Explosive cladding plate blank performance detection system

Abstract

The invention relates to an explosive composite slab performance detection system, in particular to the field of composite slab performance detection, linear and nonlinear ultrasonic complementary information is obtained through multi-physics field excitation and synchronous acquisition, a rich data basis is provided for accurate detection, and an adopted multi-modal depth feature fusion and cross-modal attention mechanism has a good application prospect. The method can effectively improve the recognition and distinguishing capability of complex interface tiny defects, especially weak combination defects, remarkably reduces the number of calibration samples and model adjustment time required for detection of a new material plate blank based on a rapid adaptive mechanism of meta-learning, enhances generalization, and enables a final physical constraint quantization network to be more accurate. The defect parameter inversion result is ensured to conform to the data statistics law and the acoustic physics principle, so that efficient and reliable detection from qualitative positioning to accurate quantitative evaluation on the interface defect of the explosive cladding plate blank is realized on the whole.

Description

Technical Field

[0001] This invention relates to the field of composite slab performance testing, and more specifically, to an explosive composite slab performance testing system. Background Technology

[0002] Exploded composite slabs, as a type of layered composite material that metallurgically bonds two or more dissimilar metal materials through explosive welding, are widely used in key pressure-bearing components and corrosion-resistant structures in fields such as petrochemicals, marine engineering, power equipment, and aerospace due to their ability to comprehensively leverage the performance advantages of each component. These components often endure extremely complex mechanical, temperature, and corrosive environments during service. The quality of their interfacial bonding is a fundamental factor determining the overall safety and service life of the equipment. Defects such as micron-level unbonded areas, weakly bonded areas, or oxide inclusions at the interface are initially insignificant, but under the combined effects of subsequent intense rolling deformation, phase transformation stress during heat treatment, and fatigue loads and media corrosion during long-term service, they are highly susceptible to defect propagation, ultimately leading to interlayer delamination failure and catastrophic accidents. Therefore, comprehensive and accurate non-destructive testing must be performed on the slabs before they leave the factory to prevent materials with potential defects from entering the manufacturing process. This places extremely stringent requirements on the identification sensitivity, quantification accuracy, and reliability of the testing technology.

[0003] Currently, ultrasonic nondestructive testing of the interface quality of such multilayer heterogeneous composite materials mainly relies on manually set amplitude thresholds to judge the A-scan echo signal, or on using time-frequency analysis methods such as wavelet transform to extract signal features for comprehensive interpretation by the testing personnel. This technical approach is effective for obvious defects such as macroscopic delamination, but it faces fundamental challenges when dealing with the specific object of exploded composite slabs. First, due to the significant differences in density and sound velocity among the various metal materials constituting the slab, severe acoustic impedance mismatch occurs. During the propagation of ultrasound, complex reflections, refractions, and mode conversions occur between the interfaces, generating strong structural noise that completely drowns out the weak defect echo signal, resulting in an extremely low signal-to-noise ratio. Second, weak bonding defects manifest as poor interfacial acoustic coupling rather than complete separation, with minimal changes in echo amplitude, which differs from the echo of the normal interfacial region. The differences in acoustic properties are only reflected in subtle features such as signal phase, spectral morphology, or nonlinear response. Traditional threshold judgment methods that rely on a single echo amplitude or transit time are extremely insensitive to these features, resulting in a high rate of missed detections. Furthermore, existing detection models or algorithms based on specific material pairings and optimized parameters have severely insufficient generalization ability. When the material combination, thickness ratio, or process parameters of the slab under inspection change, its acoustic response characteristics change accordingly, and the original model immediately becomes invalid. Extensive recalibration and parameter adjustments are necessary, which severely restricts detection efficiency and automation. Therefore, developing an intelligent signal processing and analysis method that can adaptively and automatically and accurately separate and quantify minute interface defects, especially weak bonding defects, from complex ultrasonic signals with high noise and multiple interferences has become an urgent need to break through current technical bottlenecks and achieve high-reliability quality inspection of exploded composite slabs. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an explosive composite slab performance testing system. The system utilizes a data acquisition module, a feature fusion module, an adaptive recognition module, and a quantization inversion module to solve the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes a data acquisition module, a feature fusion module, an adaptive recognition module, and a quantization inversion module connected in sequence, wherein; The data acquisition module is used to control the ultrasonic probe array to apply multi-physics field excitation to the slab under test when the test is started, and to simultaneously acquire the generated linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data in FMC format with full matrix capture. The linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data, along with the probe spatial coordinates and excitation parameters, are encapsulated into a raw multimodal data packet and sent to the feature fusion module. It is also used to acquire a calibration sample dataset with clear state labels for the known state region of the new material slab and send it to the adaptive recognition module. Feature fusion module: Receives the original multimodal data packets, performs phased array synthetic aperture focusing on the linear ultrasonic array waveform data to reconstruct the spatial acoustic image sequence, and performs high-order spectral analysis on the nonlinear ultrasonic harmonic response data to extract nonlinear feature vectors. A dual-channel depth feature extraction network is used to extract depth features from the spatial acoustic image sequence and the nonlinear feature vectors, respectively, to obtain spatial image feature vectors and spectral feature vectors. Finally, a cross-modal attention fusion unit fuses the spatial image feature vectors and spectral feature vectors to generate a fused feature vector, which is then sent to the adaptive recognition module. Adaptive Recognition Module: This module has a built-in defect recognition classifier pre-trained based on a model-independent meta-learning framework. After receiving the fused feature vector and the fused feature vector with state labels corresponding to the calibration sample dataset, the adaptive recognition module uses the fused feature vector with state labels to fine-tune the defect recognition classifier and generate an optimized classifier. The optimized classifier is then used to process the fused feature vector of the region to be tested, outputting a defect probability map and the corresponding defect category label. The defect probability map and defect category label are then associated with the linear ultrasonic array waveform data at the corresponding position in the original multimodal data packet and sent to the quantization inversion module. Quantization inversion module: It is used to locate the defect area based on the received defect probability map and defect category label, and extract the associated linear ultrasonic array waveform data; it uses a physical information constrained quantization regression network trained with the embedded vector of the extracted linear ultrasonic array waveform data and defect category label as input and a composite loss function including physical consistency loss term to process and output the quantitative physical parameters of the defect. In a preferred embodiment, the specific process of controlling the ultrasonic probe array to apply multi-physics field excitation to the slab under test and simultaneously acquiring the generated linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data in full matrix capture FMC format in the data acquisition module is as follows: First, two orthogonally encoded excitation signals are generated. The first signal is a linear scanning excitation signal group used to excite a wave field that can be processed by synthetic aperture. The ultrasonic probe array contains multiple independent array elements. The second signal is a high-energy, wideband nonlinear detection excitation signal used to excite the nonlinear response of the material. The linear scanning excitation signal group is loaded onto each element of the ultrasonic probe array and excited according to the full matrix capture protocol to generate a linear ultrasonic field. At the same time, the nonlinear detection excitation signal is loaded onto a pre-specified transmitting element or a group of parallel transmitting elements in the ultrasonic probe array for transmission. All receiving elements of the ultrasonic probe array synchronously acquire ultrasonic echo signals containing a mixture of linear and nonlinear response components. Utilizing the orthogonality of the two excitation signals in the frequency domain, the acquired ultrasonic echo signals are subjected to frequency domain separation processing to extract the linear response frequency band signal and the nonlinear response frequency band signal, respectively. Inverse transformations are performed on the extracted linear response frequency band signal and the nonlinear response frequency band signal to obtain the corresponding time-domain waveform data. This time-domain waveform data is the linear ultrasonic array waveform data and the nonlinear ultrasonic harmonic response data.

