Multi-mode signal fusion detection method and system of electromagnetic ultrasonic transducer

Abstract

The invention relates to the technical field of nondestructive testing, and discloses a multi-modal signal fusion detection method and system for an electromagnetic ultrasonic transducer, and the method comprises the steps: exciting and receiving ultrasonic signals of a plurality of propagation modes in a detected medium based on the electromagnetic ultrasonic transducer, obtaining multi-modal original data containing different physical characteristics, time synchronization and preprocessing are carried out; feature parameters of the internal state of the medium are extracted, normalization is carried out according to physical attributes, and a unified multi-modal feature space is constructed; setting a dynamic weight strategy based on the sensitivity distribution of the characteristic parameters and the structure parameters, and executing information fusion to form a multi-modal characteristic vector; carrying out comprehensive interpretation analysis to generate a detection conclusion containing abnormal region coordinate data and spatial boundary data; and outputting the corresponding visual representation information. According to the method, signals of different propagation modes are cooperatively utilized, and the defect identification capability and the space positioning precision in a complex structure are improved.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and more specifically, to a method and system for multimodal signal fusion detection of electromagnetic ultrasonic transducers. Background Technology

[0002] With the increasing demands for safety and reliability of key components in industries such as high-end equipment manufacturing, energy transportation, and aerospace, the need for detecting internal defects, microcracks, and structural anomalies in materials is becoming increasingly urgent. Traditional ultrasonic nondestructive testing methods typically rely on piezoelectric transducers for excitation and reception, but they have the following shortcomings: strong dependence on couplants: piezoelectric transducers usually require couplants to assist signal transmission during the testing process, which is not conducive to applications in complex environments such as high temperature, high speed, or strong corrosion; single mode: traditional methods often focus on ultrasonic signals with a single propagation mode, which is easily affected by noise interference, resulting in insufficient stability and accuracy in defect identification; limited feature utilization: existing detection methods are mostly limited to single-dimensional feature analysis in the time or frequency domains, failing to fully explore the comprehensive physical information contained in multimodal signals.

[0003] Electromagnetic ultrasonic transducers (EMATs) are gradually becoming an important tool in advanced nondestructive testing due to their non-contact nature and ability to directly excite and receive ultrasonic waves on high-temperature or rough surfaces. However, most existing detection methods based on electromagnetic ultrasonic transducers are still limited to single-mode signal analysis and lack the synergistic utilization of signals from different propagation modes.

[0004] Therefore, it is necessary to design a multimodal signal fusion detection method and system for electromagnetic ultrasonic transducers to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a multi-mode signal fusion detection method and system for electromagnetic ultrasonic transducers, aiming to solve the problems of lack of coordinated utilization of signals of different propagation modes, poor ability to identify defects in complex structures, and low spatial positioning accuracy.

[0006] In one aspect, the present invention proposes a multi-modal signal fusion detection method for electromagnetic ultrasonic transducers, comprising: Based on the electromagnetic ultrasonic transducer, ultrasonic signals with multiple propagation modes are excited and received in the measured medium to obtain multimodal raw signal data containing different physical properties; the multimodal raw signals are then subjected to time synchronization processing and data preprocessing. Based on the multimodal raw signal data after the data preprocessing operation, feature parameters of the internal state of the measured medium are extracted; and the feature parameters are normalized according to physical properties and integrated into a unified multimodal feature space. In the multimodal feature space, a dynamic weight allocation strategy is set according to the sensitivity distribution of the feature parameters to specific structural parameters of the tested medium, and information fusion is performed to form a multimodal feature vector; the multimodal feature vector is comprehensively interpreted and analyzed to generate a detection conclusion; the detection conclusion is the coordinate data of the abnormal area of ​​the internal structure of the tested medium and the corresponding spatial boundary data. Based on the detection results, output visual representation information corresponding to the coordinate data and spatial boundary data of the abnormal region.

