Defect positioning method and system for bulletproof vest composite interlayer based on multi-modal nondestructive testing, electronic equipment and storage medium

Abstract

The application provides a bulletproof vest composite interlayer defect positioning method and system based on multi-modal nondestructive testing, electronic equipment and storage medium, relating to the technical field of nondestructive testing. The application synchronously applies ultrasonic guided waves, pulse eddy current and infrared thermal wave excitation signals to the bulletproof vest composite interlayer, collects multi-modal response signals at each position, and generates an original multi-modal response signal set. The original multi-modal response signal set is input into a pre-trained deep convolutional neural network to determine the target frequency band parameters and target noise suppression parameters of each modal response signal, and output adaptive modulation parameter groups. The original signal set is subjected to frequency domain filtering and noise suppression processing to generate an optimized signal set. The spatial frequency characteristics are input into a cross-modal feature fusion model. The attention mechanism in the model adaptively associates defect sensitive features of different modalities to generate a three-dimensional defect distribution map of the composite interlayer. Accordingly, millimeter-level precision defect positioning is achieved, which can improve the precision and efficiency of bulletproof vest composite interlayer defect positioning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nondestructive testing, and in particular to a bulletproof vest composite interlayer defect positioning method and system based on multi-modal nondestructive testing, an electronic device, and a storage medium. BACKGROUND

[0002] The bulletproof vest composite interlayer is the core protective structure of bulletproof equipment. Whether it has defects such as delamination, micro-cracks, and foreign inclusions inside directly determines the impact resistance and protection reliability of the bulletproof vest. If the defects are not detected in time, the interlayer structure may be broken instantly when impacted by a bullet or a bullet fragment, resulting in a failure of protection. Since the bulletproof vest composite interlayer is usually made of fiber-reinforced composite materials and a buffer layer, traditional destructive testing will directly damage the product. Therefore, nondestructive testing technology is needed to identify defects. At the same time, single modal detection technology has obvious limitations. It is necessary to combine multi-modal detection means to cover different types and depths of defects. In order to ensure the accuracy of subsequent protection performance evaluation, the defect positioning needs to reach millimeter-level accuracy to meet the technical requirements of bulletproof vest production quality inspection, regular maintenance during service, and other scenarios.

[0003] At present, the mainstream targeted solution for the above-mentioned needs is the "ultrasonic guided wave-infrared thermal wave dual-modal collaborative detection solution". This solution applies ultrasonic guided wave excitation and infrared thermal wave excitation to the bulletproof vest composite interlayer simultaneously, and uses a special sensor to collect the response signals of the two modalities. The collected response signals are subjected to frequency band filtering and conventional noise removal using preset fixed parameters, and then time-domain peak features and temperature change features are extracted, respectively. The two types of features are spliced into fusion features at a fixed ratio, which are input into a pre-trained support vector machine model to identify the defect area and generate a two-dimensional defect distribution image of the composite interlayer. Finally, according to the "two-dimensional image pixel-actual depth" conversion relationship set by humans in advance, the two-dimensional defect area is mapped to the defect position information in the three-dimensional space.

[0004] This existing solution has three key defects: first, the signal processing parameters lack dynamic adaptability. The fixed frequency band filtering and conventional noise removal parameters cannot be adjusted according to the response intensity differences of the composite interlayer at different positions, which may lead to excessive filtering of useful signals or noise residue, thereby reducing the accuracy of subsequent feature extraction. Second, the feature fusion method is too simple. The fixed ratio splicing method does not establish an adaptive correlation between the ultrasonic guided wave and the infrared thermal wave modalities, which cannot fully utilize the complementarity of the two modalities, and may lead to missing or false extraction of defect sensitive features. Third, the three-dimensional positioning accuracy is insufficient. The mapping of the two-dimensional image to the three-dimensional space relies on the conversion relationship set by humans in advance, which lacks adaptability to the actual thickness and material differences of the composite interlayer, resulting in a positioning error of more than 10 millimeters in the depth direction, which cannot meet the millimeter-level accuracy requirement of bulletproof vest defect positioning, and affects the accurate judgment of the defect hazard level. SUMMARY

[0005] The application aims to provide a bulletproof vest composite interlayer defect positioning method and system based on multi-modal non-destructive testing, an electronic device and a storage medium, to solve the problem of low precision and efficiency of bulletproof vest composite interlayer defect positioning caused by the lack of dynamic adaptation of parameters, the lack of self-adaptation of feature fusion and low three-dimensional positioning precision in the prior art.

[0006] To solve the above technical problems, in a first aspect, the application provides a bulletproof vest composite interlayer defect positioning method based on multi-modal non-destructive testing, comprising:

[0007] Synchronously applying ultrasonic guided wave signals, pulsed eddy current signals and infrared thermal wave excitation signals to the bulletproof vest composite interlayer, collecting multi-modal response signals at each position of the bulletproof vest composite interlayer, and generating an original multi-modal response signal set;

[0008] Inputting the original multi-modal response signal set into a pre-trained deep convolutional neural network, determining target frequency band parameters and target noise suppression parameters of each modal response signal, and outputting an adaptive modulation parameter group;

[0009] Based on the adaptive modulation parameter group, performing frequency domain filtering and noise suppression processing on the original multi-modal response signal set to generate an optimized signal set, and extracting spatial frequency features of each modal response signal in the optimized signal set;

[0010] Inputting the spatial frequency features into a cross-modal feature fusion model, adaptively associating defect sensitive features of different modalities through an attention mechanism in the cross-modal feature fusion model, generating a three-dimensional defect distribution map of the composite interlayer, and based on the three-dimensional defect distribution map, performing millimeter-level precision defect positioning.

[0011] In a second aspect, the application provides a bulletproof vest composite interlayer defect positioning system based on multi-modal non-destructive testing, comprising:

[0012] The acquisition module is configured to synchronously apply ultrasonic guided wave signals, pulsed eddy current signals and infrared thermal wave excitation signals to the bulletproof vest composite interlayer, collect multi-modal response signals at each position of the bulletproof vest composite interlayer, and generate an original multi-modal response signal set;

[0013] The output module is configured to input the original multi-modal response signal set into a pre-trained deep convolutional neural network, determine target frequency band parameters and target noise suppression parameters of each modal response signal, and output an adaptive modulation parameter group;

[0014] The optimization module is configured to perform frequency domain filtering and noise suppression processing on the original multi-modal response signal set based on the adaptive modulation parameter group to generate an optimized signal set, and extract spatial frequency features of each modal response signal in the optimized signal set.

[0015] generating a three-dimensional defect distribution map of the composite interlayer by using an attention mechanism in the cross-modal feature fusion model to adaptively associate defect-sensitive features of different modalities, and performing defect positioning with millimeter-level precision based on the three-dimensional defect distribution map.

[0016] In a third aspect, the present application provides an electronic device, comprising:

[0017] a memory for storing a computer program;

[0018] a processor for executing the computer program to implement the steps of the method for positioning defects in a composite interlayer of a bulletproof vest based on multi-modal non-destructive testing according to the first aspect described above.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored therein, wherein the computer program, when executed by a processor, can implement the steps of the method for positioning defects in a composite interlayer of a bulletproof vest based on multi-modal non-destructive testing according to the first aspect described above.

[0020] In the present application, a method for positioning defects in a composite interlayer of a bulletproof vest based on multi-modal non-destructive testing is provided, comprising: simultaneously applying ultrasonic guided wave signals, pulsed eddy current signals and infrared thermal wave excitation signals to the composite interlayer of the bulletproof vest, collecting multi-modal response signals at each position of the composite interlayer of the bulletproof vest, and generating an original multi-modal response signal set; inputting the original multi-modal response signal set into a pre-trained deep convolutional neural network, determining target frequency band parameters and target noise suppression parameters of each modal response signal, and outputting an adaptive modulation parameter group; performing frequency domain filtering and noise suppression processing on the original multi-modal response signal set based on the adaptive modulation parameter group, generating an optimized signal set, and extracting spatial frequency features of each modal response signal in the optimized signal set; inputting the spatial frequency features into a cross-modal feature fusion model, adaptively associating defect-sensitive features of different modalities by using an attention mechanism in the cross-modal feature fusion model, generating a three-dimensional defect distribution map of the composite interlayer, and performing defect positioning with millimeter-level precision based on the three-dimensional defect distribution map.

[0021] The bulletproof vest composite interlayer defect positioning method based on multi-modal nondestructive testing provided in the application can realize comprehensive detection of different types and depths of defects in the composite interlayer, provide a complete original data basis for subsequent signal processing and defect positioning, by synchronously applying ultrasonic guided wave signals, pulse eddy current signals and infrared thermal wave excitation signals to the bulletproof vest composite interlayer, collecting multi-modal response signals at each position to generate an original multi-modal response signal set, inputting the original multi-modal response signal set into a pre-trained deep convolutional neural network, determining target frequency band parameters and target noise suppression parameters of each modal response signal and outputting an adaptive modulation parameter group, which can dynamically adapt the signal processing parameters to the characteristics of each modal signal and avoid processing deviations caused by fixed parameters, performing frequency domain filtering and noise suppression processing on the original multi-modal response signal set based on the adaptive modulation parameter group to generate an optimized signal set, and extracting spatial frequency characteristics of each modal response signal, which can effectively remove signal noise, retain useful signal components and improve signal quality, providing accurate feature data for subsequent cross-modal feature fusion, inputting the spatial frequency characteristics into a cross-modal feature fusion model, adaptively associating different modal defect sensitive features with the help of an attention mechanism, generating a three-dimensional defect distribution map and realizing millimeter-level precision defect positioning, which can fully utilize the complementarity of multi-modal features, improve defect recognition accuracy and positioning precision, and meet the high-precision requirements of bulletproof vest composite interlayer defect detection.

[0022] Further, the original multi-modal response signal set is input into a pre-trained deep convolutional neural network, the processing channels corresponding to the ultrasonic guided wave signals, the pulse eddy current signals and the infrared thermal wave excitation signals are automatically matched through modal recognition of the input layer, then the corresponding modal signals are subjected to multi-layer convolution through the channels respectively to obtain ultrasonic frequency domain characteristics, eddy current frequency domain characteristics and thermal wave frequency domain characteristics, the response intensity values are calculated based on the fluctuation amplitudes and durations of the frequency domain characteristics in the corresponding frequency bands, the frequency band parameters of the modal response signals are determined in combination with a preset intensity level and intensity range mapping relationship, the target noise suppression parameters are determined according to the frequency band parameters and the filtering modes corresponding to the modal, and finally the adaptive modulation parameter group is obtained by integrating all the target frequency band parameters and target noise suppression parameters. The exclusive processing channels are matched through modal recognition to ensure that each modal signal is subjected to targeted processing and cross interference is avoided, the accurate frequency domain characteristics are extracted through multi-layer convolution to provide a reliable basis for parameter determination, the frequency band parameters are determined through the response intensity values and the preset mapping relationship, the noise suppression parameters are determined in combination with the modal exclusive filtering modes, the adaptive modulation parameter group can accurately adapt to the actual characteristics of each modal signal, effectively solve the limitations of fixed parameter processing, provide high-quality parameter support for subsequent efficient frequency domain filtering and noise suppression processing, and further improve the reliability of the overall detection process. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0024] Figure 1 A flowchart of a composite sandwich defect positioning method for body armor based on multi-modal nondestructive testing provided by the embodiments of the present application;

[0025] Figure 2 A specific implementation diagram of a composite sandwich defect positioning method for body armor based on multi-modal nondestructive testing provided by the embodiments of the present application;

[0026] Figure 3 A structural diagram of a composite sandwich defect positioning system for body armor based on multi-modal nondestructive testing provided by the embodiments of the present application. DETAILED DESCRIPTION

[0027] In order to solve the problems of no dynamic adaptation of parameters, no self-adaptation of feature fusion, and low three-dimensional positioning accuracy in the prior art, the embodiments of the present application provide a composite sandwich defect positioning method for body armor based on multi-modal nondestructive testing, which adopts the following design concept: first, simultaneously detect the composite sandwich of the body armor with three different signals, collect the feedback of each position of the sandwich to the three signals, and form an original signal set; then, put these original signals into a pre-trained deep convolutional neural network, let the network automatically analyze each signal, determine the frequency band that each signal should focus on and the interference parameters that should be filtered out, and obtain a set of processing parameters that can be flexibly adjusted according to the actual situation of the signal; then, use this set of flexible processing parameters to optimize the original signal, remove the interference and retain the useful signal, and extract the key information reflecting the defect from the optimized signal; finally, put these key information into a special fusion model, the model will automatically combine the advantages of different signals, draw a three-dimensional defect distribution map, and then accurately find the position of the defect according to the map, with an error controlled within millimeter level. This method solves the signal processing problem caused by fixed parameters by flexibly adjusting the processing parameters; solves the problem of insufficient information utilization and easy misjudgment by combining and intelligently fusing three signals; solves the problem of insufficient positioning accuracy by directly generating a three-dimensional defect map, which can better meet the needs of body armor defect detection.