[0006] In a preferred embodiment, the specific process of encapsulating the linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data, along with the probe spatial coordinates and excitation parameters, into an original multimodal data packet is as follows: For a detection area, a complete set of linear ultrasonic array waveform data and the corresponding nonlinear ultrasonic harmonic response data obtained from this acquisition are combined as a data set. This data set, together with the spatial position coordinates of all ultrasonic probe array elements corresponding to the acquisition of this data set, and all excitation control parameters used to generate the linear scanning excitation signal set and the nonlinear detection excitation signal, including at least the first amplitude value, window function, focusing delay, center frequency, second amplitude value, pulse width factor and fundamental frequency, are encapsulated together and given a unified timestamp identifier and spatial index identifier to form a structured original multimodal data package.

[0007] In a preferred embodiment, the specific process of collecting and calibrating the sample dataset is as follows: When a new material slab needs to be inspected, the data acquisition module pre-selects and scans at least one area with a known intact interface and at least one area containing artificially pre-fabricated known-type defects on the new material slab. In each selected area, the specific process of controlling the ultrasonic probe array to apply multi-physics field excitation to the slab under test and simultaneously acquiring the generated full-matrix capture FMC format linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data is repeatedly executed, thereby generating multiple corresponding data combinations. Each data combination from a known state area is associated with a specific state label, which is used to indicate the actual physical state of its corresponding area interface. All data combinations associated with state labels together constitute the calibration sample dataset.

[0008] In a preferred embodiment, the feature fusion module utilizes a dual-channel depth feature extraction network to extract depth features from the spatial acoustic image sequence and the nonlinear feature vector, respectively, to obtain the spatial image feature vector and the spectral feature vector. The specific process is as follows: A phased array synthetic aperture focusing process is performed on the linear ultrasonic array waveform data extracted from the original multimodal data packet to reconstruct a spatial acoustic image sequence; bispectral analysis is performed on the extracted nonlinear ultrasonic harmonic response data to extract nonlinear feature vectors; the spatial acoustic image sequence is input into a three-dimensional convolutional neural network encoder to output a fixed-dimensional spatial image feature vector; the nonlinear feature vector is input into a multilayer perceptron encoder to output a fixed-dimensional spectral feature vector.

[0009] In a preferred embodiment, the specific process of fusing spatial image feature vectors and spectral feature vectors through a cross-modal attention fusion unit to generate a fused feature vector is as follows: The spatial image feature vector and the spectral feature vector are input into the cross-modal attention fusion unit. The cross-modal attention fusion unit first performs a linear transformation on the spatial image feature vector through a first set of learnable weight matrices to generate a query vector, key vector, and value vector corresponding to the spatial image feature vector, respectively. At the same time, it performs a linear transformation on the spectral feature vector through a second set of learnable weight matrices to generate a query vector, key vector, and value vector corresponding to the spectral feature vector, respectively. Then, the dot product between the query vector corresponding to the spatial image feature vector and the key vector corresponding to the spectral feature vector is calculated, and the dot product result is divided by a scaling factor, which is the square root of the dimension of the key vector, to obtain the unnormalized attention score. Next, the unnormalized attention score is normalized by applying the Softmax function to obtain the attention weights that characterize the importance of each component in the spectral feature vector to the spatial image feature vector. Finally, the attention weights are weighted and summed with the value vector corresponding to the spectral feature vector to obtain the first modality interaction vector. Simultaneously, the dot product between the query vector corresponding to the spectral feature vector and the key vector corresponding to the spatial image feature vector is calculated, and the dot product result is divided by a scaling factor, which is the square root of the dimension of the key vector, to obtain the unnormalized attention score. Next, the unnormalized attention score is normalized by applying the Softmax function to obtain the attention weights that characterize the importance of each component in the spatial image feature vector to the spectral feature vector. Finally, the attention weights are weighted and summed with the value vector corresponding to the spatial image feature vector to obtain the second modality interaction vector. Finally, the first modality interaction vector and the second modality interaction vector are added together to obtain a preliminary fusion vector; a first layer normalization operation is performed on the preliminary fusion vector; the result of the first layer normalization operation is input into a feedforward neural network for processing; the output of the feedforward neural network is added to the result of the first layer normalization operation; a second layer normalization operation is performed on the result of the addition, and the result of the second layer normalization operation is used as the final fusion feature vector.

[0010] In a preferred embodiment, the adaptive recognition module fine-tunes the defect recognition classifier pre-trained based on a model-independent meta-learning framework using a fused feature vector with state labels to generate an optimized classifier. The specific process is as follows: The adaptive recognition module first loads a pre-trained defect recognition classifier with a set of initial model parameters. It then receives fused feature vectors from the feature fusion module for all test regions of the slab to be inspected, and a dataset of calibration samples with explicit state labels, converted by the feature fusion module into fused feature vectors. The fine-tuning process uses an iterative gradient descent algorithm for optimization. Each iteration of the gradient descent algorithm involves: first, calculating the predicted output of the defect recognition classifier for the fused feature vectors of all calibration samples based on the current model parameters; second, calculating the difference between the predicted outputs of all calibration samples and their corresponding true state labels, quantifying this difference using a preset cross-entropy loss function, and taking the average of all sample differences as the total loss value; third, calculating the gradient of the total loss value relative to the current model parameters using the backpropagation algorithm; and finally, multiplying the gradient by a preset fast adaptive learning rate to obtain the parameter update amount, subtracting this update amount from the current model parameters to obtain the updated model parameters. This iterative process is repeated a preset finite number of times, and the resulting model parameters constitute the optimized classifier.