[0007] Furthermore, when performing time synchronization processing and data preprocessing operations on the multimodal raw signal, the following are included: The time synchronization process involves aligning the sampling points of each modal ultrasound signal in phase based on a reference clock signal; the data preprocessing operations include performing bandpass filtering, DC component removal, and amplitude normalization. The time synchronization process and data preprocessing operation are executed sequentially, and the data preprocessing operation is limited to being performed after the time synchronization process is completed.

[0008] Furthermore, when extracting feature parameters of the internal state of the measured medium from the multimodal raw signal data after the data preprocessing operation, the following steps are included: Based on ultrasonic signals of different propagation modes, time-domain characteristic parameters, frequency-domain characteristic parameters, and instantaneous characteristic parameters directly related to the internal structure of the medium are extracted respectively. The time-domain characteristic parameters cover the peak amplitude and zero-crossing rate of the signal envelope, the frequency-domain characteristic parameters cover the energy distribution and spectral centroid of the dominant frequency component, and the instantaneous characteristic parameters cover the instantaneous phase change rate and the analytical signal envelope derived from the Hilbert transform.

[0009] Furthermore, when normalizing the feature parameters according to physical properties, the process includes: Based on the physical dimensions and dynamic range of the characteristic parameters, dimensionless normalization is performed using a linear transformation based on the minimum-maximum scale.

[0010] Furthermore, when integrating into a unified multimodal feature space, this includes: The normalized feature parameters are divided into elastic modulus-related parameters, density-related parameters, and attenuation coefficient-related parameters according to the physical property category; the divided feature parameters are integrated into the multimodal feature space composed of a multidimensional Cartesian coordinate system according to the preset physical property dimension order. Each coordinate axis of the multimodal feature space corresponds to a normalized feature parameter.

[0011] Furthermore, when setting a dynamic weight allocation strategy based on the sensitivity distribution of the characteristic parameters to specific structural parameters of the measured medium, the strategy includes: Based on the local correlation between the feature parameters and specific structural parameters, the sensitivity distribution function of each feature parameter in the multimodal feature space is determined; the dynamic weight allocation strategy adjusts the weight coefficients in real time according to the sensitivity distribution function, and the allocation of the weight coefficients follows the principle that the higher the sensitivity value, the greater the weight. The sensitivity distribution function is constructed based on the gradient change characteristics of the feature parameters in the feature space.

[0012] Furthermore, when performing information fusion to form multimodal feature vectors, the following steps are included: The information fusion is based on a weighted linear combination, which multiplies the feature parameters with their corresponding weight coefficients and then sums them to generate a multimodal feature vector that represents the comprehensive characteristics of the multimodal signal.

[0013] Furthermore, when performing comprehensive interpretation and analysis on the multimodal feature vectors to generate detection conclusions, the following steps are included: The multimodal feature vector is input into the judgment rule base, which includes anomaly discrimination threshold range and boundary recognition logic constructed based on historical detection data; The comprehensive interpretation and analysis performs multi-level comparison operations; By comparing the Euclidean distance between the feature vector and the reference vector in the normal state, abnormal regions can be identified. The spatial boundary of the abnormal region is determined by the gradient direction change of the feature vector in the feature space.

[0014] Furthermore, when outputting visual representation information corresponding to the coordinate data and spatial boundary data of the abnormal region based on the detection conclusion, it includes: Map the coordinate data of the abnormal area and the spatial boundary data in the detection conclusion to the three-dimensional geometric model of the medium under test; The coordinate data of the abnormal region is based on the superposition of bright color blocks on the surface of the three-dimensional geometric model, and the spatial boundary data is based on the outline annotation; the visualization representation information of the text annotation layer containing the location, size and feature parameter association information of the abnormal region is generated.