[0028] For those skilled in the art to better understand the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0029] The core of the present application is to provide a kind of based on multi-modal nondestructive testing's body armor composite sandwich defect positioning method, the flow chart of one specific embodiment of the present application is as shown in Figure 1 The method comprises:

[0030] S11, ultrasonic guided wave signal, pulse eddy current signal and infrared thermal wave excitation signal are synchronously applied to the body armor composite sandwich, and the multi-modal response signals of each position of the body armor composite sandwich are collected to generate the original multi-modal response signal set.

[0031] Among them, the ultrasonic guided wave signal is the sound wave signal that can propagate in the body armor composite sandwich and reflect the internal structure state of the sandwich; the pulse eddy current signal is a transient electromagnetic signal, which can reflect the changes of metal components or structures in the sandwich through electromagnetic induction; the infrared thermal wave excitation signal is an infrared signal that can cause temperature changes on the surface and inside of the sandwich, and can reflect the defects of the sandwich through temperature response; the multi-modal response signal is the signal containing the structure information of each position of the body armor composite sandwich after receiving the above three signals, each signal corresponds to a group of response data; the original multi-modal response signal set is the complete signal set formed by arranging all the multi-modal response signals collected at each position of the sandwich in the form of "position-signal type-response data".

[0032] In the embodiments of the present application, first, the detection equipment is used to synchronously emit ultrasonic guided wave signals, pulsed eddy current signals and infrared thermal wave excitation signals to the composite interlayer of the body armor, to ensure that the three signals act on the interlayer at the same time, and to avoid response data deviation caused by different signal application times. For example, when detecting the composite interlayer of a certain brand of body armor, the three signal emission modules of the detection equipment are operated at the same time, and the three signals are applied to the center area and the edge area of the surface of the interlayer. The initial intensity is set to a conventional value suitable for the material of the interlayer of the brand. Then, the sensor array is used to collect the multi-modal response signals of the composite interlayer of the body armor at each position. When collecting, the sensors are arranged at reasonable position intervals according to the preset, to ensure that the entire detection area of the interlayer is covered. One sensor that can simultaneously receive the responses of the three signals is placed at each position, and the response data of each position to the three signals is recorded respectively. Finally, the multi-modal response signals at all collection positions are arranged in a unified format to generate a set of original multi-modal response signals. For example, the response data at each position is arranged in the order of "position number to ultrasonic guided wave response data to pulsed eddy current response data to infrared thermal wave response data", to form a complete set of original signals.

[0033] S12, inputting the set of original multi-modal response signals into a pre-trained deep convolutional neural network to determine target frequency band parameters and target noise suppression parameters of the multi-modal response signals, and outputting an adaptive modulation parameter group.

[0034] The set of original multi-modal response signals is a set of multi-modal response data at each position of the interlayer generated in step S11. The pre-trained deep convolutional neural network is an intelligent analysis model trained in advance with a large amount of multi-modal response signal data similar to the composite interlayer of the body armor, and has the ability to automatically identify and process different modal signals. The target frequency band parameters of the multi-modal response signals are the signal frequency ranges containing effective information determined for the response signals of each modality. The target noise suppression parameters of the multi-modal response signals are parameters for filtering interference signals determined for the response signals of each modality. The adaptive modulation parameter group is a parameter set formed by arranging the target frequency band parameters and the target noise suppression parameters of all modal response signals according to the signal type. The processing mode can be flexibly adjusted according to the characteristics of different modal signals.

[0035] In the embodiments of the present application, as Figure 2As shown, first, the original multi-modal response signal set generated by S11 is input to the pre-trained deep convolutional neural network through the data import module, ensuring that the format of the signal set matches the network input requirements. For example, when processing the original signal set of a certain brand of body armor, the response data in the signal set is converted into a numerical format recognizable by the network, and the deep convolutional neural network trained in advance with a large number of interlayer signals of different batches of this brand is imported. Then, through the signal recognition function inside the network, the ultrasonic guided wave response signal, the pulse eddy current response signal, and the infrared thermal wave excitation response signal in the original signal set are automatically distinguished, and the target frequency band parameters and the target noise suppression parameters of each type of signal are determined. For example, the network finds the appropriate frequency range containing interlayer structure information as the target frequency band parameter and the typical characteristics of the interference signal (such as a specific amplitude pulse) as the target noise suppression parameter for the ultrasonic guided wave response signal through multi-layer analysis, and similarly processes the other two signals. Finally, through the parameter integration module, the target frequency band parameters and the target noise suppression parameters of the three types of modal response signals are arranged in the format of "signal type-target frequency band parameter-target noise suppression parameter", and the adaptive modulation parameter set is output, for example, the parameters of ultrasonic guided wave, pulse eddy current, and infrared thermal wave are arranged respectively to form a parameter set that can be directly used for subsequent signal processing.

[0036] S13, based on the adaptive modulation parameter set, performs frequency domain filtering and noise suppression processing on the original multi-modal response signal set to generate an optimized signal set, and extracts the spatial frequency characteristics of each modal response signal in the optimized signal set.

[0037] Among them, the adaptive modulation parameter set is the set output by S12, containing the target frequency band parameters and the target noise suppression parameters of each modal signal; the original multi-modal response signal set is the original signal set generated by S11; the frequency domain filtering processing is a processing method that retains the frequency range containing effective information in the signal according to the target frequency band parameter, and removes the frequency components outside the range; the noise suppression processing is a processing method that filters out the interference components in the signal according to the target noise suppression parameter; the optimized signal set is a signal set after frequency domain filtering and noise suppression processing, in which the effective information is retained sufficiently and the interference components are few; the spatial frequency characteristics are feature data extracted from the optimized signal set, reflecting the change of signal frequency with the spatial position (lateral, longitudinal) of the interlayer, which can be used for subsequent defect identification.

[0038] In the embodiments of the present application, first, the target frequency band parameters of each modal signal in the adaptive modulation parameter group are called by the frequency domain filtering module to perform frequency domain filtering processing on the corresponding modal signals in the original multi-modal response signal set, and signals in the effective frequency range are retained. For example, when processing the original signal of a certain brand of body armor, the target frequency band parameters of the ultrasonic guided wave signal in the parameter group are called to retain the ultrasonic guided wave response data in the original signal in the frequency range, and remove the frequency components outside the range. Similarly, the corresponding signals are processed according to the target frequency band parameters of the pulse eddy current and infrared thermal wave in the parameter group; then the target noise suppression parameters of each modal signal in the adaptive modulation parameter group are called by the noise suppression module to perform noise suppression processing on the signals filtered in the frequency domain, and filter the interference components. For example, for the filtered ultrasonic guided wave signal, the pulse interference exceeding the amplitude threshold in the signal is removed according to the amplitude threshold in the parameter group, the short wave interference is removed according to the corresponding duration threshold for the pulse eddy current signal, and the abnormal interference is removed according to the corresponding temperature change rate threshold for the infrared thermal wave excitation signal. The optimized signal set is generated after processing; finally, the characteristic data of the signal frequency changing with the transverse and longitudinal position of the interlayer is extracted from each modal signal in the optimized signal set by the feature extraction module, and the spatial frequency characteristics of each modal response signal are obtained. For example, the signal frequency value at each transverse position and each longitudinal position is extracted from the optimized ultrasonic guided wave signal, and the law of the frequency changing with the position (such as the frequency at a position being lower than that at the adjacent position) is recorded as the spatial frequency characteristic of the ultrasonic guided wave signal. Similarly, the spatial frequency characteristics of the other two signals are extracted.

[0039] S14, input the spatial frequency characteristics into the cross-modal feature fusion model, adaptively associate the defect sensitive features of different modalities through the attention mechanism in the cross-modal feature fusion model, generate a three-dimensional defect distribution map of the composite interlayer, and perform defect positioning with millimeter-level precision based on the three-dimensional defect distribution map.

[0040] Among them, the spatial frequency characteristics are the characteristic data reflecting the frequency of each modal signal changing with the spatial position of the interlayer extracted in step S13; the cross-modal feature fusion model is a model that can combine the characteristic data of different modalities (ultrasonic guided wave, pulse eddy current, and infrared thermal wave) for analysis, and can judge defects by comprehensively analyzing multi-modal information; the attention mechanism is a functional module in the cross-modal feature fusion model, which can automatically pay attention to the characteristic data more important for defect judgment; the defect sensitive features of different modalities are specific features in each modal signal that can reflect the defects of the interlayer; the three-dimensional defect distribution map is a graph that can show the positions of the defects in the interlayer in the transverse, longitudinal and depth directions; and the defect positioning with millimeter-level precision refers to determining the specific position of the defect in the interlayer to an accuracy of millimeter level.

[0041] In the embodiments of the present application, first, the modal spatial frequency features extracted by S13 are input into the cross-modal feature fusion model through the data input module, ensuring that the feature data format is consistent with the model input requirements, for example, when processing the feature data of a certain brand of body armor, the spatial frequency features of ultrasonic guided wave, pulse eddy current and infrared thermal wave signals are imported in the numerical format required by the model; then through the attention mechanism in the model, the defect sensitive features of different modalities are automatically analyzed and associated, for example, the model finds that the frequency of a certain position in the ultrasonic guided wave feature is abnormally reduced, and the temperature change rate of the position in the infrared thermal wave feature is abnormally slowed down, both of which are consistent with the characteristics of the interlayer defect, so the two features are associated, and the electromagnetic induction signal change of the position in the pulse eddy current feature is combined to determine the position of the defect in the interlayer depth direction; then through the graph generation function of the model, a three-dimensional defect distribution map containing the horizontal, vertical and depth positions of the defect is generated according to the associated defect sensitive features, for example, the horizontal position, vertical position and depth position of the defect are marked in the graph; finally, based on the three-dimensional defect distribution map, the specific position of the defect in the interlayer is determined through the positioning analysis module, ensuring that the positioning accuracy reaches millimeter level, for example, according to the coordinate information marked in the graph, the center position and coverage range of the defect are determined, and the error is controlled within a small millimeter range.

[0042] The present application provides the following specific examples: In the factory quality inspection link of a certain brand of bulletproof vest, the staff operates the adaptive multi-modal detection equipment, synchronously applies ultrasonic guided wave signals, pulse eddy current signals and infrared thermal wave excitation signals to the bulletproof vest composite interlayer of this brand composed of a specific fiber reinforced layer and a buffer layer, simultaneously controls the sensor array matched with the equipment, arranges the collection positions on the surface of the interlayer at a reasonable interval according to the preset, each sensor respectively collects the response data of the three signals at the corresponding position, after the collection is completed, all the response data of each position are recorded into the computer system according to the unified format, and the original multi-modal response signal set is generated. Then the staff converts the original multi-modal response signal set into a numerical format, imports it into the pre-trained deep convolutional neural network, the network automatically identifies the three modal response signals in the signal set, analyzes and determines the target frequency band parameters and target noise suppression parameters of each signal respectively, and then integrates these parameters according to the signal type and outputs the adaptive modulation parameter group. Next, the staff calls the adaptive modulation parameter group, processes the three modal signals in the original signal set through frequency domain filtering software respectively, retains the part of the effective frequency range and removes the invalid frequency components; then filters out the interference components in each signal according to the parameter group through the noise suppression software, generates the optimized signal set; then extracts the spatial frequency characteristics of each modal signal from the optimized signal set through the feature extraction software, records the law of frequency change with the interlayer position. Finally, the staff imports the extracted three modal spatial frequency characteristics into the cross-modal feature fusion model, the model determines the spatial position of the defect and generates a three-dimensional defect distribution map through the attention mechanism to associate different modal information of the defect characteristics, the staff determines the specific position of the defect in the interlayer based on the map, and completes the defect positioning operation in the whole quality inspection process.

[0043] By performing S11-S14, the embodiment of the application can comprehensively collect the multi-modal original data of the composite interlayer of the body armor by synchronously applying three signals and collecting responses at various positions, avoid detection blind spots caused by missing signal types or missing collection positions, and provide a complete data basis for subsequent processing; then, the adaptive modulation parameter set is generated by the deep convolutional neural network, which can dynamically match the processing parameters according to the actual characteristics of different modal signals, avoid the loss of effective information or the residual interference caused by fixed parameters, and improve the parameter adaptability; then, the original signals are filtered and denoised based on the adaptive parameters and spatial frequency features are extracted, which can effectively improve the signal quality and ensure that the extracted features clearly reflect the correlation between the interlayer position and the frequency, providing accurate feature support for defect identification; finally, the attention mechanism of the cross-modal feature fusion model is used to associate the defect-sensitive features and generate a three-dimensional defect distribution map, which can fully combine the advantages of different modal features, avoid misjudgment or omission caused by single modal analysis, and realize millimeter-level precision defect positioning, providing accurate location basis for judging whether the defect affects the protective performance of the body armor and formulating a repair scheme, and overall meeting the strict requirements of body armor quality inspection on defect detection comprehensiveness, accuracy and precision.