[0011] In a preferred embodiment, the specific process of using the optimized classifier to process the fused feature vector of the region to be tested, outputting a defect probability map and corresponding defect category labels, and associating the defect probability map and defect category labels with the linear ultrasonic array waveform data at the corresponding positions in the original multimodal data packet is as follows: The fused feature vectors of all test areas on the slab are sequentially input into the optimization classifier. For each input fused feature vector, the optimization classifier outputs a probability distribution vector containing multiple elements, the number of which is consistent with the preset total number of defect categories. The value of each element represents the predicted probability that the input fused feature vector belongs to a defect category, and the sum of all element values ​​is always 1. The element with the largest value is found from this probability distribution vector; the defect category corresponding to this element is the defect category label of the current test area, and the value of the largest element is the confidence level of the defect presence in the current test area. The final step is based on the spatial order of all test areas on the slab. The defect category label and its defect presence confidence of each region are spatially arranged to generate a defect probability map. Finally, the defect probability map and the defect category label corresponding to each region to be tested are used as the recognition result. Based on the spatial index information inherited from the original multimodal data package and corresponding to each fused feature vector, the recognition result of each region is associated and bound with the linear ultrasonic array waveform data segment stored in the original multimodal data package for that region, forming a set of associated data records. The associated data records contain the defect probability map, the defect category labels of all regions and their respective associated linear ultrasonic array waveform data indices, and are sent to the quantization inversion module as a complete result data package.

[0012] In a preferred embodiment, the specific process of locating the defect region and extracting the associated linear ultrasonic array waveform data based on the received defect probability map and defect category label in the quantization inversion module is as follows: First, the quantization inversion module receives and parses the result data packet from the adaptive identification module. Then, based on the defect probability map and defect category label in the result data packet, it filters out all defect regions that are determined to have defects, forming a set of defect regions to be quantized. Next, for each defect region in the set of defect regions to be quantized, the quantization inversion module uses the linear ultrasonic array waveform data index from the result data packet corresponding to that defect region to accurately locate and extract the corresponding data segment from the linear ultrasonic array waveform data of the original multimodal data packet. Finally, each extracted data segment is preprocessed to obtain the preprocessed defect region waveform signal.

[0013] In a preferred embodiment, the specific process of using a physical information-constrained quantitative regression network to process and output quantitative physical parameters of defects is as follows: First, the input features of the Physical Information Constraint Quantization Regression Network are constructed. This construction process is performed for each defect region in the set of defect regions to be quantized: feature extraction is performed on the preprocessed waveform signal of the defect region to obtain waveform features; at the same time, the defect category label corresponding to the defect region is input into a learnable embedding layer, which outputs a dense vector of fixed dimensions as the label embedding vector; finally, the waveform features and the label embedding vector are concatenated to form the final input features of the defect region. Then, the final input features are fed into the trained Physical Information Constrained Quantization Regression Network (PEFRN). Through its internal multi-layer nonlinear transformations, the PEFRN directly regresses and outputs the quantitative physical parameter vector of the defect region. This vector includes at least the equivalent size and burial depth of the defect. During training, the PEFRN employs a composite loss function for optimization. This function is a weighted sum of a data fitting loss term and a physical consistency loss term. The data fitting loss term, using a pre-defined first loss function, calculates the average difference between the predicted quantitative physical parameter vectors and the actual labeled parameters for all samples in a training batch. The physical consistency loss term forces the network-predicted parameters to conform to the physical laws of ultrasonic scattering. Finally, the quantitative physical parameter vector obtained for each defect region in the set of defect regions to be quantified is integrated with the corresponding defect category label, the defect presence confidence obtained from the defect probability map, and the spatial location information of the region to generate a structured quantitative defect report.

[0014] The beneficial effects of this invention are as follows: by using multi-physics field excitation and synchronous acquisition, complementary information of linear and nonlinear ultrasound is obtained, providing a rich data foundation for accurate detection. The multimodal deep feature fusion and cross-modal attention mechanism adopted can effectively improve the ability to identify and distinguish small defects in complex interfaces, especially weak bonding defects. The meta-learning-based fast adaptive mechanism significantly reduces the number of calibration samples and model adjustment time required for detecting new material slabs, enhancing generalization. The final physical constraint quantification network ensures that the defect parameter inversion results conform to both data statistical laws and acoustic physics principles, thus achieving efficient and reliable detection of interface defects in exploded composite slabs from qualitative location to precise quantitative assessment. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1-2 The explosion-prone composite slab performance testing system shown includes: a data acquisition module, a feature fusion module, an adaptive recognition module, and a quantization inversion module connected in sequence; wherein; The data acquisition module is used to control the ultrasonic probe array to apply multi-physics field excitation to the slab under test when the test is started, and to simultaneously acquire the generated linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data in FMC format with full matrix capture. The linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data, along with the probe spatial coordinates and excitation parameters, are encapsulated into a raw multimodal data packet and sent to the feature fusion module. It is also used to acquire a calibration sample dataset with clear state labels for the known state region of the new material slab and send it to the adaptive recognition module. Feature fusion module: Receives the original multimodal data packets, performs phased array synthetic aperture focusing on the linear ultrasonic array waveform data to reconstruct the spatial acoustic image sequence, and performs high-order spectral analysis on the nonlinear ultrasonic harmonic response data to extract nonlinear feature vectors. A dual-channel depth feature extraction network is used to extract depth features from the spatial acoustic image sequence and the nonlinear feature vectors, respectively, to obtain spatial image feature vectors and spectral feature vectors. Finally, a cross-modal attention fusion unit fuses the spatial image feature vectors and spectral feature vectors to generate a fused feature vector, which is then sent to the adaptive recognition module. Adaptive Recognition Module: This module has a built-in defect recognition classifier pre-trained based on a model-independent meta-learning framework. After receiving the fused feature vector and the fused feature vector with state labels corresponding to the calibration sample dataset, the adaptive recognition module uses the fused feature vector with state labels to fine-tune the defect recognition classifier and generate an optimized classifier. The optimized classifier is then used to process the fused feature vector of the region to be tested, outputting a defect probability map and the corresponding defect category label. The defect probability map and defect category label are then associated with the linear ultrasonic array waveform data at the corresponding position in the original multimodal data packet and sent to the quantization inversion module. Quantization Inversion Module: This module is used to locate the defect region based on the received defect probability map and defect category label, and extract the associated linear ultrasonic array waveform data. It then uses a physical information constrained quantization regression network trained with a composite loss function including a physical consistency loss term as input, taking the extracted linear ultrasonic array waveform data and the embedding vector of the defect category label as input, to process the data and output the quantitative physical parameters of the defect.