[0015] Compared with existing technologies, the advantages of this invention are as follows: It employs an electromagnetic ultrasonic transducer to acquire ultrasonic signals with multiple propagation modes, avoiding the dependence on coupling agents found in traditional piezoelectric transducers. This allows for stable operation under complex conditions such as high temperature, strong corrosion, or rough surfaces, improving the environmental adaptability and reliability of the detection. By performing time synchronization processing and data preprocessing on the multimodal raw signals, and extracting feature parameters from three dimensions—time domain, frequency domain, and instantaneous characteristics—it can comprehensively characterize the internal physical state of the measured medium, avoiding the information loss problem caused by noise interference in single-mode signals. Normalizing the feature parameters of different physical properties and integrating them into a unified multimodal feature space ensures the comparability between features of different dimensions, facilitating subsequent weight allocation and fusion calculations, thereby improving the scientific rigor and accuracy of the detection data processing. A dynamic weight allocation strategy is constructed based on the sensitivity distribution of feature parameters to structural parameters, allowing the weight coefficients to be adjusted in real time according to detection conditions and local characteristics, ensuring high sensitivity response to abnormal areas and improving the accuracy of defect identification and location. Information fusion is achieved through weighted linear combination to obtain a multimodal feature vector with comprehensive characteristics. This vector is then combined with a rule base built on historical data to perform multi-level comparisons, reducing the false alarm rate and improving the accuracy of anomaly region identification and spatial boundary determination. The detection results are visualized using a 3D geometric model, combined with highlighted color blocks, contour lines, and text annotation layers. This visualization displays the location, extent, and relevant feature parameters of the anomaly regions, providing a clear basis for subsequent structural evaluation and decision-making.

[0016] On the other hand, this application also provides a multi-modal signal fusion detection system for an electromagnetic ultrasonic transducer, used to apply the above-mentioned multi-modal signal fusion detection method for an electromagnetic ultrasonic transducer, comprising: The acquisition unit is configured to excite and receive ultrasonic signals with multiple propagation modes in the medium under test based on an electromagnetic ultrasonic transducer, acquire multimodal raw signal data containing different physical characteristics, and perform time synchronization processing and data preprocessing operations on the multimodal raw signals. The processing unit is configured to extract feature parameters of the internal state of the measured medium based on the multimodal raw signal data after the data preprocessing operation; and to normalize the feature parameters according to physical properties and integrate them into a unified multimodal feature space. The generation unit is configured to, in the multimodal feature space, set a dynamic weight allocation strategy based on the sensitivity distribution of the feature parameters to specific structural parameters of the tested medium, and perform information fusion to form a multimodal feature vector; perform comprehensive interpretation and analysis on the multimodal feature vector to generate a detection conclusion; the detection conclusion is the coordinate data of the abnormal area of ​​the internal structure of the tested medium and the corresponding spatial boundary data; The output unit is configured to output visual representation information corresponding to the coordinate data of the abnormal region and the spatial boundary data based on the detection conclusion.

[0017] It is understandable that the above-mentioned multimodal signal fusion detection method and system for electromagnetic ultrasonic transducers have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a multimodal signal fusion detection method for an electromagnetic ultrasonic transducer provided in an embodiment of the present invention; Figure 2 This is a functional block diagram of a multimodal signal fusion detection system for an electromagnetic ultrasonic transducer provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Existing electromagnetic ultrasonic testing technologies mostly focus on the analysis of single-mode signals, typically using only a single characteristic parameter in the time or frequency domain for judgment. This makes the detection results susceptible to environmental noise and signal attenuation, resulting in insufficient defect identification sensitivity, low positioning accuracy, and a high false positive rate. Current methods generally lack collaborative processing mechanisms for multi-mode signals, failing to fully exploit the complementary information contained in different propagation modes, thus limiting their application effectiveness in complex structures and variable operating conditions.

[0021] For example, in the service inspection of a high-speed railway rail, when using the traditional electromagnetic ultrasonic testing method based on single-mode ultrasonic signals for crack identification, the ultrasonic signal amplitude was severely attenuated due to the surface roughness and local permeability variations of the rail. The reflected signals of some micro-cracks were masked by background noise and ultimately failed to appear in the inspection results. This not only reduces the reliability of the inspection but may also pose structural safety hazards, indicating that the traditional single-mode testing method has significant limitations in complex scenarios.

