the th The complete decision path in a decision tree is the set of all split nodes traversed from the root node down to the final leaf node. The information gain value calculated for each split node during feature partitioning is extracted. The information gain values ​​of all nodes along the decision path are summed, and the sum is the information gain index of the sample in that decision tree. Using the cumulative approach can more comprehensively characterize the overall classification discrimination provided by the sample features in multiple node splits; It is the hyperbolic tangent activation function.

[0056] Among them, when a decision tree can highly accurately separate the weights with extremely high dynamic acoustic costs on its branch paths, When dealing with high-risk and easily confused defective samples, and when the splitting node gains a significant information gain, this insightful decision tree will be assigned a highly cost-sensitive voting factor. Conversely, the importance of decision trees that can only identify common features of most classes will be diminished.

[0057] The aforementioned relational architecture originates from the soft voting mechanism in ensemble learning and the information gain metric in information theory. Although the voting mechanism itself belongs to algorithmic logic, this invention aims to force the algorithm's decision-making direction to be anchored and guided by objective physical attributes, namely the severity of defects, by introducing pre-existing physical constraints. Traditional random forest majority voting uses an equal-vote mechanism, lacking a physically adaptive evaluation of feature fusion quality, and is prone to missing high-risk defects due to interference from normal physical fluctuations in the wavy structure. To break this rigid logic, this invention incorporates dynamic acoustic cost weights that integrate the physical characteristics of acoustic impedance and dispersion variance. In the evaluation at the decision tree level, the split information gain of the decision tree nodes is extracted. And using the hyperbolic tangent activation function Nonlinear soft constraints are imposed to prevent a single tree from monopolizing voting rights; multiplying the physical cost weight by the algorithm gain means that a decision tree that can accurately separate samples with high physical risk will be given a larger voting factor. Through the above modifications, at the macro level of the system, it is ensured that the voting logic of the classifier strictly follows the physical logic of material flaw detection.

[0058] Dynamic acoustic cost weighting Dimensionless; Information Gain It is usually measured in bits or kilobytes, while As a nonlinear normalization, the input hyperbolic tangent function Then, the output is limited to The dimensionless scalar of the interval; multiplying and summing the two, the final cost-sensitive voting factor is obtained. The weighting parameter is dimensionless.

[0059] like Figure 3 As shown in the figure, the horizontal axis represents the decision tree number, and the vertical axis represents the value of the cost-sensitive voting factor. It can be seen from the figure that there are significant differences in the cost-sensitive voting factor of different decision trees. Some decision trees have higher voting factors, while others have lower voting factors. This reflects the difference in importance of different decision trees in classification decision-making. Decision trees with higher voting factors have a stronger ability to identify high-risk defects and will occupy a higher weight in the final weighted summation classification.

[0060] S005: Obtain the final classification result to achieve defect classification and identification of the composite board interface.

[0061] It should be noted that the testing scenarios for the composite panels under test are mostly post-production quality acceptance or periodic inspections of in-service equipment. Accurate identification of the defect type at each scanning point is crucial, while strictly adhering to the safety-first principle to ensure that high-risk defects are not missed. Traditional classification results do not consider the differences in discriminative capabilities among different decision trees, failing to highlight the role of decision trees that contribute to the identification of high-risk defects. Therefore, this step, by performing a weighted summation on each multi-dimensional acoustic feature based on a cost-sensitive voting factor, integrates the discriminative results of all decision trees, while highlighting the contribution of decision trees with strong high-risk defect discrimination capabilities. Ultimately, this achieves accurate and safety-first defect classification and identification, meeting the quality control requirements of industrial scenarios.

[0062] Each multidimensional acoustic feature of the composite panel interface to be tested is input into the reconstructed random forest model. The multidimensional acoustic feature is weighted and summed according to the cost-sensitive voting factor to obtain the final classification result of the scanning point corresponding to each multidimensional acoustic feature, so as to realize the defect classification and identification of the composite panel interface.

[0063] Specifically, obtaining the final classification result for each multidimensional acoustic feature corresponding to the scanning point includes: By comparing the weighted summation results of the random forest algorithm to determine the multidimensional acoustic feature's classification into various defect categories, the defect category corresponding to the highest weighted summation result is determined as the true defect category of the scan point corresponding to the multidimensional acoustic feature. This yields the final classification result for the scan point corresponding to the multidimensional acoustic feature. The true defect categories include at least normal bonding types, minor inclusion types, and high-risk defect types. The weighted summation results satisfy the following: ; In the formula, The number of decision trees in the random forest model. For the first Cost-sensitive voting factors for decision trees in the forest prediction phase. For the first The output of the decision tree The original predicted category results of multidimensional acoustic features, This is an indicator function.