[0020] In this embodiment, it is specifically necessary to explain the process in the data acquisition module of controlling the ultrasonic probe array to apply multi-physics field excitation to the slab under test, and simultaneously acquiring the generated linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data in the full matrix capture FMC format: First, two orthogonally encoded excitation signals are generated. The first signal is a linear scanning excitation signal group used to excite a wavefield capable of synthetic aperture processing. The ultrasonic probe array contains multiple independent array elements. The excitation signal of each array element in this linear scanning excitation signal group is a narrowband pulse signal with a preset center frequency, modulated by a preset first amplitude value, a preset window function, and a focusing delay set for the corresponding array element. The specific values ​​of the first amplitude value, window function, focusing delay, and center frequency are pre-configured according to the acoustic characteristics of the material of the tested slab and the detection resolution requirements. For example, the center frequency can be selected in the range of 2MHz to 10MHz, and the focusing delay is dynamically calculated according to the array element position and detection depth. The second signal is used to excite... The high-energy, wideband nonlinear detection excitation signal of the material's nonlinear response is a single-cycle or short-pulse signal with a preset second amplitude value, a preset pulse width factor, and a preset fundamental frequency. The second amplitude value is usually higher than the first amplitude value to generate sufficient acoustic energy to excite the nonlinear effect. Its fundamental frequency can be selected as 1 / 2 or 1 / 3 of the linear excitation center frequency, and the pulse width factor determines the bandwidth of the excitation signal. The linear scanning excitation signal group and the nonlinear detection excitation signal satisfy the orthogonality condition in the time-frequency domain so that the acoustic waves excited by the two can be separated in the frequency domain and crosstalk between the transmission channels can be reduced. The orthogonality condition can be achieved by minimizing the overlap of the main lobes of the two signals, for example, through frequency separation or coded modulation. Subsequently, the linear scanning excitation signal group is loaded onto each element of the ultrasonic probe array. Through electronic switching or parallel transmission channels, each element is sequentially or in groups excited according to the full matrix capture protocol to generate a linear ultrasonic field. Simultaneously, the nonlinear detection excitation signal is loaded onto a pre-designated transmitting element or a group of parallel transmitting elements in the ultrasonic probe array for transmission. For the same physical detection point on the tested slab, the transmission time of the nonlinear detection excitation signal and the virtual transmission event formed in the linear scanning excitation signal group to achieve synthetic focusing of that physical detection point are strictly synchronized on the system clock. This synchronization accuracy must be better than 1 / 10 of a linear excitation cycle to ensure that the linear and nonlinear responses from the same physical point are aligned in the time domain. Then, while applying the linear scanning excitation signal group and the nonlinear detection excitation signal, all receiving elements in the ultrasonic probe array synchronously acquire the ultrasonic echo signal; the ultrasonic echo signal is a signal containing a mixture of linear and nonlinear response components. Utilizing the orthogonality of the two excitation signals in the frequency domain, the acquired ultrasonic echo signals are subjected to frequency domain separation processing: First, a time-domain to frequency domain transformation is performed on the ultrasonic echo signal acquired by each receiving array element to obtain its frequency domain representation. This transformation is preferably a Fast Fourier Transform (FFT). Then, in the frequency domain, the transformed signals are processed by a first complementary bandpass filter function and a second complementary bandpass filter function. The first complementary bandpass filter function is designed based on the spectral characteristics of the linear scanning excitation signal group, and its passband center frequency is similar to the preset center frequency. The first frequency bandpass filter is consistent with the frequency of the signal and is used to extract the linear response band signal from the signal represented in the frequency domain. Its passband width is usually 60%-80% of the center frequency to retain the main frequency component and suppress out-of-band noise. The second complementary bandpass filter function is designed based on the spectral characteristics of the nonlinear probe excitation signal. Its passband covers the preset fundamental frequency and its integer multiples of higher harmonic frequencies. It is used to extract the nonlinear response band signal from the signal represented in the same frequency domain. This filter function is designed as a combination of multiple sub-passbands, corresponding to the fundamental frequency, the second harmonic, and the third harmonic, respectively, to extract the nonlinear response of each order. Finally, the extracted linear response frequency band signal and nonlinear response frequency band signal are subjected to inverse frequency domain to time domain transformation to obtain the corresponding time domain waveform data. This time domain waveform data is the linear ultrasonic array waveform data and the nonlinear ultrasonic harmonic response data. The inverse transformation is preferably the inverse fast Fourier transform (IFFT). The specific process of encapsulating linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data, along with probe spatial coordinates and excitation parameters, into a raw multimodal data package is as follows: For a given detection area, a complete set of linear ultrasonic array waveform data and the corresponding nonlinear ultrasonic harmonic response data obtained from this acquisition are combined into a single data set. This data set, along with the spatial coordinates of all ultrasonic probe array elements acquired during the acquisition of this data set, and all excitation control parameters used to generate the linear scanning excitation signal set and the nonlinear detection excitation signal set (including at least the first amplitude value, window function, focusing delay, center frequency, second amplitude value, pulse width factor, and fundamental frequency), are encapsulated and assigned a unified timestamp and spatial index identifier to form a structured raw multimodal data packet. The structured data packet can use a predefined data format, for example, using a header to store the timestamp, spatial index, excitation parameters, and array geometry information, followed by segmented storage of linear and nonlinear waveform data matrices to ensure the integrity and consistency of data during subsequent transmission and retrieval between modules. The specific process of collecting and calibrating the sample dataset is as follows: When a new material slab needs to be inspected, the data acquisition module pre-selects and scans at least one region with a known intact interface and at least one region containing artificially prefabricated defects of a known type on the slab. In each selected region, the process of controlling the ultrasonic probe array to apply multi-physics field excitation to the slab under test and simultaneously acquiring the generated linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data in FMC format is repeated, thereby generating multiple corresponding data combinations. Each data combination from a known state region is associated with a specific state label, which is used to indicate the actual physical state of its corresponding interface. All data combinations associated with state labels constitute the calibration sample dataset. The type and size of the artificially prefabricated defects should cover the range of defects that may actually occur. For example, it may include unbonded areas of different diameters, weakly bonded areas of different areas, and simulated inclusions. The state labels are usually in vector or one-hot encoded form and are used to provide supervision information in subsequent model adaptation.