[0022] For this, please refer to Figure 1 As shown, a multi-modal signal fusion detection method for an electromagnetic ultrasonic transducer includes: S100: Based on an electromagnetic ultrasonic transducer, it excites and receives ultrasonic signals with multiple propagation modes in the measured medium to acquire multimodal raw signal data containing different physical characteristics; and performs time synchronization processing and data preprocessing operations on the multimodal raw signals. S200: Extract feature parameters of the internal state of the measured medium from the multimodal raw signal data after data preprocessing; normalize the feature parameters according to physical properties and integrate them into a unified multimodal feature space; S300: In the multimodal feature space, a dynamic weight allocation strategy is set according to the sensitivity distribution of feature parameters to specific structural parameters of the measured medium, and information fusion is performed to form a multimodal feature vector; the multimodal feature vector is comprehensively interpreted and analyzed to generate a detection conclusion; the detection conclusion is the coordinate data of the abnormal area of ​​the internal structure of the measured medium and the corresponding spatial boundary data. S400: Outputs visual representation information corresponding to the coordinate data and spatial boundary data of the abnormal area based on the detection results.

[0023] Specifically, step S100 uses an electromagnetic ultrasonic transducer to excite and receive ultrasonic signals of various propagation modes (such as shear waves, surface waves, and volume waves) on the surface of the measured medium in a non-contact manner. A unified reference clock is used to sample each channel to ensure time base consistency. The acquired multimodal raw signals are first precisely phase-aligned according to the reference clock (phase / sample-level synchronization), and then preprocessed sequentially. This includes configurable bandpass filtering (removing power frequency and external interference bands; the passband is set according to the selected frequency band), DC component removal, artifact removal, and amplitude normalization (e.g., dividing each channel signal by its short-time energy or peak value to suppress magnitude differences). After preprocessing, the results are sent to a buffer queue for subsequent feature calculations. In step S200... In this process, multiple feature parameters are extracted for each preprocessed modal signal according to its propagation mode: in the time domain, envelope peak, zero-crossing rate, pulse width, and arrival time are extracted; in the frequency domain, power spectral density, energy distribution of the dominant frequency component, spectral centroid, and bandwidth are calculated; for instantaneous features, the analytic signal envelope, instantaneous phase and its rate of change (phase derivative), and instantaneous frequency are obtained through Hilbert transform; to eliminate the influence of different physical dimensions, each feature parameter is dimensionless by using a minimum-maximum linear transformation according to its dimension and dynamic range, and the features are classified into categories such as elastic modulus correlation, density correlation, and attenuation coefficient correlation according to physical properties, and mapped to a multi-dimensional Cartesian coordinate system in a predetermined attribute dimension order to construct a unified multimodal feature space. In step S300, based on the local correlation between feature parameters and the parameters of the measured structure (which can be obtained through historical labeled data or model sensitivity analysis), a sensitivity distribution function is established in the feature space—for example, the response strength of a feature to the target parameter is characterized by the feature gradient magnitude or local correlation coefficient. Based on this sensitivity distribution, a dynamic weight allocation strategy is used to calculate the weight coefficients in real time (normalized exponential mapping can be used), and each feature parameter is multiplied and summed with its corresponding weight in a weighted linear combination to generate a multimodal feature vector characterizing the comprehensive physical properties. This feature vector is then input into the judgment rule base to perform multi-level comparison (including Euclidean distance criterion based on historical normal state reference vector, threshold interval comparison, and multi-model cross-validation). The presence of anomalies is determined by the Euclidean distance with the reference vector and the gradient direction change in the feature space, and the boundary contour of the abnormal region is extracted based on the gradient direction and magnitude change of the feature vector in space. In step S400, the detection results (coordinates and boundary data of abnormal areas) are mapped onto the three-dimensional geometric model of the tested medium: the abnormal areas are superimposed on the model surface as highlighted color blocks or semi-transparent voxels, the spatial boundaries are marked with outlines, and a text annotation layer is generated to display the abnormal location, size, relevant feature parameters and confidence score; at the same time, coordinate data and visualization models in standard formats (such as CSV / JSON) can be exported for subsequent evaluation or maintenance decisions.