[0044] In a possible embodiment, S12, the original multi-modal response signal set is input to the pre-trained deep convolutional neural network to determine the target frequency band parameters and the target noise suppression parameters of each modal response signal, and an adaptive modulation parameter set is output, including:

[0045] Step 121, input the original multi-modal response signal set to the input layer of the deep convolutional neural network, and the input layer automatically matches the first processing channel corresponding to the ultrasonic guided wave signal, the second processing channel corresponding to the pulse eddy current signal, and the third processing channel corresponding to the infrared thermal wave excitation signal from the deep convolutional neural network through modal recognition.

[0046] The input layer of the deep convolutional neural network is the initial module of the network receiving external data, responsible for the preliminary reception and distribution of data, the modal recognition method is a recognition method that can distinguish different types of signals, which judges the signal type through the unique characteristics of the signal such as propagation speed and signal form, the first processing channel is a path in the deep convolutional neural network that is used to process ultrasonic guided wave response signals, the second processing channel is a path that is used to process pulse eddy current response signals, and the third processing channel is a path that is used to process infrared thermal wave excitation response signals. Each channel has a structure that is adapted to the processing needs of the corresponding signal, and the entire step ultimately realizes the matching of different signals with dedicated processing channels.

[0047] In the embodiment of the present application, first, the original multi-modal response signal set is input to the input layer of the deep convolutional neural network through the data transmission module, ensuring that the data format of the input is consistent with the receiving requirements of the input layer. Then, the input layer distinguishes the types of the signals in the original multi-modal response signal set through modal recognition. According to the unique characteristics of different signals such as propagation speed and signal form, it is determined which data belongs to ultrasonic guided wave signals, which belongs to pulse eddy current signals, and which belongs to infrared thermal wave excitation signals. Finally, according to the recognized signal types, the corresponding processing channels are automatically matched from the deep convolutional neural network. The response data corresponding to the ultrasonic guided wave signals is distributed to the first processing channel, the response data corresponding to the pulse eddy current signals is distributed to the second processing channel, and the response data corresponding to the infrared thermal wave excitation signals is distributed to the third processing channel. For example, in the factory quality inspection scene of a certain brand of bulletproof vest, after the original multi-modal response signal set containing three signal response data of each position of the composite sandwich of the brand is converted into a numerical format recognizable by the input layer, the input is completed. Then, through modal recognition, it is found that some data propagates at a relatively fast speed and fluctuates periodically, which is determined to be ultrasonic guided wave signals. Some data has electromagnetic induction characteristics, which is determined to be pulse eddy current signals. Some data is related to temperature change, which is determined to be infrared thermal wave excitation signals. Finally, the signal data of each type is transmitted to the corresponding processing channel.

[0048] Step 122, using the first processing channel, the second processing channel and the third processing channel, respectively, multi-layer convolution is performed on the ultrasonic guided wave response signal, the pulse eddy current response signal and the infrared thermal wave excitation response signal to obtain ultrasonic frequency domain features, eddy current frequency domain features and thermal wave frequency domain features.

[0049] Among them, multi-layer convolution is a processing method in deep convolutional neural network, which extracts key feature information in the signal through similar multi-layer screening operation, ultrasonic frequency domain features are key information related to frequency extracted from ultrasonic guided wave response signal, which can reflect the characteristics of ultrasonic guided wave response signal at different frequencies, eddy current frequency domain features are key information related to frequency extracted from pulse eddy current response signal, thermal wave frequency domain features are key information related to frequency extracted from infrared thermal wave excitation response signal, and finally three kinds of signal corresponding frequency domain features are obtained.

[0050] In the embodiments of the present application, first, the first processing channel, the second processing channel and the third processing channel are started, and the respective received ultrasonic guided wave response signals, pulse eddy current response signals and infrared thermal wave excitation response signals are read to ensure that the corresponding signal data is accurately obtained, then the corresponding signals are subjected to multi-layer convolution processing by each channel, each layer of convolution operation can filter out more critical information according to the signal characteristics, and gradually remove irrelevant or interfering information, finally, the corresponding frequency domain features are output after the multi-layer convolution of each channel, the first processing channel outputs ultrasonic frequency domain features, the second processing channel outputs eddy current frequency domain features, and the third processing channel outputs thermal wave frequency domain features, for example, in the scene of factory quality inspection of a certain brand of bulletproof vest, after the first processing channel reads the allocated ultrasonic guided wave response signal, it first filters out the interference components with too low frequency through the first layer of convolution, then refines the signal characteristics within the effective frequency range through the second layer of convolution, and finally extracts the frequency related information that can reflect the interlayer structure through the third layer of convolution, the other two channels also process the corresponding signals according to similar logic, and finally output the respective frequency domain features.

[0051] Step 123, based on the ultrasonic frequency domain features, the eddy current frequency domain features and the thermal wave frequency domain features, determining the target frequency band parameters and the target noise suppression parameters of each modal response signal.

[0052] The target frequency band parameter is a frequency range containing effective information determined for each modal response signal, which can reflect the structure or defect condition of the composite interlayer of the bulletproof vest to the greatest extent, and the target noise suppression parameter is a parameter determined for each modal response signal for filtering interference signals, which can remove interference components in the signal that do not belong to the target frequency band and may affect defect judgment, and the filtering method corresponding to different modalities is a signal filtering method designed according to the respective characteristics of the three signals, such as some suitable for filtering interference with fixed frequency, and some suitable for filtering interference with large fluctuation amplitude. The whole step finally determines the target frequency band parameters and the target noise suppression parameters of each modal response signal.

[0053] In the embodiments of the present application, firstly, the fluctuation amplitude and duration of each of the three frequency domain features are analyzed. The frequency band with large fluctuation amplitude and stable duration usually contains more effective information. Then, the target frequency band parameters of the modal response signals are determined according to the analysis results. The frequency band containing effective information is determined as the target frequency band parameter. Finally, the target noise suppression parameters are determined according to the target frequency band parameters of the modal response signals and the corresponding filtering mode of each modal. The filtering mode of different modal is adapted to the signal characteristics. For example, in the scene of factory quality inspection of a certain brand of bulletproof vest, it is found by analyzing the ultrasonic frequency domain feature that a certain frequency band has large fluctuation amplitude and stable duration. It is determined that the frequency band contains more interlayer structure information and is determined as the target frequency band parameter of the ultrasonic guided wave signal. Similarly, the target frequency band parameters of the other two signals are determined. Then, the corresponding target noise suppression parameters are determined according to the target frequency band parameters of each signal and the adapted filtering mode. For example, the filtering mode suitable for removing high-amplitude interference corresponds to setting an interference amplitude threshold, and the filtering mode suitable for removing short-time interference corresponds to setting an interference duration threshold.

[0054] Step 124, integrating the target frequency band parameters and the target noise suppression parameters of all modal response signals to obtain an adaptive modulation parameter group.

[0055] The target frequency band parameters of all modal response signals refer to the frequency range parameters of the three signals respectively containing effective information. The target noise suppression parameters of all modal response signals refer to the parameters of the three signals respectively used for filtering interference signals. The adaptive modulation parameter group is a parameter set formed by arranging the target frequency band parameters and the target noise suppression parameters of the three signals according to a unified rule. The set can provide adaptive processing parameters according to the characteristics of different modal signals. The set can be directly called for subsequent signal processing. The whole step finally obtains a complete adaptive modulation parameter group.

[0056] In the embodiments of the present application, firstly, the target frequency band parameters and the target noise suppression parameters of the three signals are collected to ensure that no parameters of any modal signal are missed. Then, a unified arrangement rule is determined. Usually, the parameters are arranged in the order of “modal signal type-target frequency band parameter-target noise suppression parameter” to ensure that the parameter set structure is clear and convenient for subsequent calling. Finally, all the collected parameters are integrated together according to the set arrangement rule to form a complete adaptive modulation parameter group. For example, in the scene of factory quality inspection of a certain brand of bulletproof vest, the two types of parameters of the three signals determined previously are collected. Then, the parameters are arranged in the order of signal type, target frequency band parameter and target noise suppression parameter. Finally, all the parameters are sequentially recorded according to the rule. After integration, the adaptive modulation parameter group is output.

[0057] The application provides the following specific examples: in the factory quality inspection process of a certain brand of body armor, the staff first converts the previously collected original multi-modal response signal set of the composite interlayer of the brand of body armor into a format recognizable by the input layer of the deep convolutional neural network, and then inputs the converted data set into the input layer of the network; the input layer starts the modal recognition function, distinguishes three types of signals by analyzing the propagation speed and morphological characteristics of the signals, and sends the signal data of each type to the corresponding processing channel. Subsequently, each processing channel reads the corresponding signal data and performs multi-layer convolution processing, the first processing channel processes ultrasonic guided wave response signals, the second processing channel processes pulse eddy current response signals, and the third processing channel processes infrared thermal wave excitation response signals. Each layer of convolution filters key information and removes interference, and finally outputs the frequency domain characteristics of the three types of signals. Next, the staff analyzes the fluctuation amplitude and duration of the three types of frequency domain characteristics, determines the frequency range of each signal containing effective information as the target frequency band parameter, and then determines the corresponding target noise suppression parameter in combination with the filtering method adapted for each signal. Finally, the staff collects the target frequency band parameters and target noise suppression parameters of all signals, integrates them according to the rule of "modal signal type-target frequency band parameter-target noise suppression parameter", and forms an adaptive modulation parameter group suitable for signal processing of the composite interlayer of the brand of body armor, preparing for the subsequent signal processing link.

[0058] By performing steps 121-124, the embodiments of the application accurately distinguish different types of signals through modal recognition, and match each signal with a dedicated processing channel, avoiding interference problems caused by mixed processing of different signals, reducing manual operations, and improving the automation level of signal processing. Through multi-layer convolution processing of the corresponding signals in each dedicated channel, frequency domain characteristics related to defect detection can be accurately extracted, irrelevant interference can be gradually removed, and the effectiveness and relevance of the characteristics can be ensured. The target frequency band parameters and target noise suppression parameters determined based on the frequency domain characteristics can accurately adapt to the characteristics of each signal, avoiding the loss of effective information or the residual interference caused by fixed parameters. Finally, the integrated adaptive modulation parameter group structure is clear and convenient for subsequent calling, laying a reliable foundation for subsequent unified and accurate frequency domain filtering and noise suppression processing of the original multi-modal response signal set, and ensuring the accuracy and orderly progress of the entire defect positioning process.

[0059] In one possible embodiment, step 123, based on the ultrasonic frequency domain characteristics, eddy current frequency domain characteristics and thermal wave frequency domain characteristics, determines the target frequency band parameters and target noise suppression parameters of each modal response signal, comprising:

[0060] a1, based on the fluctuation amplitude and duration of each characteristic in the corresponding frequency band, calculating the response intensity value corresponding to each characteristic.

[0061] Wherein, each feature refers to ultrasonic frequency domain feature, eddy current frequency domain feature and thermal wave frequency domain feature. These features are key information related to frequency extracted from the corresponding modal signal, which can reflect the characteristics of the signal at different frequencies; the fluctuation amplitude is the size of the signal change in the corresponding frequency band of each feature, the more obvious the change, the greater the fluctuation amplitude; the duration is the time length of each feature maintaining this kind of fluctuation state in the corresponding frequency band, which can reflect the stability of the fluctuation; the response intensity value is a value calculated by combining the fluctuation amplitude and duration of each feature, which is used to measure the degree of containing effective information of the feature in the corresponding frequency band. The whole step will finally generate the response intensity value corresponding to each feature.

[0062] In the embodiments of the present application, the frequency bands corresponding to the ultrasonic frequency domain feature, the eddy current frequency domain feature and the thermal wave frequency domain feature are first determined to ensure that each feature corresponds to a frequency range that can reflect effective information. Then the fluctuation amplitude and duration of each feature in the corresponding frequency band are obtained. The signal analysis tool is used to read the size of the signal change and the duration of the change in the corresponding frequency band of each feature. Finally, the response intensity value of each feature is calculated by combining the values of the fluctuation amplitude and the duration using a preset formula. The two factors are given reasonable weights to balance the influence during calculation. For example, in the scene of factory quality inspection of a certain brand of bulletproof vest, the frequency bands corresponding to the three features are first determined, then the fluctuation amplitude (a value representing the size of the change) and the duration (the whole signal acquisition or a specific period) of each feature in the corresponding frequency band are read, and then the fluctuation amplitude and the duration are combined according to the weights using a preset formula to calculate the response intensity value of each feature. If the fluctuation amplitude of a certain feature is large and the duration is long, the calculated response intensity value will be higher.