[0021] In this embodiment, it is specifically necessary to explain the process by which the dual-channel depth feature extraction network extracts depth features from the spatial acoustic image sequence and the nonlinear feature vector in the feature fusion module, thereby obtaining the spatial image feature vector and the spectral feature vector: First, linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data are extracted from the received raw multimodal data packets. Next, phased array synthetic aperture focusing (SAP) processing is performed on the extracted linear ultrasonic array waveform data: based on preset imaging grid points, for each transmitter-receiver combination in the linear ultrasonic array waveform data, the theoretical transit time of the sound wave from the transmitter element position to the imaging grid point and back to the receiver element position is calculated according to the propagation speed of the sound wave in the tested material. Based on the theoretical transit time, the signal amplitude is extracted from the corresponding signal. Finally, the signal amplitudes extracted from all transmitter-receiver combination signals for the same imaging grid point are superimposed and summed to obtain the pixel value of that imaging grid point. By performing this operation on all preset imaging grid points, a spatial acoustic image sequence is reconstructed. The spatial acoustic image sequence is a three-dimensional data volume, with two dimensions representing the width and depth of the imaging area, respectively. The third dimension is associated with the sound wave propagation time or a specified echo time window, which typically covers the time zone from the echo from the upper surface of the slab to the echo from the lower interface. To ensure complete interface information is included, a higher-order spectral analysis is performed on the extracted nonlinear ultrasonic harmonic response data to extract nonlinear feature vectors. This higher-order spectral analysis is a bispectral analysis, the process of which is as follows: First, a discrete-time Fourier transform is performed on the nonlinear ultrasonic harmonic response data to obtain the frequency domain representation of the signal; then, the bispectrum of the signal is calculated based on this frequency domain representation. The bispectrum is a two-dimensional function, the value of which is obtained by multiplying the frequency domain representation of the signal at one fundamental frequency component, the frequency domain representation at another frequency component, and the complex conjugate of the frequency domain representation at the sum of the fundamental frequency component and the other frequency component. The fundamental frequency component is the preset fundamental frequency of the nonlinear detection excitation signal in the data acquisition module, while the other frequency component traverses its harmonic frequency range; finally, preset statistical features are extracted from the calculated bispectrum. The statistical features include bispectral diagonal slice integrals or bispectral region energy. These statistical features are combined to form a fixed-dimensional nonlinear feature vector. This fixed dimension is preset according to the number of selected statistical features, for example, selecting the first N significant bispectral region energy values ​​to form an N-dimensional vector. The spatial acoustic image sequence reconstructed through phased array synthetic aperture focusing is input into a three-dimensional convolutional neural network encoder. This encoder automatically learns and extracts high-level abstract features characterizing the spatial geometry, distribution, and evolution of defects with depth or time by sequentially performing multiple layers of three-dimensional convolution, three-dimensional pooling, and nonlinear activation function operations. The final output is a fixed-dimensional spatial image feature vector. The encoder can contain multiple convolutional blocks, each consisting of a three-dimensional convolutional layer, a three-dimensional batch normalization layer, and an activation function layer. The preferred activation function is the ReLU function. Finally, a global average pooling layer flattens the 3D feature map into a spatial image feature vector. Simultaneously, the nonlinear feature vector extracted through bispectral analysis is input into a multilayer perceptron encoder. This encoder performs nonlinear mapping and dimensional transformation on the input nonlinear feature vector by sequentially executing multilayer fully connected operations and nonlinear activation function operations, ultimately outputting a fixed-dimensional spectral feature vector. The multilayer perceptron encoder may contain two or three hidden layers, each followed by an activation function, preferably a ReLU or Sigmoid function. The final output layer maps the feature dimension to a dimension that matches the spatial image feature vector. The specific process of fusing spatial image feature vectors and spectral feature vectors through a cross-modal attention fusion unit to generate a fused feature vector is as follows: The spatial image feature vector and spectral feature vector are input into a cross-modal attention fusion unit. This unit first performs a linear transformation on the spatial image feature vector using a first set of learnable weight matrices to generate query vectors, key vectors, and value vectors corresponding to the spatial image feature vectors. Simultaneously, it performs a linear transformation on the spectral feature vector using a second set of learnable weight matrices to generate query vectors, key vectors, and value vectors corresponding to the spectral feature vectors. The query vector, key vector, and value vector have the same dimension, and their dimension is less than or equal to the dimension of the input feature vector. For example, if the input feature vector is 256-dimensional, the projected query / key / value vector can be set to 64-dimensional. Then, the dot product between the query vector corresponding to the spatial image feature vector and the key vector corresponding to the spectral feature vector is calculated, and the dot product result is divided by a scaling factor, which is the square root of the dimension of the key vector, to obtain the unnormalized attention score. Next, the unnormalized attention score is normalized by applying the Softmax function to obtain the attention weights that characterize the importance of each component in the spectral feature vector to the spatial image feature vector. Finally, the attention weights are weighted and summed with the value vector corresponding to the spectral feature vector to obtain the first modality interaction vector. Simultaneously, the dot product between the query vector corresponding to the spectral feature vector and the key vector corresponding to the spatial image feature vector is calculated, and the dot product result is divided by a scaling factor, which is the square root of the dimension of the key vector, to obtain the unnormalized attention score. Next, the unnormalized attention score is normalized by applying the Softmax function to obtain the attention weights that characterize the importance of each component in the spatial image feature vector to the spectral feature vector. Finally, the attention weights are weighted and summed with the value vector corresponding to the spatial image feature vector to obtain the second modality interaction vector. Finally, the first modal interaction vector and the second modal interaction vector are added together to obtain a preliminary fusion vector. A first-layer normalization operation is performed on the preliminary fusion vector. The result of the first-layer normalization operation is input into a feedforward neural network for processing. The feedforward neural network typically consists of two linear transformation layers and an intermediate activation function layer to enhance the model's nonlinear expressive power. The output of the feedforward neural network is added to the result of the first-layer normalization operation. A second-layer normalization operation is performed on the result of the addition, and the result of the second-layer normalization operation is used as the final fusion feature vector. The dimension of the fusion feature vector is usually consistent with the output dimensions of the spatial image feature vector and the spectral feature vector, or it is designed and adjusted to a specific dimension for use by subsequent modules.