[0024] Step S100 involves using an electromagnetic ultrasonic transducer to excite and receive ultrasonic signals with multiple propagation modes, achieving comprehensive acquisition of the multidimensional physical properties of the measured medium. Simultaneously, time synchronization processing ensures precise alignment of each modal signal on the time axis, guaranteeing time-base consistency for subsequent data analysis. Data preprocessing operations include filtering, DC component removal, and amplitude normalization, aiming to remove noise and signal offset, improving the reliability and stability of feature extraction. Step S200 extracts feature parameters related to the internal state of the medium from the preprocessed multimodal signals, including time-domain, frequency-domain, and instantaneous features, to comprehensively reflect the internal structural information of the medium. Subsequently, the feature parameters are normalized according to physical properties and mapped to a unified multimodal feature space, thereby eliminating the influence of different dimensions and providing a unified platform for feature fusion. Step S300 further enhances information utilization efficiency in the feature space: based on the sensitivity distribution of each feature parameter to specific structural parameters, weights are dynamically allocated to form a weighted fusion multimodal feature vector, realizing the synergistic effect of different features; a comprehensive interpretation and analysis of this feature vector is performed, and abnormal regions can be identified and their spatial boundaries determined through multi-level comparison, Euclidean distance calculation, and gradient direction analysis, achieving precise positioning. Step S400 maps the detection results onto a three-dimensional geometric model, visually displaying abnormal regions with highlighted color blocks or contour lines, while generating a text annotation layer to record the location, size, and related feature information of the abnormality, realizing the visualization and interpretability of the detection results.

[0025] The working principle and process of this application are as follows: In step S100, an electromagnetic ultrasonic transducer non-contactly excites and receives ultrasonic signals of various propagation modes, including shear waves, longitudinal waves, and surface waves, on the surface of the measured medium. By acquiring multimodal raw signals with different physical properties, a comprehensive perception of the internal structure information of the medium is achieved. The acquired signals are first subjected to time synchronization processing to ensure that the phase of each modal signal is aligned under the reference clock, so as to eliminate the influence of sampling time differences on subsequent analysis. Subsequently, data preprocessing operations are performed, including bandpass filtering, DC component removal, and amplitude normalization, to remove noise and interference and unify the signal magnitude, providing a reliable basis for feature extraction. In step S200, feature parameters are extracted from the preprocessed multimodal signals, including time-domain features (such as envelope peak value, zero-crossing rate), frequency-domain features (such as the energy distribution of the dominant frequency component, spectral centroid), and instantaneous features (such as instantaneous phase change rate, Hilbert transform analytic signal envelope), comprehensively characterizing the internal state of the medium. The extracted feature parameters are normalized according to their physical properties to eliminate dimensional differences and integrated into a unified multimodal feature space, enabling effective comparison and fusion of various features within the same coordinate system. In step S300, a dynamic weight allocation strategy is set based on the sensitivity distribution of each feature parameter to specific structural parameters of the tested medium, adjusting the weights of different features in the information fusion process in real time. The weighted feature parameters are linearly combined to form a multimodal feature vector, fully utilizing the complementary information of each modal signal. Subsequently, this feature vector undergoes comprehensive interpretation and analysis. Through multi-level comparison with historical reference vectors, Euclidean distance calculation, and feature space gradient analysis, the identification of abnormal regions and precise positioning of spatial boundaries are achieved, generating coordinate data of abnormal regions within the tested medium's internal structure and corresponding spatial boundary data. In step S400, the detection results are mapped onto the three-dimensional geometric model of the tested medium. Abnormal regions are displayed through highlighted color blocks or semi-transparent voxels, and spatial boundaries are marked with contour lines. Simultaneously, a text annotation layer is generated to record the abnormal location, size, and related feature parameter information, achieving a visual representation of the detection results.