[0063] a2, obtain the intensity level corresponding to the response intensity value from the preset mapping relationship between intensity levels and intensity ranges, and take the product of the reference frequency band parameter and the frequency band coefficient corresponding to the intensity level as the frequency band parameter of the corresponding modal response signal, each feature including: ultrasonic frequency domain feature, eddy current frequency domain feature and thermal wave frequency domain feature.

[0064] The mapping relationship between the preset intensity level and the intensity range is a rule set in advance, which corresponds different intervals of the response intensity value to different intensity levels, and is used for quickly matching the level to which the response intensity value belongs; the intensity level is a division standard for representing the high and low of the response intensity value, and different levels correspond to different parameter adjustment coefficients; the reference frequency band parameter is a basic frequency range set in advance for each modal response signal, and is determined based on the general characteristics of the modal signal; the frequency band coefficient corresponding to the intensity level is an adjustment value matched with each intensity level, and is used for adjusting the reference frequency band parameter according to the high and low of the response intensity value; the frequency band parameter of the corresponding modal response signal is a final frequency range obtained by multiplying the reference frequency band parameter and the frequency band coefficient, and is used for subsequent signal filtering; each feature refers to an ultrasonic frequency domain feature, an eddy current frequency domain feature and a thermal wave frequency domain feature, and the whole step finally generates the frequency band parameter of each modal response signal.

[0065] In the embodiment of the present application, the mapping relationship between the preset intensity level and the intensity range is first called, the intensity level corresponding to different intervals of the response intensity value is determined, then the feature response intensity values calculated in step a1 are compared with the intervals in the mapping relationship to determine the intensity level to which each response intensity value belongs, then the reference frequency band parameter corresponding to each modal response signal and the frequency band coefficient matched with each intensity level are obtained, and finally the reference frequency band parameter of each modal is multiplied by the frequency band coefficient corresponding to the intensity level to obtain the frequency band parameter of the modal response signal. For example, in the scene of factory quality inspection of a certain brand of bulletproof vest, the preset mapping relationship is first called to determine the intensity level and the matched frequency band coefficient corresponding to each interval, then the feature response intensity values obtained in a1 are compared with the intervals to determine the level, then the reference frequency band parameter of each modal is obtained, and finally the reference parameter is multiplied by the frequency band coefficient to obtain the frequency band parameter of each modal for subsequent filtering.

[0066] a3, determining the target noise suppression parameter of each modal response signal according to the frequency band parameter of the modal response signal and the filtering mode corresponding to the modal; different modes correspond to different filtering modes.

[0067] The frequency band parameter of the modal response signal is the frequency range used for filtering of each modal obtained in step a2, which determines the effective frequency part to be retained in each modal signal; the filtering mode corresponding to the modal is a special interference filtering method designed for the characteristics of each modal response signal, and different modes adopt different filtering modes due to different signal types, such as limiting signal amplitude to filter interference or limiting signal duration to filter interference; the target noise suppression parameter of each modal response signal is a specific standard determined in combination with the frequency band parameter and the filtering mode of the modal, which is used to determine the characteristics of the interference signal to be removed during filtering, and the whole step finally generates the target noise suppression parameter of each modal response signal.

[0068] In the embodiments of the present application, first, the filtering mode corresponding to each modal response signal is determined, which is pre-set based on the type of modal signal, then the frequency band parameters of each modal response signal obtained in step a2 are acquired, the effective frequency range that needs to be retained for each modal is determined, then the typical characteristics of the effective signal in the frequency band parameters of each modal are analyzed, the normal range of these characteristics is determined through historical detection data or signal characteristics, and finally, the target noise suppression parameter is determined according to the filtering mode corresponding to the modal and the typical characteristic range of the effective signal, that is, the characteristic standard of the interference signal that needs to be filtered out is determined, for example, in the scene of factory quality inspection of a certain brand of bulletproof vest, first, the filtering mode corresponding to each modal is determined, then the frequency band parameters of each modal obtained in a2 are acquired, then the typical characteristic range of the effective signal is analyzed, and finally, the filtering standard of the interference signal is determined in combination with the filtering mode to form the target noise suppression parameter.

[0069] The present application provides the following specific examples: in the factory quality inspection process of a certain brand of bulletproof vest, the staff first determines the frequency band corresponding to the ultrasonic frequency domain feature, the eddy current frequency domain feature and the thermal wave frequency domain feature respectively, reads the fluctuation amplitude and duration of each feature in the corresponding frequency band through a signal analysis tool, then calculates the response intensity value of each feature according to the preset formula with weights, and if the fluctuation amplitude of a certain feature is large and the duration is long, the response intensity value obtained is higher. Then the preset intensity level and intensity range mapping relationship is called, each feature response intensity value is compared with the interval in the mapping relationship, the intensity level to which each response intensity value belongs is determined, the reference frequency band parameter of each modal response signal and the frequency band coefficient matched with each intensity level are acquired, and the reference frequency band parameter is multiplied by the corresponding frequency band coefficient to obtain the frequency band parameter of each modal. Finally, the filtering mode corresponding to each modal is determined, the frequency band parameters of each modal obtained before are called, the typical characteristic range of the effective signal in each frequency band is analyzed, and then the characteristic standard of the interference signal that needs to be filtered out is determined according to the filtering mode to form the target noise suppression parameter of each modal, which prepares for subsequent signal processing.

[0070] By performing a1~a3, the embodiment of the present application calculates the response intensity value by combining the fluctuation amplitude and the duration of each feature through step a1, which can comprehensively reflect the degree of effective information contained in the feature, avoid the deviation caused by single factor judgment, and provide accurate quantitative reference for subsequent parameter setting; through step a2, the response intensity value is associated with the intensity level, and the reference frequency band parameter is adjusted by the frequency band coefficient, so that the modal frequency band parameter can be flexibly adapted to the effective information degree of the corresponding feature, avoiding the problem that the fixed parameter cannot fully retain the effective signal or contains too much interference; through step a3, the target noise suppression parameter is determined by combining the modal frequency band parameter and the exclusive filtering mode, so that the noise suppression parameter can accurately match the modal signal characteristics and the effective frequency range, specifically remove the interference and retain the effective signal, ensuring the rationality and accuracy of the subsequent signal filtering and noise suppression processing, laying a foundation for generating high-quality optimized signal set and improving the accuracy of subsequent defect positioning.

[0071] In a possible embodiment, S13, based on the adaptive modulation parameter set, performs frequency domain filtering and noise suppression processing on the original multi-modal response signal set to generate an optimized signal set, comprising:

[0072] Step 131, for each modal response signal in the original multi-modal response signal set, according to the corresponding target frequency band parameter, retaining the frequency band part of each modal response signal that is higher than or equal to the corresponding target frequency band parameter, and removing the frequency band part of each modal response signal that is lower than the corresponding target frequency band parameter, to complete the frequency band filtering, each modal response signal including: ultrasonic guided wave response signal, pulse eddy current response signal and infrared thermal wave excitation response signal.

[0073] Among them, the frequency band filtering is a processing method that retains effective signals and removes invalid signals according to frequency range. The whole step finally completes the frequency band filtering processing of each modal response signal.

[0074] In the embodiment of the present application, first, for the ultrasonic guided wave response signal, the pulse eddy current response signal and the infrared thermal wave excitation response signal in the original multi-modal response signal set, find the corresponding target frequency band parameter of each signal, ensure that each signal can match the adaptive frequency range lower limit standard, then perform frequency band screening on each modal response signal, retain the part of the signal whose frequency is higher than or equal to the corresponding target frequency band parameter, and remove the part whose frequency is lower than the parameter, to complete the frequency band filtering of each modal response signal. For example, in the A brand body armor factory quality inspection scene, the target frequency band parameter of the ultrasonic guided wave response signal is a certain frequency band lower limit, the target frequency band parameter of the pulse eddy current response signal is another frequency band lower limit, and the target frequency band parameter of the infrared thermal wave excitation response signal is another frequency band lower limit, then retain the part above the corresponding frequency band lower limit for each signal respectively, and remove the part below the lower limit.

[0075] Step 132, for the ultrasonic guided wave response signal, a wavelet threshold denoising algorithm is adopted, the threshold coefficient in the corresponding target noise suppression parameter is taken as the threshold reference of wavelet decomposition, and the high-frequency noise in the ultrasonic guided wave response signal is attenuated.

[0076] The ultrasonic guided wave response signal is the signal filtered by the frequency band in step 131, which may still contain high-frequency interference; the wavelet threshold denoising algorithm is a processing method for weakening high-frequency interference in the signal, which distinguishes and attenuates high-frequency noise by setting a judgment standard; the threshold coefficient in the target noise suppression parameter is a value for setting the judgment standard, which is determined in advance for the ultrasonic guided wave response signal; the high-frequency noise is the interference signal in the ultrasonic guided wave response signal with higher frequency but no effective information, and the whole step finally attenuates the high-frequency noise in the ultrasonic guided wave response signal.

[0077] In the embodiments of the present application, the target noise suppression parameter for the ultrasonic guided wave response signal is first called, the threshold coefficient for setting the judgment standard is obtained therefrom, and the coefficient is ensured to adapt to the noise condition of the current signal. Then, the wavelet threshold denoising algorithm is started, the obtained threshold coefficient is taken as the judgment standard in the algorithm, so that the algorithm can distinguish the high-frequency noise and the effective high-frequency signal according to the standard. Finally, the ultrasonic guided wave response signal is processed by the algorithm, the high-frequency part determined as noise is attenuated, and the high-frequency signal containing effective information is retained. For example, in the A brand body armor factory inspection scene, after the adaptive threshold coefficient is called, it is set as the judgment standard, the algorithm will identify the high-frequency interference without effective structural information and weaken it, while retaining the high-frequency effective signal reflecting the interlayer structure.

[0078] Step 133, for the pulse eddy current response signal, an adaptive Kalman filter algorithm is adopted, and the covariance matrix in the corresponding target noise suppression parameter is taken as the filter iteration initial value to dynamically correct the eddy current interference noise in the pulse eddy current response signal.

[0079] The pulse eddy current response signal is the signal filtered by the frequency band in step 131, which may contain eddy current interference noise affecting the accuracy; the adaptive Kalman filter algorithm is a filter method that can dynamically adjust the processing mode to adapt to the signal changes, and can correct the interference in the signal in real time; the covariance matrix in the target noise suppression parameter is a data set describing the noise distribution characteristics of the pulse eddy current response signal, which is used to provide an initial reference for the filter algorithm; the filter iteration initial value is the initial data when the filter algorithm starts to run, and the covariance matrix will be used as the initial value; the eddy current interference noise is the noise in the pulse eddy current response signal generated by electromagnetic interference, which affects the accuracy of the signal, and the whole step finally dynamically corrects this kind of interference noise in the pulse eddy current response signal.

[0080] In the embodiments of the present application, first, the target noise suppression parameter for the pulse eddy current response signal is called, from which the covariance matrix containing the distribution information of the common noise of this type of signal is extracted, then the adaptive Kalman filtering algorithm is started, and the extracted covariance matrix is taken as the initial value of the filtering iteration of the algorithm, thereby providing a reference basis for the initial operation of the algorithm, and finally the algorithm dynamically adjusts the filtering strategy according to the real-time changes of the signal, continuously corrects the eddy current interference noise in the signal, for example, in the A brand body armor factory quality inspection scene, the covariance matrix records the distribution rule of the noise of this type of signal, and the algorithm takes this as the starting point to analyze the current signal, and when new electromagnetic interference occurs, the algorithm will automatically calculate the correction amount to remove the interference and keep the signal stable.

[0081] In step 134, for the infrared thermal wave excitation response signal, a combination of Gaussian filtering and median filtering is used to smooth the thermal noise fluctuation in the infrared thermal wave excitation response signal through Gaussian filtering, and to remove the salt and pepper noise in the infrared thermal wave excitation response signal through median filtering.

[0082] In step 134, for the infrared thermal wave excitation response signal, a combination of Gaussian filtering and median filtering is used to smooth the thermal noise fluctuation in the infrared thermal wave excitation response signal through Gaussian filtering, and to remove the salt and pepper noise in the infrared thermal wave excitation response signal through median filtering.