[0022] In this embodiment, it is specifically necessary to explain the process by which the adaptive recognition module fine-tunes the defect recognition classifier pre-trained based on the model-independent meta-learning framework using fused feature vectors with state labels to generate an optimized classifier. The adaptive recognition module first loads a pre-trained defect recognition classifier with a set of initial model parameters. It then receives fused feature vectors from the feature fusion module for all test regions of the slab to be inspected, and a dataset of labeled samples with defined state labels, converted by the feature fusion module into fused feature vectors. The fine-tuning process uses an iterative optimization algorithm based on gradient descent. Each iteration of the gradient descent algorithm involves: first, calculating the predicted output of the defect recognition classifier for the fused feature vectors of all labeled samples based on the current model parameters; then, calculating the difference between the predicted outputs of all labeled samples and their corresponding true state labels, which is determined by a preset cross-entropy loss. The loss function is quantized, and the average of the differences among all samples is taken as the total loss value. Then, the gradient of the total loss value with respect to the current model parameters is calculated using the backpropagation algorithm. Finally, a preset fast adaptive learning rate is used to multiply the gradient to obtain the parameter update amount, which is then subtracted from the current model parameters to obtain the updated model parameters. The fast adaptive learning rate is usually set between 0.001 and 0.01 to ensure the stability and fast convergence of the fine-tuning process. This iterative process is repeated a preset finite number of times, usually several to a dozen times. The model parameters obtained after the iteration constitute the optimized classifier. This fine-tuning method with a small number of iterations allows the system to adapt to new material slabs within minutes, greatly improving detection efficiency. The specific process of using an optimized classifier to process the fused feature vector of the region under test, outputting a defect probability map and corresponding defect category labels, and associating the defect probability map and defect category labels with the linear ultrasonic array waveform data at the corresponding positions in the original multimodal data package is as follows: The fused feature vectors of all test areas on the slab are sequentially input into the optimization classifier. For each input fused feature vector, the optimization classifier outputs a probability distribution vector containing multiple elements, the number of which is consistent with the preset total number of defect categories. The value of each element represents the predicted probability that the input fused feature vector belongs to a defect category, and the sum of all element values ​​is always 1. The element with the largest value is found from this probability distribution vector; the defect category corresponding to this element is the defect category label of the current test area, and the value of the largest element is the confidence level of the defect presence of the current test area. A confidence threshold can be set, for example, 0.7. When the value of the largest element is lower than this threshold, the area can be marked as 'questionable' or require further verification. Based on the spatial order of all test areas on the slab, the defect category label and / or the defect presence confidence level of each area are spatially arranged to generate a defect probability map. The image can be a single-channel image, where the pixel values ​​represent the defect category number; or it can be a multi-channel image, where one channel stores the maximum confidence value and another channel stores the corresponding category label. Finally, the defect probability map and the defect category label corresponding to each region to be tested are used as the recognition result. Based on the spatial index information inherited from the original multimodal data package and strictly corresponding to each fused feature vector, the recognition result of each region is associated and bound with the linear ultrasonic array waveform data segment stored in the original multimodal data package for that region, forming a set of associated data records. This associated data record contains the defect probability map, the defect category labels of all regions, and their respective associated linear ultrasonic array waveform data indices, and is sent to the quantization inversion module as a complete result data package. The defect probability map can be a single-channel image, where the pixel values ​​represent the defect category number; or it can be a multi-channel image, where one channel stores the maximum confidence value and another channel stores the corresponding category label.

[0023] In this embodiment, it is specifically necessary to explain the process in the quantization inversion module of locating the defect region and extracting the associated linear ultrasonic array waveform data based on the received defect probability map and defect category label: First, the quantization inversion module receives and parses the result data packet from the adaptive identification module. Then, based on the defect probability map and defect category label in the result data packet, it filters out all defect regions that are determined to have defects, forming a set of defect regions to be quantized. Next, for each defect region in the set of defect regions to be quantized, the quantization inversion module uses the linear ultrasonic array waveform data index from the result data packet corresponding to that defect region to accurately locate and extract the corresponding data segment from the original multimodal data packet. Finally, each extracted data segment is preprocessed, which includes at least signal normalization and time window truncation centered on the defect echo, to obtain the preprocessed defect region waveform signal. Signal normalization typically scales the waveform data amplitude to the range of [-1,1] or [0,1] to eliminate acquisition gain differences. The time window truncation operation requires setting the starting point and window length of the window function according to the estimated arrival time of the defect echo. The window length must be sufficient to contain the main scattering signal of the defect, for example, set to 10 to 20 times the period corresponding to the probe center frequency. The specific process of using a physical information-constrained quantitative regression network to process and output quantitative physical parameters of defects is as follows: First, the input features of the Physical Information Constraint Quantization Regression Network are constructed. This construction process is performed for each defect region in the set of defect regions to be quantized: Feature extraction is performed on the preprocessed waveform signal of the defect region to obtain waveform features. Feature extraction includes directly using the time-domain waveform sequence or converting the time-domain waveform sequence to the frequency domain or time-frequency domain to extract features, such as extracting the signal's spectral envelope, time-frequency plot, or wavelet coefficients as waveform features. Simultaneously, the defect category label corresponding to the defect region is input into a learnable embedding layer. This embedding layer outputs a dense vector of fixed dimensions as the label embedding vector. The output dimension of the embedding layer can be set according to the number of defect categories, for example, 8-dimensional or 16-dimensional. Finally, the waveform features and the label embedding vector are concatenated to form the final input features of the defect region. Then, the final input features are input into the trained Physical Information Constraint Quantization Regression Network. The constrained quantization regression network directly regresses the quantitative physical parameter vector of the defect region through its internal multi-layer nonlinear transformations. This vector includes at least the equivalent size and burial depth of the defect, and may also include parameters such as the defect's orientation and shape factor. During the training phase, the physical information constrained quantization regression network employs a composite loss function for optimization. This composite loss function is a weighted sum of a data fitting loss term and a physical consistency loss term. The data fitting loss term uses a pre-defined first loss function to calculate the average difference between the quantitative physical parameter vectors predicted by the physical information constrained quantization regression network and the true labeled parameters for all samples in a training batch. The first loss function is the Huber loss function, which includes a pre-defined threshold parameter to balance sensitivity to large and small errors. This threshold parameter is typically set in the range of 0.1 to 1.0, for example, set to 0.5. The physical consistency loss term is used to force the parameters predicted by the network to satisfy the physical laws of ultrasonic scattering. Its specific calculation method is as follows for each training sample in the training batch: First, extract or calculate a first feature vector representing the defect scattering field mode from the intermediate layer features of the physical information-constrained quantization regression network. This first feature vector is obtained based on the input features of the current sample and the current parameters of the physical information-constrained quantization regression network. Simultaneously, based on the quantitative physical parameter vector predicted by the physical information-constrained quantization regression network for the current training sample, the known acoustic properties of the material, and the probe parameters, calculate a theoretical second feature vector of the scattering field based on a preset ultrasonic scattering physical model. The ultrasonic scattering physical model can be an analytical model based on the Kirchhoff approximation or the Born approximation. Then, calculate the sum of squares of the differences in each corresponding dimension between the first and second feature vectors. Finally... The average of the sum of squares of all training samples in the training batch is taken as the value of the physical consistency loss term for that training batch. When constructing the composite loss function, the value of the data fitting loss term and the value of the physical consistency loss term are multiplied by a balance coefficient greater than zero and then added together. This balance coefficient is used to adjust the strength of the physical constraints and can be adjusted between 0.1 and 10.0 based on experimental results. Finally, the quantitative physical parameter vector obtained for each defect region in the set of defect regions to be quantified is integrated with the defect category label corresponding to that defect region, the defect existence confidence obtained from the defect probability map, and the spatial location information of that region to generate a structured quantitative defect report. The quantitative defect report presents the category, location, size, and burial depth information of each defect in the form of a list or labeled image. The quantitative defect report is the final output of the system and can be used to guide quality assessment and subsequent process decisions.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0030] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A performance testing system for exploded composite slabs, characterized in that, Specifically, it includes: The data acquisition module, feature fusion module, adaptive recognition module, and quantization inversion module are connected sequentially, among which; The data acquisition module is used to control the ultrasonic probe array to apply multi-physics field excitation to the slab under test when the test is started, and to simultaneously acquire the generated linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data in FMC format with full matrix capture. The linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data, along with the probe spatial coordinates and excitation parameters, are encapsulated into a raw multimodal data packet and sent to the feature fusion module. It is also used to acquire a calibration sample dataset with clear state labels for the known state region of the new material slab and send it to the adaptive recognition module. Feature fusion module: Receives the original multimodal data packets, performs phased array synthetic aperture focusing on the linear ultrasonic array waveform data to reconstruct the spatial acoustic image sequence, and performs high-order spectral analysis on the nonlinear ultrasonic harmonic response data to extract nonlinear feature vectors. A dual-channel depth feature extraction network is used to extract depth features from the spatial acoustic image sequence and the nonlinear feature vectors, respectively, to obtain spatial image feature vectors and spectral feature vectors. Finally, a cross-modal attention fusion unit fuses the spatial image feature vectors and spectral feature vectors to generate a fused feature vector, which is then sent to the adaptive recognition module. Adaptive Recognition Module: This module has a built-in defect recognition classifier pre-trained based on a model-independent meta-learning framework. After receiving the fused feature vector and the fused feature vector with state labels corresponding to the calibration sample dataset, the adaptive recognition module uses the fused feature vector with state labels to fine-tune the defect recognition classifier and generate an optimized classifier. The optimized classifier is then used to process the fused feature vector of the region to be tested, outputting a defect probability map and the corresponding defect category label. The defect probability map and defect category label are then associated with the linear ultrasonic array waveform data at the corresponding position in the original multimodal data packet and sent to the quantization inversion module. Quantization Inversion Module: This module is used to locate the defect region based on the received defect probability map and defect category label, and extract the associated linear ultrasonic array waveform data. It then uses a physical information constrained quantization regression network trained with a composite loss function including a physical consistency loss term as input, taking the extracted linear ultrasonic array waveform data and the embedding vector of the defect category label as input, to process the data and output the quantitative physical parameters of the defect.