[0026] As a preferred embodiment, the solution of this application is implemented as follows: For example, during routine safety inspections of in-service rails, electromagnetic ultrasonic transducers are placed on the rail surface to excite and receive ultrasonic signals of various propagation modes, such as longitudinal waves, shear waves, and surface waves, thereby obtaining multimodal raw signals containing different physical characteristics. Subsequently, the signals undergo time synchronization processing and data preprocessing, including bandpass filtering to remove environmental noise, DC component removal, and amplitude normalization, to ensure the reliability and comparability of each modal signal. Feature parameters related to the internal structure of the rail are extracted from the preprocessed signals, such as the envelope peak and zero-crossing rate in the time domain, the dominant frequency energy distribution and spectral centroid in the frequency domain, and the phase change rate and analytical signal envelope of instantaneous features. These feature parameters are normalized and integrated into a unified multimodal feature space, providing a standardized data foundation for subsequent analysis. The system dynamically assigns weights based on the sensitivity distribution of feature parameters to local structural parameters of the rail, and forms a multimodal feature vector through weighted linear combination. Multi-level comparative analysis is performed on the feature vector, including Euclidean distance calculation and gradient direction analysis with historical normal state reference vectors, thereby accurately identifying the location and spatial boundary of microcracks inside the rail. The detection results are mapped onto the three-dimensional geometric model of the rail, with abnormal areas marked by highlighted blocks, spatial boundaries displayed by contour lines, and additional text annotations recording crack size, depth, and related feature parameters. This process effectively identifies microcracks that are not directly observable to the naked eye.

[0027] Through the above technical solution, this application achieves high-precision detection of internal structural anomalies in the tested medium by acquiring and fusing multimodal signals using an electromagnetic ultrasonic transducer. By exciting and receiving ultrasonic signals of multiple propagation modes and performing time synchronization and preprocessing on the original multimodal signals, the accuracy and consistency of the data are ensured. Time-domain, frequency-domain, and instantaneous feature parameters are extracted and their physical properties are normalized to construct a unified multimodal feature space, providing a standardized foundation for information fusion. Weights are dynamically allocated based on the sensitivity distribution of features and structural parameters, forming a weighted multimodal feature vector. Comprehensive interpretation and analysis accurately identify abnormal regions and spatial boundaries. The detection results are mapped to a three-dimensional geometric model, achieving highly visualized representation. This application utilizes complementary information from multimodal signals to improve detection sensitivity and accuracy; optimizes feature fusion effects through dynamic weight allocation, reducing false alarms and false negatives; and displays the location and spatial range of anomalies.

[0028] This application further proposes time synchronization processing and data preprocessing operations for multimodal raw signals, including: Time synchronization processing involves phase alignment of the sampling points of each modal ultrasound signal based on a reference clock signal; data preprocessing operations include performing bandpass filtering, DC component removal, and amplitude normalization. The time synchronization process and data preprocessing are executed sequentially, and the data preprocessing operation is limited to being performed after the time synchronization process is completed.

[0029] Specifically, when processing the multimodal raw signals, a time synchronization operation is performed. The core of this operation is to precisely align the sampling points of each modal ultrasonic signal with the reference clock signal to eliminate potential time delays or phase deviations between different acquisition channels. This ensures complete consistency of the modal signals on the time axis in subsequent analysis, thereby avoiding feature extraction errors or signal fusion deviations caused by time asynchrony. After time synchronization, data preprocessing operations are performed, including bandpass filtering to remove low-frequency environmental noise and high-frequency interference from the ultrasonic signal, ensuring the signal spectrum is concentrated in the effective frequency band; DC component removal to eliminate signal baseline drift, making signal amplitude fluctuations more accurately reflect the internal characteristics of the medium; and amplitude normalization to unify the amplitude range of each modal signal, avoiding inconsistencies in characteristic dimensions due to differences in sensor responses. These preprocessing operations are performed in a strict sequence, meaning they must be performed after time synchronization to ensure signal consistency in the time domain and amplitude, providing high-quality, comparable input data.