[0083] In the embodiments of the present application, first, the infrared thermal wave excitation response signal is processed by Gaussian filtering, which makes the thermal noise fluctuation in the signal due to random changes in temperature become smooth, reducing the degree of fluctuation of the signal, and then the signal processed by Gaussian filtering is processed by median filtering, which finds and removes isolated points of mutation (i.e. salt and pepper noise) in the signal, and replaces these points with normal signal values around them. Through such combined processing, noise suppression of the infrared thermal wave excitation response signal is completed, for example, in the A brand body armor factory quality inspection scene, Gaussian filtering will smooth the small range of fluctuations in the signal caused by local temperature sudden rise and sudden drop, making the overall trend more stable, and then the median filtering will identify isolated points with amplitude far exceeding the surrounding points, replace them with the median value of the adjacent normal signal, and eliminate abnormality caused by device interference.

[0084] Step 135, signal intensity normalization processing is performed on each modal response signal after noise suppression processing, and the amplitude range of all modal response signals is uniformly mapped to a preset interval to eliminate the magnitude difference between different modal response signals.

[0085] Among them, each modal response signal after noise suppression processing is the ultrasonic guided wave response signal, the pulse eddy current response signal and the infrared thermal wave excitation response signal processed by steps 132 to 134, and the amplitude range of these signals may have a large difference; signal intensity normalization processing is a processing method for uniformly adjusting the amplitude range of different signals to the same interval; the amplitude range refers to the interval between the maximum amplitude and the minimum amplitude in the signal; the preset interval is a fixed interval for uniformly adjusting the amplitude of each signal, and is usually selected in a range convenient for subsequent processing; the magnitude difference refers to the size difference of different modal response signals due to the difference in amplitude range, and the whole step will ultimately eliminate this magnitude difference and unify the amplitude range of each signal.

[0086] In the embodiments of the present application, the preset interval for uniformly adjusting the amplitude of each modal response signal is first determined, which is selected in a range convenient for subsequent signal comparison and analysis, then the amplitude range of the ultrasonic guided wave response signal, the pulse eddy current response signal and the infrared thermal wave excitation response signal after noise suppression processing is obtained respectively, and finally the amplitude of each signal is mapped to the preset interval by calculation. The calculation method is to subtract the minimum amplitude of the signal from each amplitude in the signal, and then divide by the difference between the maximum amplitude and the minimum amplitude of the signal. For example, in the A brand body armor factory quality inspection scene, the preset interval is selected in a fixed range convenient for comparison, after obtaining the amplitude range of each signal, the amplitude of each signal is uniformly adjusted to the interval to avoid that the amplitude of a certain type of signal is too large to cover the effective information of other signals.

[0087] Step 136, the normalized ultrasonic guided wave response signal, the normalized pulse eddy current response signal and the normalized infrared thermal wave excitation response signal are aligned in time axis and space axis to ensure that the spatial coordinates and time stamps of different modal response signals at the same sampling point correspond one by one, and an optimized signal set is formed.

[0088] The normalized modal response signals are the ultrasonic guided wave response signal, the pulsed eddy current response signal and the infrared thermal wave excitation response signal processed by step 135, which can be different in time and space. The alignment of the time axis and the space axis is a processing method for adjusting signals to keep different signals consistent in time and space. The same sampling point refers to the signal data point of the same time and space position in each modal response signal. The space coordinates and the time stamp are information describing the position and collection time of the sampling point. The space coordinates reflect the position of the sampling point on the composite sandwich of the body armor, and the time stamp reflects the time of collecting the data. One-to-one correspondence means that the space coordinates and the time stamp of each sampling point are completely matched in different modal signals. The optimized signal set is a high-quality and synchronized signal set formed by integrating the aligned modal response signals. The whole step finally forms the optimized signal set.

[0089] In the embodiments of the present application, the normalized ultrasonic guided wave response signal, the pulsed eddy current response signal and the infrared thermal wave excitation response signal are first obtained, as well as the space coordinates and the time stamp of each sampling point of each signal. Then, the time axis alignment processing is performed to adjust the time stamp of each modal signal sampling point, so that the collection time of the same space position sampling point in the three signals is consistent. Then, the space axis alignment processing is performed to ensure that the space coordinates of the same time stamp sampling point in the three signals are completely matched. Finally, the three modal response signals after the alignment of the time axis and the space axis are integrated together to form an optimized signal set containing synchronized signal data. For example, in the A brand body armor factory quality inspection scene, after obtaining the sampling points of each signal and the corresponding position and time information, the time stamp is adjusted to make the sampling time of the same position consistent, and then the space coordinates are adjusted to make the sampling position of the same time match. Finally, the synchronized sampling point data is associated and integrated to form an optimized signal set.

[0090] The application provides the following specific examples: in the A brand body armor factory quality inspection process, the staff first finds the corresponding target frequency band parameters for three kinds of modal response signals in the original multi-modal response signal set, retains the part above the corresponding frequency band lower limit for each signal, removes the part below the lower limit, and completes the frequency band filtering. Then, for the ultrasonic guided wave response signal filtered by the frequency band, the corresponding threshold coefficient is called, the wavelet threshold denoising algorithm is started, the coefficient is taken as the judgment standard, and the high frequency noise without effective information is attenuated; for the pulse eddy current response signal, the covariance matrix is taken as the initial value of the adaptive Kalman filtering algorithm, and the eddy current noise caused by electromagnetic interference is dynamically corrected; for the infrared thermal wave excitation response signal, the gentle interference caused by random temperature fluctuations is first smoothed by Gaussian filtering, and then the isolated mutation points caused by device interference are replaced by median filtering, and the two kinds of noise are removed. Then, a preset interval for comparison is determined, the amplitude range of each of the three signals is obtained, the amplitudes of each signal are uniformly mapped to the interval by calculation, and the size difference of different signals is eliminated. Finally, the normalized signals and the position and time information of each sampling point are obtained, the time stamp is adjusted to make the sampling time at the same position consistent, the spatial coordinates are adjusted to make the sampling position at the same time match, the synchronized sampling point data are associated and integrated, and an optimized signal set is formed.

[0091] By performing steps 131-136, the embodiments of the application can effectively retain the parts containing effective structure information in each modal signal and remove the low frequency interference without effective information, providing a pure basis for subsequent noise processing; by steps 132 to 134, the denoising method suitable for the characteristics of different modal signals is selected, which can accurately attenuate or remove the exclusive interference in each signal (such as high frequency noise of ultrasonic guided wave, electromagnetic interference of pulse eddy current, and two kinds of noise of infrared thermal wave), and retain the effective information to the greatest extent; by the normalization processing of step 135, the weight imbalance problem caused by the amplitude difference of different modal signals is eliminated, and it is ensured that each signal has equal reference value in subsequent processing; by the time and space alignment and integration of step 136, the sampling points of different modal signals are completely synchronized in time and space, forming a high-quality optimized signal set, continuously improving the signal quality and synchronization, and laying a reliable data foundation for subsequent extraction of spatial frequency characteristics and cross-modal feature fusion, and realization of accurate defect positioning.

[0092] In one possible embodiment, step 136, the normalized ultrasonic guided wave response signal, the normalized pulse eddy current response signal and the normalized infrared thermal wave excitation response signal are subjected to time axis and space axis alignment processing to ensure that different modal response signals correspond one by one in space coordinates and time stamps at the same sampling point, and are integrated to form an optimized signal set, comprising:

[0093] b1. Taking the geometric center of the ballistic vest composite interlayer as the coordinate origin, a three-dimensional coordinate system is constructed, and the coordinates of the sampling points of each modal response signal are converted to the three-dimensional coordinate system to obtain three-dimensional coordinate values, wherein the x-axis and the y-axis correspond to the interlayer plane direction, and the z-axis corresponds to the interlayer thickness direction.

[0094] Wherein, the geometric center of the ballistic vest composite interlayer is the combination point of the midpoint of the interlayer plane direction and the middle position of the thickness direction, which is used as the coordinate origin of the three-dimensional coordinate system; the three-dimensional coordinate system is a three-dimensional reference system for describing the spatial position of each point on the interlayer, wherein the x-axis and the y-axis correspond to the plane direction of the interlayer (such as the x-axis along the length direction of the interlayer and the y-axis along the width direction of the interlayer), and the z-axis corresponds to the thickness direction of the interlayer (from the surface to the inside of the interlayer); the sampling point coordinates of each modal response signal are the position data of the sensors on the interlayer when collecting each signal; the three-dimensional coordinate values are the values obtained after converting the sampling point coordinates to the three-dimensional coordinate system, which can accurately represent the spatial position of the sampling points. The whole step finally obtains the three-dimensional coordinate values of all sampling points.

[0095] In the embodiments of the present application, the geometric center of the ballistic vest composite interlayer is first determined, the midpoint of the plane and the midpoint of the thickness (half of the thickness) are calculated by measuring the length, width and thickness of the interlayer, and this point is taken as the coordinate origin of the three-dimensional coordinate system; then the coordinate axes of the three-dimensional coordinate system are defined, the x-axis is set along the length direction of the interlayer, the y-axis is set along the width direction of the interlayer, and the z-axis is set along the thickness direction of the interlayer; finally, the sampling point coordinates of each modal response signal are converted to the coordinate system to calculate the three-dimensional coordinate values. For example, in the A brand ballistic vest factory inspection scene, the midpoints of the interlayer plane and thickness are found as the origin, the directions of the three coordinate axes are defined, and then the position data of each sampling point on the interlayer is converted to three-dimensional values with the origin as the reference, ensuring that the positions of all sampling points can be represented in the same coordinate system.

[0096] b2. Extracting the timestamp information of each modal response signal, calibrating the timestamps of different modal signals according to a preset precision clock synchronization protocol, so that the time reference error of all modal response signals is controlled within a preset error range.

[0097] Wherein, the timestamp information of each modal response signal is the collection time of the corresponding sampling point data recorded when collecting each signal; the preset precision clock synchronization protocol is a synchronization rule set in advance to ensure the accuracy of time recording of different signals, which is used to unify the time reference of each modal signal; the time reference error is the deviation between the timestamps of different modal signals and the unified time reference; the preset error range is the maximum value of the allowed time reference error set in advance, and the whole step finally controls the time reference error of all modal response signals within this range.

[0098] In the embodiments of the present application, first, the timestamp information corresponding to each sampling point is extracted from the collected data of each modal response signal, ensuring that the collection time of each sampling point is accurately extracted; then a preset precision clock synchronization protocol is started, and the timestamp of one of the signals is taken as a standard to adjust the timestamps of the other modal signals, so that the timestamps of different modal signals corresponding to the same sampling point remain consistent; finally, the error of the adjusted time reference is checked to ensure that the deviation of the timestamps of all modal signals from the reference does not exceed the preset error range, and if it does, it is adjusted again until it meets the requirements. For example, in the A brand body armor factory inspection scene, the collection time of each sampling point of the three signals is first extracted, then the time of one of the signals is taken as a standard to adjust the times of the other two signals to be consistent, and finally the adjusted time deviation is confirmed to meet the requirements.

[0099] b3, grid processing is performed on the three-dimensional coordinate values to divide the body armor composite interlayer into a plurality of three-dimensional grid units of a preset size, each three-dimensional grid unit corresponding to a unique spatial index.

[0100] The grid processing performed on the three-dimensional coordinate values is a processing method of dividing the three-dimensional space of the body armor composite interlayer into a plurality of small space units according to a fixed size; the preset size is the length, width and height of each small space unit set in advance (such as a fixed length, width and thickness suitable for detection accuracy); the three-dimensional grid unit is each small space unit formed after grid processing, each unit corresponding to a small area of the interlayer; the spatial index is a unique identifier assigned to each three-dimensional grid unit, used for quickly locating and distinguishing different grid units, and the entire step will eventually divide the interlayer into a plurality of three-dimensional grid units with unique spatial indexes.

[0101] In the embodiments of the present application, first, the preset size of the three-dimensional grid unit is determined, and the length, width and height of each grid unit are set according to the size of the body armor composite interlayer and the detection accuracy requirement, to ensure that the unit size can cover all areas of the interlayer and meet the detection accuracy; then, based on the three-dimensional coordinate system constructed in step b1, the three-dimensional space of the interlayer is divided into a plurality of uniform small units by dividing the three-dimensional space of the interlayer according to the preset size along the x-axis, y-axis and z-axis directions from the coordinate origin; finally, a unique spatial index is assigned to each three-dimensional grid unit, and each unit is labeled according to a unified serial number rule to ensure that each grid unit can be uniquely identified by the index. For example, in the A brand body armor factory inspection scene, the size of the grid unit suitable for the interlayer of this brand is first set, then the space is divided according to the coordinate system direction, and finally a unique label is assigned to each small unit to facilitate subsequent searching.

[0102] b4, mapping the sampling points of each modal response signal to the corresponding three-dimensional grid cells by using the nearest neighbor interpolation method, if there is a three-dimensional grid cell lacking a sampling point, then taking the three-dimensional grid cell lacking the sampling point as the center, and performing weighted calculation according to the signal values of the adjacent 8 grid cells.