2. The explosive composite slab performance testing system according to claim 1, characterized in that: In the data acquisition module, the specific process of controlling the ultrasonic probe array to apply multi-physics field excitation to the slab under test and simultaneously acquiring the generated linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data in full matrix capture FMC format is as follows: First, two orthogonally encoded excitation signals are generated. The first signal is a linear scanning excitation signal group used to excite a wave field that can be processed by synthetic aperture. The ultrasonic probe array contains multiple independent array elements. The second signal is a high-energy, wideband nonlinear detection excitation signal used to excite the nonlinear response of the material. A linear scanning excitation signal group is loaded onto each element of the ultrasonic probe array and excited according to the full matrix capture protocol to generate a linear ultrasonic field. At the same time, a nonlinear detection excitation signal is loaded onto a pre-specified transmitting element or a group of parallel transmitting elements in the ultrasonic probe array for transmission. All receiving elements of the ultrasonic probe array synchronously acquire ultrasonic echo signals containing a mixture of linear and nonlinear response components. Utilizing the orthogonality of the two excitation signals in the frequency domain, the acquired ultrasonic echo signals are subjected to frequency domain separation processing to extract the linear response frequency band signal and the nonlinear response frequency band signal, respectively. Inverse transformations are performed on the extracted linear response frequency band signal and the nonlinear response frequency band signal to obtain the corresponding time-domain waveform data. This time-domain waveform data is the linear ultrasonic array waveform data and the nonlinear ultrasonic harmonic response data.

3. The explosive composite slab performance testing system according to claim 2, characterized in that: The specific process of encapsulating the linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data, along with the probe spatial coordinates and excitation parameters, into a raw multimodal data package is as follows: For a detection area, a complete set of linear ultrasonic array waveform data and the corresponding nonlinear ultrasonic harmonic response data obtained from this acquisition are combined as a data set. This data set, together with the spatial position coordinates of all ultrasonic probe array elements corresponding to the acquisition of this data set, and all excitation control parameters used when generating the linear scanning excitation signal set and the nonlinear detection excitation signal, including at least the first amplitude value, window function, focusing delay, center frequency, second amplitude value, pulse width factor and fundamental frequency, are encapsulated together and given a unified timestamp identifier and spatial index identifier to form a structured original multimodal data package.

4. The explosive composite slab performance testing system according to claim 3, characterized in that: The specific process for collecting and calibrating the sample dataset is as follows: When a new material slab needs to be inspected, the data acquisition module pre-selects and scans at least one area with a known intact interface and at least one area containing artificially pre-fabricated known-type defects on the new material slab. In each selected area, the specific process of controlling the ultrasonic probe array to apply multi-physics field excitation to the slab under test and simultaneously acquiring the generated full-matrix capture FMC format linear ultrasonic array waveform data and nonlinear ultrasonic harmonic response data is repeatedly executed, thereby generating multiple corresponding data combinations. Each data combination from a known state area is associated with a specific state label, which is used to indicate the actual physical state of its corresponding area interface. All data combinations associated with state labels together constitute the calibration sample dataset.

5. The explosive composite slab performance testing system according to claim 4, characterized in that: In the feature fusion module, the specific process of extracting depth features from the spatial acoustic image sequence and nonlinear feature vector using a dual-channel depth feature extraction network to obtain spatial image feature vector and spectral feature vector is as follows: Phased array synthetic aperture focusing (SAP) processing is performed on the linear ultrasonic array waveform data extracted from the original multimodal data packets to reconstruct a spatial acoustic image sequence; bispectral analysis is performed on the extracted nonlinear ultrasonic harmonic response data to extract nonlinear feature vectors; the spatial acoustic image sequence is input into a three-dimensional convolutional neural network encoder to output a fixed-dimensional spatial image feature vector; the nonlinear feature vector is input into a multilayer perceptron encoder to output a fixed-dimensional spectral feature vector.