[0030] As a preferred embodiment, the solution of this application is implemented as follows: For example, when detecting internal defects in steam pipelines, time synchronization and data preprocessing of multimodal ultrasonic signals are crucial. Electromagnetic ultrasonic transducers arranged on the pipeline surface collect multimodal signals such as longitudinal waves and shear waves. Due to slight time differences in the acquisition of different transducer channels and propagation modes, precise phase alignment of the sampling points of each modal signal must be performed through time synchronization processing based on a reference clock signal to ensure the temporal consistency of signals from different channels and avoid deviations in subsequent feature analysis. Data preprocessing is performed on the synchronized signals, including bandpass filtering to remove low-frequency environmental noise and high-frequency mechanical vibration interference, concentrating the signal spectrum in the effective detection frequency band; DC component removal to eliminate sensor baseline drift, ensuring that signal fluctuations accurately reflect the internal material characteristics of the pipeline; and amplitude normalization processing to unify the amplitude range of each modal signal, eliminating amplitude inconsistencies caused by differences in transducer sensitivity or uneven coupling. The entire processing flow strictly follows the order of time synchronization followed by data preprocessing to ensure the consistency of each modal signal in the time domain, frequency domain, and amplitude.

[0031] Through the above technical solutions, this application eliminates the time delay between different mode signals based on the phase alignment of the reference clock signal, ensuring that each mode is accurately superimposed and compared under the same time reference; bandpass filtering can remove environmental noise and high-frequency interference, so that the signal is focused on the effective frequency band; DC component removal eliminates baseline drift and improves the true reflection capability of the signal; amplitude normalization processing unifies the magnitude of each mode signal and avoids characteristic deviations caused by differences in sensor sensitivity.

[0032] This application further proposes methods for extracting characteristic parameters of the internal state of the measured medium from multimodal raw signal data after data preprocessing, including: Based on the ultrasonic signals of different propagation modes, time-domain characteristic parameters, frequency-domain characteristic parameters, and instantaneous characteristic parameters directly related to the internal structure of the medium are extracted respectively. The time-domain characteristic parameters cover the peak amplitude and zero-crossing rate of the signal envelope, the frequency-domain characteristic parameters cover the energy distribution and spectral centroid of the dominant frequency component, and the instantaneous characteristic parameters cover the instantaneous phase change rate and the analytical signal envelope derived from the Hilbert transform.

[0033] Specifically, in the process of extracting internal state characteristic parameters of the measured medium from the multimodal raw signals after data preprocessing, detailed analysis is performed on ultrasonic signals of different propagation modes. For time-domain features, the peak amplitude and zero-crossing rate of the signal envelope are extracted. These parameters directly reflect the inhomogeneity, presence of defects, and energy distribution of the material within the medium, thus providing important evidence for identifying cracks, pores, or inclusions. Frequency-domain features, by analyzing the energy distribution and spectral centroid of the dominant frequency component of the signal, reveal the propagation characteristics of ultrasound in the medium and the differences in response to different structural parameters, helping to distinguish local stiffness, density changes, and attenuation characteristics of the material. Instantaneous features further utilize Hilbert transform to obtain the analytical signal envelope and instantaneous phase change rate, thereby capturing phase fluctuations and instantaneous energy changes caused by minute structural changes. This is particularly crucial for the rapid detection of micro-cracks or micro-defects such as interface delamination. By simultaneously integrating time-domain, frequency-domain, and instantaneous feature parameters, a comprehensive and refined characterization of the internal state of the measured medium can be formed, providing a rich and reliable data foundation for subsequent multimodal feature space integration, dynamic weighting, and information fusion.

[0034] As a preferred embodiment, the solution of this application is implemented as follows: For example, when performing non-destructive testing on turbine blades, the extraction of feature parameters using multimodal ultrasonic signals can effectively identify micro-cracks or material defects inside the blade. Time-domain feature parameters, such as the peak amplitude of the signal envelope, can reflect the local reflection intensity of ultrasonic waves when they encounter defects inside the blade, while the zero-crossing rate can be used to analyze the periodic structure and internal discontinuities of the material, thereby determining the presence of cracks or pores. Frequency-domain feature parameters, by analyzing the energy distribution and spectral centroid of the signal's dominant frequency components, can reveal the energy concentration in different frequency bands, helping to determine local stiffness or density changes in the material; for example, crack regions can cause spectral energy to shift to higher or lower frequencies. Instantaneous feature parameters, using the analytic signal envelope and instantaneous phase change rate obtained by Hilbert transform, can capture instantaneous energy fluctuations and phase anomalies caused by crack propagation or interface delamination. The combined use of these feature parameters can comprehensively characterize the internal state of the turbine blade and provide high-precision defect location information.