[0103] wherein the nearest neighbor interpolation method is a method of distributing sampling point data to grid cells, that is, finding the three-dimensional grid cell closest to the sampling point, and distributing the signal value of the sampling point to this cell; mapping is the process of associating the signal value of the sampling point to the corresponding three-dimensional grid cell; the three-dimensional grid cell lacking a sampling point is a grid cell that does not have any sampling point directly corresponding to it and has not been assigned a signal value; the adjacent 8 grid cells are 8 grid cells adjacent to the grid cell lacking a sampling point in the x, y, z three-axis directions; and the weighted calculation is to assign different weights to the signal values of each adjacent cell according to the distance of the adjacent 8 cells from the center cell, and then calculate the weighted average value as the signal value of the center cell. The entire step finally assigns signal values to all grid cells.

[0104] In the embodiments of the present application, first, the nearest neighbor interpolation method is used to process the sampling points of each modal response signal. For each sampling point, the distance between its three-dimensional coordinate value and the center coordinate of all three-dimensional grid cells is calculated, the nearest grid cell is found, and the signal value of the sampling point is assigned to this cell to complete the signal mapping of the existing sampling points. Then, all three-dimensional grid cells are checked to find the cells lacking sampling points that have not been assigned signal values. Finally, for each cell lacking a sampling point, the 8 adjacent grid cells are determined, the signal values of the 8 cells are obtained, the weighted average value is calculated as the signal value of the cell lacking a sampling point according to the distance, and the signal values of all grid cells are ensured to have corresponding signal values. For example, in the A brand body armor factory inspection scene, first, the signal of each sampling point is assigned to the nearest grid cell, and then the cells without signals are found, and the signals of the surrounding 8 cells are weighted to calculate and complete, so that all cells have signals.

[0105] b5, generating a space-time alignment verification matrix, the elements in the space-time alignment verification matrix are signal existence identifiers of each modal response signal at the same three-dimensional grid cell and time stamp, when the missing rate of the signal existence identifier exceeds a preset missing rate threshold, re-executing the above alignment processing procedures b1~b5 until the missing rate of the signal existence identifier is less than or equal to the preset missing rate threshold, integrating in the order of three-dimensional grid cell position, and forming an optimized signal set.

[0106] The space-time alignment check matrix is a table for checking whether there is a signal value of each modality signal at the same grid cell and time stamp, the rows of the matrix correspond to the spatial indexes of the three-dimensional grid cells, the columns correspond to the time stamps, and the elements are signal existence identifiers; the signal existence identifier is a symbol for marking whether there is a signal value of the grid cell at a specific time stamp; the missing rate is the proportion of the number of "0" in the matrix to the total number of elements; the preset missing rate threshold is the maximum missing rate allowed in advance; the re-execution of the above alignment processing procedure is to execute steps b1 to b4 again when the missing rate exceeds the threshold; the integration is to associate the grid cell signal values and time stamps of each modality signal according to the position order of the three-dimensional grid cells; the optimized signal set is a signal set formed after integration, which is space-time aligned and data complete, and the whole step will finally form the optimized signal set.

[0107] In the embodiments of the present application, first, a space-time alignment check matrix is constructed, the rows and columns of the matrix are determined, for each cell, according to whether there is a signal value of each modality signal at the grid cell and time stamp, "1" or "0" is filled in as a signal existence identifier; then the missing rate is calculated, the number of "0" in the matrix is counted, and the missing rate is obtained by dividing the total number of elements of the matrix (the number of rows x the number of columns); then the missing rate is compared with the preset missing rate threshold, if the missing rate exceeds the threshold, steps b1 to b4 are re-executed to optimize the data integrity; if the missing rate is less than or equal to the threshold, the grid cell signal values and time stamps of each modality signal are integrated together according to the spatial index order of the three-dimensional grid cells, to form an optimized signal set, for example, in the A brand body armor factory inspection scene, first, the matrix is built to check which grid cells and time stamps have no signal, the missing rate is calculated, if there are too many missing, the previous steps are reprocessed, if it meets the requirements, all signals are integrated according to the cell order to form a complete optimized signal set.

[0108] The application provides the following specific examples: in the A brand body armor factory quality inspection process, the staff first measures the planar size and thickness of the A brand body armor composite sandwich, finds the combination point of the planar midpoint and the thickness midpoint as the origin of the three-dimensional coordinate system, defines the directions of the three coordinate axes, and then converts the positions of the sampling points of the three modal signals on the sandwich to three-dimensional coordinate values in the coordinate system. Then the collection time of each sampling point of the three modal signals is extracted, and the time of one of the signals is taken as the unified reference to adjust the time stamps of the other two signals to ensure that the time records of the same sampling point are consistent and the time deviation meets the preset requirements. Then, according to the sandwich size and detection accuracy, a suitable three-dimensional grid unit size is set, and the sandwich space is divided into multiple small units according to the coordinate system direction, and each unit is assigned a unique space index. Then the signal value of each sampling point is assigned to the nearest grid unit by the nearest neighbor interpolation method, and the signal value of the unit without signal is calculated by distance weighting through the signals of the eight adjacent units around it. Finally, a space-time alignment verification matrix is constructed to check whether the signals under each grid unit and time stamp exist, calculate the signal missing ratio, confirm that the missing rate meets the requirements, and then integrate the signal values of the three modal signals with the time stamps according to the position order of the grid units to form a complete optimized signal set.

[0109] By performing b1-b5, the embodiments of the application construct a unified three-dimensional coordinate system through the b1 step, so that the positions of all sampling points can be described under the same reference, avoiding spatial positioning confusion; the time stamps are calibrated through the b2 step to ensure that the time records of different modal signals are consistent, providing protection for the alignment of the time dimension; through the grid processing of the b3 step, the dispersed sampling points are corresponded to fixed space units, facilitating subsequent data management and completion; through the b4 step, the signal values of the missing sampling points are completed, avoiding the detection blind area caused by data vacancy; through the verification and integration of the b5 step, the integrity of the signal in space and time is ensured, forming a high-quality optimized signal set, realizing the double alignment of signal space and time and data completion, and providing a clear structure and complete data reliable data source for subsequent cross-modal feature fusion and accurate identification of body armor composite sandwich defects.

[0110] In a possible embodiment, S14, through the attention mechanism in the cross-modal feature fusion model, adaptively associates the defect sensitive features of different modalities, generates a three-dimensional defect distribution map of the composite sandwich, and based on the three-dimensional defect distribution map, performs millimeter-level precision defect positioning, including:

[0111] Step 141, the spatial frequency characteristics of the modal response signal are frequency domain segmented to obtain spatial frequency characteristics of multiple frequency bands, and a cosine similarity algorithm is used to calculate the matching degree of the spatial frequency characteristics of each frequency band and a preset defect characteristic template, the defect characteristic template being constructed based on spatial frequency characteristics of a preset number of known defects and containing characteristic parameters of different types and sizes of defects in each frequency band, the types including delamination, cavity and crack, and the characteristic parameters including frequency distribution and amplitude variation rate.

[0112] The spatial frequency characteristics of the modal response signal are key information extracted from the optimized signal set and capable of reflecting the change of signal frequency with the spatial position of the composite sandwich of the body armor; the frequency domain segmentation is a processing mode of splitting the spatial frequency characteristics into multiple small frequency bands according to the frequency range, each frequency band corresponding to a specific frequency interval; the cosine similarity algorithm is a method of calculating the similarity degree of two characteristics, the higher the similarity degree, the closer the matching degree value calculated to 1; the matching degree is a value obtained by the algorithm and used to represent the similarity degree of the spatial frequency characteristics of a frequency band and the defect characteristic template; the defect characteristic template is a reference template constructed based on the spatial frequency characteristics of a large number of known defects and containing characteristic parameters of several common defects such as delamination, cavity and crack in different frequency bands, the parameters including frequency distribution (distribution of signal frequency in the frequency band) and amplitude variation rate; the entire step finally obtains the matching degree of the spatial frequency characteristics of each frequency band and the defect characteristic template.

[0113] In the embodiments of the present application, the spatial frequency characteristics of the modal response signal are first frequency domain segmented, and according to the overall frequency range of the signal and the detection requirements, it is divided into multiple continuous frequency bands, ensuring that each frequency band can independently reflect the characteristics of a certain frequency range, for example, in the A brand body armor factory quality inspection scene, the spatial frequency characteristics are divided into multiple frequency bands according to several continuous frequency ranges; secondly, the defect characteristic template prepared in advance is called, which records the frequency distribution and amplitude variation rate parameters of delamination, cavity and crack in these frequency bands; finally, for the spatial frequency characteristics of each frequency band, the cosine similarity algorithm and the parameters of the corresponding frequency band in the defect characteristic template are used to calculate the matching degree, and the similarity between the characteristics is quantified by a specific formula during calculation, for example, the matching degree of a certain frequency band characteristic and the corresponding frequency band of the template is calculated, and the matching degrees of other frequency bands are calculated in the same way.

[0114] Step 142, the spatial frequency characteristics with a matching degree greater than or equal to a corresponding threshold value in each frequency band are marked as candidate characteristics.

[0115] Wherein, the corresponding threshold is the minimum matching degree standard for judging whether the frequency band feature is likely to belong to the defect, which is set in advance for each frequency band. Different frequency bands may have different thresholds because the defect features are different. The candidate feature refers to the frequency band spatial frequency feature with a matching degree greater than or equal to the corresponding threshold. These features have the potential to indicate the existence of defects. The entire step finally marks all features that meet the conditions as candidate features.

[0116] In the embodiments of the present application, firstly, for the matching degree of each frequency band spatial frequency feature obtained in step 141, the corresponding preset threshold of the frequency band is retrieved, to ensure that each frequency band has its own appropriate judgment standard. For example, in the A brand body armor factory inspection scene, corresponding thresholds are set for different frequency bands. Secondly, the matching degree of each frequency band is compared with the corresponding threshold. If the matching degree of a certain frequency band is greater than or equal to the threshold, the spatial frequency feature of this frequency band is marked as a candidate feature. If the matching degree is less than the threshold, the feature of this frequency band is excluded. For example, in a certain modality, part of the frequency bands meet the standard and are marked as candidate features, and the other frequency bands that do not meet the standard are excluded.

[0117] Step 143, calculate the time domain difference degree of the candidate feature and the spatial frequency feature of the normal region in the same time interval. If the time domain difference degree is greater than the preset difference threshold, and the spatial coordinate deviation of the candidate feature corresponding to each modality is less than the preset deviation threshold, the candidate feature is determined as a defect sensitive feature.

[0118] Wherein, the time domain difference degree is a value used to measure the difference in changes of the candidate feature and the spatial frequency feature of the normal region in the same time interval. The greater the difference, the higher the time domain difference degree. The spatial frequency feature of the normal region is the spatial frequency feature extracted from the area of the body armor composite interlayer known to have no defects, which is used as a reference for comparison. The time interval is the same signal acquisition time period selected for calculating the time domain difference degree. The preset difference threshold is a minimum threshold set in advance to judge whether the time domain difference is large enough. The spatial coordinate deviation refers to the deviation between the spatial position coordinates corresponding to the candidate features of different modalities, which can reflect whether the features of each modality point to the same spatial region. The preset deviation threshold is the maximum allowed spatial coordinate deviation set in advance. The defect sensitive feature refers to the candidate feature that meets the conditions of the time domain difference degree being greater than the preset difference threshold and the spatial coordinate deviation corresponding to the candidate feature of each modality being less than the preset deviation threshold. Such features can clearly indicate the existence of defects. The entire step finally determines the defect sensitive features.

[0119] Step 144, calculate the mutual correlation value between the defect sensitive features of different modalities through the attention mechanism in the cross-modality feature fusion model, and determine the weight coefficient of the defect sensitive features of each modality according to the mutual correlation value.

[0120] The cross-modal feature fusion model is a model for integrating defect-sensitive features of different modalities, which can fully utilize the complementary information of each modality. The attention mechanism is the core functional module in the model, which can automatically focus on the modality features more important for judging defects. It measures the importance of modalities by calculating the correlation values between different modalities. The correlation value is a numerical value representing the closeness of the correlation between different modalities. The closer the correlation, the higher the correlation value. The weight coefficient is a coefficient determined according to the correlation value, which represents the importance of each modality defect-sensitive feature in the fusion process. The higher the correlation value, the greater the weight coefficient. The weight coefficient of each modality defect-sensitive feature is obtained through the above steps.

[0121] Step 145, according to the weight coefficient, the defect-sensitive features of different modalities are weighted and fused to form a fusion feature.

[0122] The weighted fusion is a processing method that weights and calculates the defect-sensitive features of different modalities according to the weight coefficient and then integrates them. Specifically, the feature data of each modality is multiplied by its corresponding weight coefficient, and then all the weighted feature data are integrated according to a predetermined rule. The fusion feature is a comprehensive feature obtained by weighted fusion, which contains the key information of all modality defect-sensitive features. This feature can more comprehensively and accurately reflect the defect situation. The fusion feature is finally obtained through the above steps.