[0064] The above relationship originates from the majority voting and weighted voting mechanism of the random forest algorithm in ensemble learning, aiming to synthesize the prediction results of all decision trees in the entire forest and output the final classification decision; the majority voting formula of the traditional random forest is a simple frequency summation, which is equivalent to all trees having equal weights. In highly unbalanced industrial defect detection scenarios, such an absolutely fair mechanism can easily lead to the minority's high-risk defects being overwhelmed by the majority's normal interface fluctuations in voting. This invention abandons equal weighting and replaces the implicit constant weights in the traditional formula with cost-sensitive voting factors reconstructed in the pre-step process. Due to the core multiplier By deeply integrating the underlying physical laws such as the dramatic changes in acoustic impedance at the composite panel interface and the dispersion variance of ultrasonic echoes, the above relationship, when performing a weighted summation, is essentially performing an accumulation of physical hazard levels. The indicator function in the formula... As a precise Boolean switch, when the first Predicted category of trees With target category When consistent Otherwise In conjunction with the summation calculation, the judgment results of all trees are weighted and accumulated according to their ability to identify physical defects. Through the above reconstruction, the voting logic of the algorithm is changed, so that the output mathematical classification boundary can penetrate complex normal interface physical fluctuation signals, which is in line with the objective law and industrial needs of prioritizing safety and never missing high-risk physical defects in material flaw detection.

[0065] Cost-sensitive voting factor obtained beforehand Dimensionless weight parameters; indicator function The output is a pure numerical value. or It is also dimensionless; The number of decision trees is a constant; therefore, after multiplying and summing the terms, the final weighted sum is obtained. It is a dimensionless comprehensive judgment score, used for subsequent horizontal comparison and judgment of scores in various categories.

[0066] Specifically, the method for classifying and identifying defects at the interface of the composite board includes: Based on the final classification results of all scan points on the interface of the composite board under test, composite board-level determination is performed according to the principle of safety priority: The composite board under test is deemed to be qualified if only the normal bonding type exists. In response to the detection of at least one high-risk defect type, the composite board under test is determined to be unqualified; When only minor inclusions are present and their proportion is not higher than the threshold, the composite board under test is determined to be a qualified product. When only a small amount of inclusions are present and the proportion is higher than the threshold, the composite board under test is determined to be a downgraded product.

[0067] In this embodiment, the threshold is set to 5%. When the proportion of minute inclusions is ≤5%, it will not affect the structural load-bearing capacity and corrosion resistance of the composite plate, and will meet the usage requirements of mainstream industrial scenarios such as chemical industry and pressure vessels. Alternatively, the grade of qualified products can be further divided with reference to the relevant industry standards, which will not be elaborated here.

[0068] like Figure 4 As shown, the left side represents the predicted defect category distribution of the traditional random forest model, while the right side represents the predicted defect category distribution of the cost-sensitive random forest model reconstructed in this invention. It can be seen from the figure that the traditional random forest model did not predict high-risk defects in the test set, while the model of this invention predicted high-risk defects. At the same time, the prediction ratio of small inclusions was slightly improved, while the ratio of normal combinations was slightly decreased. This demonstrates that this invention can effectively identify high-risk defects that the traditional model cannot identify, thus improving the accuracy of composite board interface defect classification and identification.

[0069] This embodiment also discloses a composite plate interface defect classification and identification system based on echo feature fusion, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a composite plate interface defect classification and identification method based on echo feature fusion according to the present invention is implemented.

[0070] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A method for classifying and identifying interface defects in composite plates based on echo feature fusion, characterized in that, include: Acquire ultrasonic echo signals from each scanning point on the composite plate interface, extract multidimensional acoustic features of the ultrasonic echo signals to form a training sample set and add category labels; Extract the nearest neighbor relationships of each sample in the training sample set in the feature space, and calculate the local acoustic feature aliasing degree by combining the dispersion variance of the ultrasonic echo signal. The standard acoustic impedance value of the composite plate matrix and the actual inverted acoustic impedance value of the composite plate interface are obtained. Combined with the local acoustic feature aliasing degree and the scarcity of the corresponding sample category in the global data, dynamic acoustic cost weights are generated. By tracing the node splitting records of each decision tree in the random forest model, extracting the information gain index obtained by each decision tree when splitting samples, and combining the information gain index with the dynamic acoustic cost weight, a cost-sensitive voting factor is constructed to complete the reconstruction of the random forest model. Each multidimensional acoustic feature of the composite panel interface to be tested is input into the reconstructed random forest model. The multidimensional acoustic feature is weighted and summed according to the cost-sensitive voting factor to obtain the final classification result of the scanning point corresponding to each multidimensional acoustic feature, so as to realize the defect classification and identification of the composite panel interface.