6. The explosive composite slab performance testing system according to claim 5, characterized in that: The specific process of fusing spatial image feature vectors and spectral feature vectors through a cross-modal attention fusion unit to generate a fused feature vector is as follows: The spatial image feature vector and the spectral feature vector are input into the cross-modal attention fusion unit. The cross-modal attention fusion unit first performs a linear transformation on the spatial image feature vector through a first set of learnable weight matrices to generate a query vector, key vector, and value vector corresponding to the spatial image feature vector, respectively. At the same time, it performs a linear transformation on the spectral feature vector through a second set of learnable weight matrices to generate a query vector, key vector, and value vector corresponding to the spectral feature vector, respectively. Then, the dot product between the query vector corresponding to the spatial image feature vector and the key vector corresponding to the spectral feature vector is calculated, and the dot product result is divided by a scaling factor, which is the square root of the dimension of the key vector, to obtain the unnormalized attention score. Next, the unnormalized attention score is normalized by applying the Softmax function to obtain the attention weights that characterize the importance of each component in the spectral feature vector to the spatial image feature vector. Finally, the attention weights are weighted and summed with the value vector corresponding to the spectral feature vector to obtain the first modality interaction vector. Simultaneously, the dot product between the query vector corresponding to the spectral feature vector and the key vector corresponding to the spatial image feature vector is calculated, and the dot product result is divided by a scaling factor, which is the square root of the dimension of the key vector, to obtain the unnormalized attention score. Next, the unnormalized attention score is normalized by applying the Softmax function to obtain the attention weights that characterize the importance of each component in the spatial image feature vector to the spectral feature vector. Finally, the attention weights are weighted and summed with the value vector corresponding to the spatial image feature vector to obtain the second modality interaction vector. Finally, the first modal interaction vector and the second modal interaction vector are added to obtain a preliminary fusion vector; the first layer normalization operation is performed on the preliminary fusion vector; the result of the first layer normalization operation is input into a feedforward neural network for processing; The output of the feedforward neural network is added to the result of the first layer normalization operation; A second layer normalization operation is performed on the summed result, and the result of the second layer normalization operation is used as the final fused feature vector.

7. The explosive composite slab performance testing system according to claim 6, characterized in that: In the adaptive recognition module, the specific process of fine-tuning the defect recognition classifier pre-trained based on the model-independent meta-learning framework using the fused feature vector with state labels to generate an optimized classifier is as follows: The adaptive recognition module first loads a pre-trained defect recognition classifier with a set of initial model parameters. It then receives fused feature vectors from the feature fusion module for all test regions of the slab to be inspected, and a dataset of labeled samples with defined state labels, converted by the feature fusion module into fused feature vectors. The fine-tuning process uses an iterative gradient descent algorithm for optimization. Each iteration of the gradient descent algorithm involves: first, calculating the predicted output of the defect recognition classifier for the fused feature vectors of all labeled samples based on the current model parameters; second, calculating the difference between the predicted outputs of all labeled samples and their corresponding true state labels, quantified by a pre-defined cross-entropy loss function, and taking the average of all sample differences as the total loss value; third, calculating the gradient of the total loss value relative to the current model parameters using the backpropagation algorithm; and finally, multiplying the gradient by a pre-defined fast adaptive learning rate to obtain the parameter update amount, and subtracting this update amount from the current model parameters to obtain the updated model parameters. This iterative process is repeated a predetermined number of times, and the model parameters obtained after the iterations are completed constitute the optimized classifier.

8. The explosive composite slab performance testing system according to claim 7, characterized in that: The specific process of using the optimized classifier to process the fused feature vector of the region under test, outputting a defect probability map and corresponding defect category labels, and associating the defect probability map and defect category labels with the linear ultrasonic array waveform data at the corresponding positions in the original multimodal data package is as follows: The fused feature vectors of all test areas on the slab are sequentially input into the optimization classifier. For each input fused feature vector, the optimization classifier outputs a probability distribution vector containing multiple elements, the number of which is consistent with the preset total number of defect categories. The value of each element represents the predicted probability that the input fused feature vector belongs to a defect category, and the sum of all element values ​​is always 1. The element with the largest value is found from this probability distribution vector; the defect category corresponding to this element is the defect category label of the current test area, and the value of the largest element is the confidence level of the defect presence in the current test area. The final step is based on the spatial order of all test areas on the slab. The defect category label and its defect presence confidence of each region are spatially arranged to generate a defect probability map. Finally, the defect probability map and the defect category label corresponding to each region to be tested are used as the recognition result. Based on the spatial index information inherited from the original multimodal data package and corresponding to each fused feature vector, the recognition result of each region is associated and bound with the linear ultrasonic array waveform data segment stored in the original multimodal data package for that region, forming a set of associated data records. The associated data records contain the defect probability map, the defect category labels of all regions and their respective associated linear ultrasonic array waveform data indices, and are sent to the quantization inversion module as a complete result data package.

9. The explosive composite slab performance testing system according to claim 8, characterized in that: In the quantization inversion module, the specific process of locating the defect region and extracting the associated linear ultrasonic array waveform data based on the received defect probability map and defect category label is as follows: First, the quantization inversion module receives and parses the result data packet from the adaptive identification module. Then, based on the defect probability map and defect category label in the result data packet, it filters out all defect regions that are determined to have defects, forming a set of defect regions to be quantized. Next, for each defect region in the set of defect regions to be quantized, the quantization inversion module uses the linear ultrasonic array waveform data index from the result data packet corresponding to that defect region to accurately locate and extract the corresponding data segment from the linear ultrasonic array waveform data of the original multimodal data packet. Finally, each extracted data segment is preprocessed to obtain the preprocessed defect region waveform signal.

10. The explosive composite slab performance testing system according to claim 9, characterized in that: The specific process of using a physical information-constrained quantitative regression network to process and output quantitative physical parameters of defects is as follows: First, the input features of the Physical Information Constraint Quantization Regression Network are constructed. This construction process is performed for each defect region in the set of defect regions to be quantized: feature extraction is performed on the preprocessed waveform signal of the defect region to obtain waveform features; at the same time, the defect category label corresponding to the defect region is input into a learnable embedding layer, which outputs a dense vector of fixed dimensions as the label embedding vector; finally, the waveform features and the label embedding vector are concatenated to form the final input features of the defect region. Then, the final input features are fed into the trained Physical Information Constrained Quantization Regression Network (PEFRN). Through its internal multi-layer nonlinear transformations, the PEFRN directly regresses and outputs the quantitative physical parameter vector of the defect region. This vector includes at least the equivalent size and burial depth of the defect. During training, the PEFRN employs a composite loss function for optimization. This function is a weighted sum of a data fitting loss term and a physical consistency loss term. The data fitting loss term, using a pre-defined first loss function, calculates the average difference between the predicted quantitative physical parameter vectors and the actual labeled parameters for all samples in a training batch. The physical consistency loss term forces the network-predicted parameters to conform to the physical laws of ultrasonic scattering. Finally, the quantitative physical parameter vector obtained for each defect region in the set of defect regions to be quantified is integrated with the corresponding defect category label, the defect presence confidence obtained from the defect probability map, and the spatial location information of the region to generate a structured quantitative defect report.

Images