[0123] Step 146, the fusion feature is layered according to the thickness direction of the body armor composite interlayer to obtain the feature distribution of different depth layers. The feature distribution of each depth layer is used to represent the defect information of the depth layer.

[0124] The thickness direction refers to the vertical direction of the body armor composite interlayer from the surface to the interior, which can reflect the different depths of the interlayer. The layered processing refers to the processing method of splitting the fusion feature into features corresponding to different depth layers according to the interval of the thickness direction. The feature distribution of the depth layer refers to the distribution of the fusion feature on each depth layer after splitting. This distribution can represent the defect information of the depth layer. The feature distribution of different depth layers is finally obtained through the above steps.

[0125] Step 147, the feature distribution of each depth layer is stacked in turn to construct a three-dimensional structure containing three-dimensional space information. The area corresponding to the defect feature is marked in the three-dimensional structure to obtain a three-dimensional defect distribution map.

[0126] The superimposition refers to a processing manner that the feature distribution of each layer is sequentially superimposed together according to the order of the depth layer (from the surface layer to the inner layer); the three-dimensional structure is a structure formed after superimposition, which can reflect the three-dimensional spatial defect information of the interlayer; the three-dimensional defect distribution map is a graph obtained after marking the area corresponding to the defect feature in the three-dimensional structure, and the graph can directly show the position and range of the defect in the three directions of the interlayer transverse direction, longitudinal direction and depth direction; the whole step finally obtains the three-dimensional defect distribution map.

[0127] Step 148, determining the position of the defect in the composite interlayer based on the three-dimensional coordinate information of the marked area in the three-dimensional defect distribution map.

[0128] The three-dimensional coordinate information of the marked area refers to the transverse (x-axis), longitudinal (y-axis) and depth (z-axis) coordinate range corresponding to the defect area marked in the three-dimensional defect distribution map; the defect position refers to the specific position of the defect in the composite interlayer determined by analyzing the three-dimensional coordinate information; the whole step finally determines the position of the defect in the composite interlayer.

[0129] The following specific examples are provided: in the A brand body armor factory quality inspection process, the staff first extracts the spatial frequency features of three modalities from the optimized signal set, divides them into multiple frequency bands according to the frequency range, then retrieves the preset defect feature template, and calculates the matching degree of each frequency band feature and the template by a specific algorithm; then retrieve the threshold value corresponding to each frequency band, mark the frequency band features that meet the matching degree as candidate features; then select a fixed time interval, extract the time domain data of the candidate features and the normal region features, calculate the time domain difference degree of the two, and then calculate the spatial coordinate deviation of the different modal candidate features after meeting the two conditions, and determine the features that meet both conditions as defect sensitive features; then input these sensitive features into the cross-modal fusion model, calculate the mutual correlation value between modalities through the attention mechanism of the model, and determine the weight coefficient of each modality according to the correlation value; then use the weight coefficient to weight the sensitive features of each modality, add the same type of parameters to obtain the fusion features; then set the layering interval according to the total thickness of the interlayer and the detection requirement, split the fusion features into feature distributions of different depth layers; superimpose these layer distributions to form a three-dimensional structure in order from the surface layer to the inner layer, mark the corresponding area by comparing the defect template, and form a three-dimensional defect distribution map; finally, extract the coordinate range of the defect area, calculate the center coordinates, determine the specific position and coverage range of the defect, and complete the whole defect positioning process.

[0130] By performing steps 141-148, the embodiment of the application narrows down the range of potential defect features through frequency domain segmentation and matching degree calculation in step 141, providing a basis for subsequent screening; step 142 quickly screens out candidate features through threshold comparison, reducing invalid data; step 143 accurately determines defect sensitive features by combining temporal differences and spatial coordinate deviations, reducing the risk of misjudgment; step 144 objectively allocates modal weights using the attention mechanism to avoid subjective bias; step 145 integrates multi-modal key information through weighted fusion to form more comprehensive fusion features; step 146 decomposes the fusion features to different depth levels through hierarchical processing, laying the foundation for three-dimensional positioning; step 147 forms an intuitive three-dimensional defect distribution map by superimposing and marking, facilitating the observation of defect spatial distribution; step 148 accurately calculates the defect coordinates and range to determine the specific location, ensuring the accuracy of defect recognition and achieving three-dimensional accurate positioning of defects, providing reliable support for judging the impact of defects on the protective performance of body armor and developing repair schemes.

[0131] In one possible embodiment, step 147, the feature distributions of each depth level are sequentially superimposed to construct a three-dimensional structure containing three-dimensional spatial information, and the areas corresponding to the defect features are marked in the three-dimensional structure to obtain a three-dimensional defect distribution map, including:

[0132] c1, the feature distributions of each depth level are sorted from shallow to deep according to depth, and each feature distribution corresponds to a different actual depth position.

[0133] Wherein, the feature distributions of each depth level are defect feature distribution conditions corresponding to different depths of the composite interlayer of the body armor, which are previously split from the fusion features, and these distributions contain feature parameters that can reflect whether there is a defect at the corresponding depth; the sorting from shallow to deep according to depth is a processing method that arranges the interlayer depths in order from the shallow layer close to the surface of the interlayer to the deep layer close to the interior of the interlayer; the actual depth position refers to the real position range corresponding to each feature distribution in the thickness direction of the interlayer; the entire step will finally obtain feature distributions sorted from shallow to deep according to depth, and each feature distribution is explicitly associated with its corresponding actual depth position.

[0134] c2, sequentially superimpose each feature distribution sorted according to depth along the depth direction, and in the superimposition process, the horizontal and vertical positions in the spatial coordinates of different depth levels correspond to each other to form a three-dimensional structure containing multiple depth levels.

[0135] The depth direction stacking refers to a processing manner of stacking the sorted feature distribution along the interlayer thickness direction (from shallow to deep) in sequence; the transverse and longitudinal positions correspond to each other, which means that the feature distribution of different depth levels is ensured to be consistent in the transverse and longitudinal coordinates reflecting the interlayer plane position, that is, the same transverse and longitudinal coordinates correspond to the same plane point of the interlayer; the three-dimensional structure is a three-dimensional structure formed after stacking, which can reflect the feature distribution of the interlayer in the transverse, longitudinal and depth directions at the same time; the whole step finally constructs a three-dimensional structure containing all depth levels and corresponding coordinates.

[0136] c3、In the three-dimensional structure, check whether the defect feature of each actual depth position falls into the preset feature reference range one by one, and classify the actual depth positions whose defect features fall into the preset feature reference range into the consistent area. The preset feature reference range is determined based on the feature parameters of known defects and covers the feature threshold values of different defect types.

[0137] Wherein, each actual depth position is the actual depth range of each depth level in the three-dimensional structure; the defect feature refers to the parameter in the feature distribution of each actual depth position that can indicate the defect, such as the frequency distribution of the signal, the amplitude change rate, etc.; the preset feature reference range is a range determined in advance based on the feature parameters of a large number of known defects (such as delamination, cavity, crack), which covers the feature parameter threshold values corresponding to different defect types; the consistent area refers to the three-dimensional structure area corresponding to the actual depth position whose defect feature falls into the preset feature reference range; the whole step finally classifies the actual depth positions meeting the conditions in the three-dimensional structure into the consistent area.

[0138] c4, marking all consistent areas in the three-dimensional structure as a whole to form a three-dimensional diagram containing the marked areas, and generating a three-dimensional defect distribution map based on the three-dimensional diagram.

[0139] Wherein, the three-dimensional diagram refers to a graph that visually displays the marked consistent areas on the three-dimensional structure, which can clearly show the position of the consistent area; the three-dimensional defect distribution map is generated based on the three-dimensional diagram, which can show the position and range of the defect in the interlayer in the transverse, longitudinal and depth directions in detail, and also supplements information such as coordinate scale and defect type; the whole step finally generates a three-dimensional defect distribution map that can intuitively reflect the three-dimensional distribution of the defect.

[0140] The application provides the following specific examples: in the A brand body armor factory quality inspection process, the staff first targets the previously split depth layer feature distribution, combined with the total thickness of the A brand interlayer and the previous layer spacing, determines the actual depth range corresponding to each feature distribution, and then sorts these feature distributions in order from shallow to deep, ensuring that each feature distribution is associated with a clear actual depth position. Then determine the uniform rule that the horizontal coordinate corresponds to the length direction of the interlayer and the vertical coordinate corresponds to the width direction, stack the sorted feature distributions in the depth direction in turn, align the horizontal and vertical coordinates of each feature distribution during stacking, and construct a three-dimensional structure containing all depth layers. Then call the preset feature reference range, which covers the feature parameter threshold of defects such as delamination, voids and cracks, check the feature parameters of each actual depth position in the three-dimensional structure one by one, and classify the areas corresponding to the depth positions falling within the reference range as conforming areas. Finally, mark all conforming areas using a specific color filling method, generate a three-dimensional illustration based on the marked three-dimensional structure, and supplement coordinate scales and defect type annotations on the illustration to generate a three-dimensional defect distribution map that can visually display the three-dimensional position and range of defects.

[0141] By performing c1-c4, the embodiments of the application sort the feature distribution by actual depth and associate the depth position through the c1 step, ensuring that the subsequent processing follows the real depth order of the interlayer, avoiding structural misplacement caused by chaotic order; through the c2 step, the feature distribution is stacked along the depth direction and the horizontal and vertical coordinates are kept corresponding, a three-dimensional structure consistent with the actual space of the interlayer is constructed, providing a three-dimensional structure basis for defect positioning; through the c3 step, the feature parameters are compared with the preset reference range, and the conforming areas that may exist defects are accurately screened out, reducing the deviation of manual judgment and preliminarily distinguishing the defect types; through the c4 step, the conforming areas are marked and a three-dimensional defect distribution map is generated, which converts abstract data into intuitive graphics and clearly presents the three-dimensional position and range of defects, ensuring the accuracy of defect area identification, and providing reliable basis for subsequent evaluation of the influence of defects on the protective performance of body armor and the development of accurate repair schemes, ensuring efficient progress of the quality inspection process.

[0142] Figure 3 The structure diagram of a body armor composite interlayer defect positioning system based on multi-modal non-destructive testing provided by the embodiments of the application is shown in Figure 3 The system comprises:

[0143] The acquisition module 31 is used for synchronously applying ultrasonic guided wave signals, pulse eddy current signals and infrared thermal wave excitation signals to the body armor composite interlayer, collecting multi-modal response signals at each position of the body armor composite interlayer, and generating an original multi-modal response signal set.

[0144] The output module 32 is configured to input the original multi-modal response signal set into the pre-trained deep convolutional neural network, determine the target frequency band parameters and target noise suppression parameters of the response signals of each mode, and output an adaptive modulation parameter group.

[0145] The optimization module 33 is configured to perform frequency domain filtering and noise suppression processing on the original multi-modal response signal set based on the adaptive modulation parameter group, generate an optimized signal set, and extract spatial frequency features of the response signals of each mode in the optimized signal set.

[0146] The generation module 34 is configured to input the spatial frequency features into the cross-modal feature fusion model, adaptively associate the defect sensitive features of different modes through the attention mechanism in the cross-modal feature fusion model, generate a three-dimensional defect distribution map of the composite interlayer, and perform millimeter-level precision defect positioning based on the three-dimensional defect distribution map.

[0147] The multi-modal nondestructive testing based body armor composite interlayer defect positioning system according to the embodiments of the present application is used to implement the multi-modal nondestructive testing based body armor composite interlayer defect positioning method described above, and therefore the specific embodiments in the multi-modal nondestructive testing based body armor composite interlayer defect positioning system can be seen from the embodiment part of the multi-modal nondestructive testing based body armor composite interlayer defect positioning method described above, and the specific embodiments can be referred to the description of the corresponding embodiment part, which will not be repeated here.

[0148] The present application also provides an electronic device, comprising: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the multi-modal nondestructive testing based body armor composite interlayer defect positioning method described above.

[0149] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the multi-modal nondestructive testing based body armor composite interlayer defect positioning method described above.

[0150] The embodiments of the present application also provide a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the steps in the multi-modal nondestructive testing based body armor composite interlayer defect positioning method described above.

[0151] Those skilled in the art will further realize that the mere concepts, teachings, and embodiments described herein are merely meant to provide examples of the various aspects of the present application and that various modifications, equivalents and alternatives are intended to fall within the scope of the present application. Accordingly, the appended claims are intended to embrace all such alterations, modifications, and additions of the present application.