2. The method for classifying and identifying interface defects of composite plates based on echo feature fusion according to claim 1, characterized in that, The process of extracting multidimensional acoustic features from ultrasonic echo signals to form a training sample set and adding category labels includes: Short-time Fourier transform and wavelet packet decomposition are performed on the ultrasonic echo signal to extract an initial acoustic feature set including echo peak value, center frequency shift, high-frequency attenuation coefficient, and frequency band energy ratio. The initial acoustic feature set is standardized to eliminate dimensional differences, resulting in multidimensional acoustic features; Based on the verification results of the destructive test, defect category labels are added to each multidimensional acoustic feature, and the multidimensional acoustic features with added category labels are mapped to the feature space to form a training sample set.

3. The method for classifying and identifying interface defects of composite plates based on echo feature fusion according to claim 1, characterized in that, The local acoustic feature aliasing degree satisfies: ； In the formula, For the first Algebraic local acoustic features of each sample According to the first The number of nearest neighbor samples selected for each sample in the feature space. and The first The sample and the first Multidimensional acoustic features of neighboring samples and These are the normalized th The sample and the first The dispersion variance of the ultrasound echo signal corresponding to the scanning point of each nearest neighbor sample. To prevent hyperparameters with a denominator of 0, It is a natural exponential function. This is the Euclidean distance symbol.

4. The method for classifying and identifying interface defects of composite plates based on echo feature fusion according to claim 1, characterized in that, The dynamic acoustic cost weights satisfy: ； In the formula, For the first Dynamic acoustic cost weights for each sample For the first Algebraic local acoustic features of each sample This is the standard acoustic impedance value of the composite board matrix. For the first The actual inverted acoustic impedance value of each sample at the interface of the composite plate. The number of samples in the training sample set. For the first The number of categories to which each sample belongs. It is the natural logarithm function.

5. The method for classifying and identifying interface defects of composite plates based on echo feature fusion according to claim 1, characterized in that, The cost-sensitive voting factor satisfies: ； In the formula, For the first Cost-sensitive voting factors for decision trees in the forest prediction phase. For the first The training sample set of decision trees, For the training sample set, the first Dynamic acoustic cost weights for each sample For the first The decision tree in the partitioning of the first The information gain metric of the node containing the sample during the splitting process. It is the hyperbolic tangent activation function.

6. A method for classifying and identifying interface defects in composite plates based on echo feature fusion according to claim 1 or 3, characterized in that, The dispersion variance of the ultrasonic echo signal is obtained as follows: Perform short-time Fourier transform on the ultrasonic echo signals at each scanning point on the composite plate interface, and calculate the initial standard deviation of the frequency components within each time window; The ratio of the initial standard deviation to the global dispersion variance standard deviation of the training sample set is calculated to obtain the dispersion variance of the dimensionless ultrasonic echo signal.

7. A method for classifying and identifying interface defects of composite plates based on echo feature fusion according to claim 1 or 4, characterized in that, The actual inverted acoustic impedance value of the composite plate interface is obtained as follows: The reflectivity is calculated by measuring the reflected wave amplitude and incident wave amplitude at the interface corresponding to the scanning point of each sample using an ultrasonic testing device. By substituting the reflectivity into the reflection coefficient formula, the actual inverted acoustic impedance value of the composite plate interface is obtained.

8. The method for classifying and identifying interface defects of composite plates based on echo feature fusion according to claim 1, characterized in that, The final classification result obtained for each multidimensional acoustic feature corresponding to the scan point includes: By comparing the weighted summation results of the random forest calculations that determine the multidimensional acoustic feature to belong to each defect category, the defect category corresponding to the highest weighted summation result is determined as the true defect category of the scan point corresponding to the multidimensional acoustic feature, thus obtaining the final classification result of the scan point corresponding to the multidimensional acoustic feature. The actual defect categories include at least normal bonding types, minor inclusion types, and high-risk defect types.

9. The method for classifying and identifying interface defects of composite plates based on echo feature fusion according to claim 8, characterized in that, The method for classifying and identifying defects at the interface of composite panels includes: Based on the final classification results of all scan points on the interface of the composite board under test, composite board-level determination is performed according to the principle of safety priority: The composite board under test is deemed to be qualified if only the normal bonding type exists. In response to the detection of at least one high-risk defect type, the composite board under test is determined to be unqualified; When only minor inclusions are present and their proportion is not higher than the threshold, the composite board under test is determined to be a qualified product. When only a small amount of inclusions are present and the proportion is higher than the threshold, the composite board under test is determined to be a downgraded product.

10. A composite plate interface defect classification and identification system based on echo feature fusion, characterized in that, include: The processor and memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a composite plate interface defect classification and identification method based on echo feature fusion according to any one of claims 1-9.

Images