[0152] The above provides a kind of based on multi-modal nondestructive testing's bulletproof vest composite sandwich defect positioning method, system, electronic equipment and storage medium provided in the application in detail.This paper applies specific example to the principle and implementation mode of the application are described, the above example is only for helping to understand the method of the application and its core idea.It should be pointed out that, for the ordinary skilled in the art, without departing from the principles of the application, the application can be improved and modified, these improvements and modifications also fall within the scope of the application.

Claims

1. A method for defect positioning of a ballistic vest composite sandwich based on multi-modal non-destructive testing, comprising: synchronously applying ultrasonic guided wave signals, pulsed eddy current signals and infrared thermal wave excitation signals to the ballistic vest composite sandwich, collecting multi-modal response signals at each position of the ballistic vest composite sandwich, and generating an original multi-modal response signal set; inputting the original multi-modal response signal set into a pre-trained deep convolutional neural network to determine target frequency band parameters and target noise suppression parameters of each modal response signal, and outputting an adaptive modulation parameter group; based on the adaptive modulation parameter group, performing frequency domain filtering and noise suppression processing on the original multi-modal response signal set to generate an optimized signal set, and extracting spatial frequency features of each modal response signal in the optimized signal set; inputting the spatial frequency features into a cross-modal feature fusion model, adaptively associating defect sensitive features of different modalities through an attention mechanism in the cross-modal feature fusion model, generating a three-dimensional defect distribution map of the composite sandwich, and based on the three-dimensional defect distribution map, performing defect positioning with millimeter-level precision; the spatial frequency features are frequency domain segmented to obtain spatial frequency features of multiple frequency bands, and a cosine similarity algorithm is used to calculate the matching degree of the spatial frequency features of each frequency band and a preset defect feature template, the defect feature template is constructed based on spatial frequency features of a preset number of known defects, and contains feature parameters of defects of different types and sizes in each frequency band, the types include delamination, cavity and crack, and the feature parameters include frequency distribution and amplitude change rate; spatial frequency features with a matching degree greater than or equal to a corresponding threshold value in each frequency band are marked as candidate features; calculating the time domain difference between the candidate features and the spatial frequency features of the normal region in the same time interval, if the time domain difference is greater than a preset difference threshold value, and the spatial coordinate deviation of the candidate features corresponding to each modality is less than a preset deviation threshold value, the candidate features are determined as defect sensitive features; the interrelation values between the defect sensitive features of different modalities are calculated through the attention mechanism in the cross-modal feature fusion model, and the weight coefficients of the defect sensitive features of each modality are determined according to the interrelation values; the defect sensitive features of different modalities are weighted and fused according to the weight coefficients to form a fusion feature; the fusion feature is layered according to the thickness direction of the ballistic vest composite sandwich to obtain feature distributions of different depth layers, and each depth layer feature distribution is used to represent the defect information of the depth layer; the feature distributions of each depth layer are sequentially superimposed to construct a three-dimensional structure containing three-dimensional spatial information, and the areas corresponding to the defect features are marked in the three-dimensional structure to obtain a three-dimensional defect distribution map; based on the three-dimensional coordinate information of the marked areas in the three-dimensional defect distribution map, the position of the defect in the composite sandwich is determined.

2. The method of claim 1, wherein, the inputting the original multi-modal response signal set into a pre-trained deep convolutional neural network to determine target frequency band parameters and target noise suppression parameters of each modal response signal, and outputting an adaptive modulation parameter group, comprises: inputting the original multi-modal response signal set into an input layer of a deep convolutional neural network, the input layer automatically matching a first processing channel corresponding to the ultrasonic guided wave signal, a second processing channel corresponding to the pulse eddy current signal, and a third processing channel corresponding to the infrared thermal wave excitation signal from the deep convolutional neural network by mode recognition; performing multi-layer convolution on the ultrasonic guided wave response signal, the pulse eddy current response signal, and the infrared thermal wave excitation response signal respectively by using the first processing channel, the second processing channel, and the third processing channel to obtain ultrasonic frequency domain features, eddy current frequency domain features, and thermal wave frequency domain features; determining target frequency band parameters and target noise suppression parameters of each modal response signal based on the ultrasonic frequency domain features, the eddy current frequency domain features, and the thermal wave frequency domain features; integrating the target frequency band parameters and the target noise suppression parameters of all modal response signals to obtain an adaptive modulation parameter group.

3. The method of claim 2, wherein, The method for determining the target frequency band parameters and the target noise suppression parameters of each modal response signal based on the ultrasonic frequency domain features, the eddy current frequency domain features, and the thermal wave frequency domain features comprises: calculating a response intensity value corresponding to each feature based on the fluctuation amplitude and the duration of each feature in the corresponding frequency band; obtaining an intensity level corresponding to the response intensity value from a preset mapping relationship between intensity levels and intensity ranges, and taking the product of a reference frequency band parameter and a frequency band coefficient corresponding to the intensity level as the frequency band parameter of the corresponding modal response signal, wherein the features include the ultrasonic frequency domain features, the eddy current frequency domain features, and the thermal wave frequency domain features; determining the target noise suppression parameters of each modal response signal according to the frequency band parameters of the modal response signals and in combination with the filtering mode corresponding to the modal; and different filtering modes correspond to different modals.

4. The method of claim 1, wherein, The method for performing frequency domain filtering and noise suppression processing on the original multi-modal response signal set based on the adaptive modulation parameter group to generate an optimized signal set comprises: for each modal response signal in the original multi-modal response signal set, retaining a frequency band part of the modal response signal that is higher than or equal to the corresponding target frequency band parameter and removing a frequency band part of the modal response signal that is lower than the corresponding target frequency band parameter according to the corresponding target frequency band parameter to complete frequency band filtering, wherein the modal response signals include ultrasonic guided wave response signals, pulse eddy current response signals, and infrared thermal wave excitation response signals; for the ultrasonic guided wave response signal, using a wavelet threshold denoising algorithm, taking a threshold coefficient in the corresponding target noise suppression parameter as a threshold reference for wavelet decomposition, and attenuating high-frequency noise in the ultrasonic guided wave response signal; for the pulse eddy current response signal, using an adaptive Kalman filtering algorithm, taking a covariance matrix in the corresponding target noise suppression parameter as a filter iteration initial value, and dynamically correcting eddy current interference noise in the pulse eddy current response signal. For the infrared thermal wave excitation response signal, a combination of Gaussian filtering and median filtering is adopted, the thermal noise fluctuation in the infrared thermal wave excitation response signal is smoothed through Gaussian filtering, and the salt and pepper noise in the infrared thermal wave excitation response signal is removed through median filtering; The signal intensity normalization processing is performed on each modal response signal after the noise suppression processing, the amplitude range of all modal response signals is uniformly mapped to a preset interval, so as to eliminate the order difference between different modal response signals; The normalized ultrasonic guided wave response signal, the normalized pulse eddy current response signal and the normalized infrared thermal wave excitation response signal are aligned in time axis and space axis, so as to ensure that the spatial coordinates and time stamps of different modal response signals at the same sampling point correspond to each other, and an optimized signal set is integrated.

5. The method of claim 4, wherein, The normalized ultrasonic guided wave response signal, the normalized pulse eddy current response signal and the normalized infrared thermal wave excitation response signal are aligned in time axis and space axis, so as to ensure that the spatial coordinates and time stamps of different modal response signals at the same sampling point correspond to each other, and an optimized signal set is integrated, including: A three-dimensional space coordinate system is constructed with the geometric center of the body armor composite interlayer as the coordinate origin, the coordinates of the sampling points of each modal response signal are converted to the three-dimensional space coordinate system to obtain three-dimensional coordinate values, wherein the x-axis and the y-axis correspond to the interlayer plane direction, and the z-axis corresponds to the interlayer thickness direction; The timestamp information of each modal response signal is extracted, and the timestamps of different modal signals are calibrated according to a preset precision clock synchronization protocol, so that the time reference error of all modal response signals is controlled within a preset error range; The three-dimensional coordinate values are subjected to grid processing, and the body armor composite interlayer is divided into a plurality of three-dimensional grid units of a preset size, and each three-dimensional grid unit corresponds to a unique spatial index; The nearest neighbor interpolation method is used to map the sampling points of each modal response signal to the corresponding three-dimensional grid unit, and if there is a three-dimensional grid unit lacking a sampling point, the three-dimensional grid unit lacking a sampling point is taken as the center, and the signal values of the adjacent 8 grid units are weighted and calculated; A space-time alignment verification matrix is generated, the elements in the space-time alignment verification matrix are signal existence identifiers of each modal response signal at the same three-dimensional grid unit and timestamp, when the missing rate of the signal existence identifier exceeds a preset missing rate threshold, the above alignment processing procedure is re-executed until the missing rate of the signal existence identifier is less than or equal to the preset missing rate threshold, and the optimized signal set is integrated in the order of three-dimensional grid unit position.

6. The method of claim 1, wherein, The feature distribution of each depth layer is sequentially superimposed to construct a three-dimensional structure containing three-dimensional space information, the area consistent with the defect feature is marked in the three-dimensional structure, and a three-dimensional defect distribution map is obtained, including: The feature distribution of each depth layer is sorted according to the depth from shallow to deep, and each feature distribution corresponds to a different actual depth position; The sorted feature distributions are sequentially stacked along the depth direction, and in the stacking process, the transverse and longitudinal positions in the spatial coordinates of different depth levels correspond to each other, thereby forming a three-dimensional structure including multiple depth levels; In the three-dimensional structure, it is checked whether the defect features of each actual depth position fall within a preset feature reference range, and the actual depth positions of the defect features falling within the preset feature reference range are classified as consistent regions. The preset feature reference range is determined based on the feature parameters of known defects and covers feature thresholds of different defect types; All the consistent regions in the three-dimensional structure are marked as a whole to form a three-dimensional diagram including marked regions, and the three-dimensional defect distribution map is generated based on the three-dimensional diagram.

7. An anti-ballistic vest composite interlayer defect positioning system based on multi-modal non-destructive testing, comprising: A collection module for synchronously applying ultrasonic guided wave signals, pulsed eddy current signals and infrared thermal wave excitation signals to the anti-ballistic vest composite interlayer, collecting multi-modal response signals at each position of the anti-ballistic vest composite interlayer, and generating an original multi-modal response signal set; An output module for inputting the original multi-modal response signal set into a pre-trained deep convolutional neural network, determining target frequency band parameters and target noise suppression parameters of each modal response signal, and outputting an adaptive modulation parameter group; An optimization module for performing frequency domain filtering and noise suppression processing on the original multi-modal response signal set based on the adaptive modulation parameter group, generating an optimized signal set, and extracting spatial frequency features of each modal response signal in the optimized signal set; A generation module for inputting the spatial frequency features into a cross-modal feature fusion model, adaptively associating defect sensitive features of different modalities through an attention mechanism in the cross-modal feature fusion model, generating a three-dimensional defect distribution map of the composite interlayer, and performing millimeter-level precision defect positioning based on the three-dimensional defect distribution map; the spatial frequency features of the modal response signals are frequency domain segmented to obtain spatial frequency features of multiple frequency bands, and a cosine similarity algorithm is used to calculate the matching degree of each frequency band spatial frequency feature and a preset defect feature template, the defect feature template is constructed based on a preset number of known defect spatial frequency features and contains feature parameters of different types and sizes of defects in each frequency band, the types include delamination, void and crack, and the feature parameters include frequency distribution and amplitude change rate; Spatial frequency features with a matching degree greater than or equal to a corresponding threshold in each frequency band are marked as candidate features; The time domain difference between the candidate features and the spatial frequency features of the normal region in the same time interval is calculated, and if the time domain difference is greater than a preset difference threshold and the spatial coordinate deviation of the candidate features of each modality is less than a preset deviation threshold, the candidate features are determined as defect sensitive features; The cross-correlation values between the defect sensitive features of different modalities are calculated through the attention mechanism in the cross-modal feature fusion model, and the weight coefficients of the defect sensitive features of each modality are determined according to the cross-correlation values; The defect sensitive features of different modalities are weighted and fused according to the weight coefficients to form a fusion feature. The fusion features are layered according to the thickness direction of the composite interlayer of the body armor, to obtain feature distributions at different depth levels, and each feature distribution at a depth level is used to represent defect information at the depth level; The feature distributions at the depth levels are sequentially superimposed to construct a three-dimensional structure containing three-dimensional space information, and a region corresponding to the defect features is marked in the three-dimensional structure to obtain a three-dimensional defect distribution map; Based on three-dimensional coordinate information of the marked region in the three-dimensional defect distribution map, the position of the defect in the composite interlayer is determined.

8. An electronic device, comprising: It comprises: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the body armor composite interlayer defect positioning method based on multi-modal non-destructive testing according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program can implement the body armor composite interlayer defect positioning method based on multi-modal non-destructive testing according to any one of claims 1 to 6 when executed by the processor.

Images