[0035] Through the above technical solutions, this application improves the characterization accuracy and detection sensitivity of the internal structure of the measured medium by extracting time-domain, frequency-domain, and instantaneous characteristic parameters from the multimodal raw signal. Time-domain characteristic parameters, such as the peak amplitude and zero-crossing rate of the signal envelope, can directly reflect the discontinuities and local defects inside the material, providing reliable indicators for cracks, pores, or delamination. Frequency-domain characteristic parameters reveal the local stiffness, density, and attenuation characteristics of the medium through the energy distribution of the dominant frequency component and the center of spectral centroid, which helps to distinguish between normal material regions and abnormal regions. Instantaneous characteristic parameters use the instantaneous phase change rate and Hilbert transform to analyze the signal envelope and capture the instantaneous energy fluctuations and phase anomalies caused by minute structural changes, thereby enhancing the sensitivity to microcracks or interface defects.

[0036] This application further proposes that when normalizing feature parameters according to physical properties, the following steps are included: Based on the physical dimensions and dynamic range of the characteristic parameters, a dimensionless normalization is performed using a linear transformation based on the minimum-maximum scale.

[0037] Specifically, when normalizing the extracted feature parameters, a minimum-maximum scale linear transformation method is used to map each feature to a uniform dimensionless interval based on its physical dimensions and dynamic range. For example, assuming the original value range of a feature parameter is Xmin and Xmax, normalization maps the feature to the interval [0, 1]. This process eliminates differences between different feature dimensions and magnitudes, allowing various feature parameters to be compared and analyzed at a uniform scale.

[0038] Through the above technical solution, this application avoids excessive influence of large-amplitude features on the weight of subsequent multimodal feature fusion by normalization processing, and also ensures that small-amplitude features can be fully utilized in the information fusion process, thereby improving the accuracy and reliability of comprehensive interpretation.

[0039] This application further proposes integration into a unified multimodal feature space, including: The normalized feature parameters are divided into elastic modulus-related parameters, density-related parameters, and attenuation coefficient-related parameters according to the physical property category. The divided feature parameters are then integrated into a multimodal feature space composed of a multidimensional Cartesian coordinate system according to the preset physical property dimension order. In this multimodal feature space, each coordinate axis corresponds to a normalized feature parameter.

[0040] Specifically, in the process of integrating into a unified multimodal feature space, the normalized feature parameters are classified according to their physical properties, into three categories: elastic modulus-related, density-related, and attenuation coefficient-related features. This classification clearly distinguishes information from different physical properties, facilitating the identification of the contribution of various features to the internal structure of the medium during subsequent analysis. These features are mapped onto a multidimensional Cartesian coordinate system according to a preset physical property dimension order, with each coordinate axis corresponding to a normalized feature parameter, thus forming a multidimensional, unified feature space. In this feature space, each feature parameter not only maintains the independence of its physical properties but also intuitively expresses its interrelationships and trends in space. This construction of a multimodal feature space allows for comprehensive analysis and dynamic weighting of features from different modes and with different physical properties within a unified framework, helping to enhance the accuracy of anomaly region identification and achieve high-precision, comprehensive monitoring of the internal structure of complex media.

[0041] As a preferred embodiment, the solution of this application is implemented as follows: When performing non-destructive testing on engine blades, the normalized feature parameters can be classified according to physical properties: elastic modulus-related parameters are used to reflect the stiffness distribution of the blade material, density-related parameters are used to characterize the local compactness of the material, and attenuation coefficient-related parameters are used to reveal the energy attenuation caused by internal defects or microcracks. These classified feature parameters are integrated into a multi-dimensional Cartesian coordinate system according to a preset physical property dimension order. Each coordinate axis corresponds to a normalized feature. For example, the elastic modulus axis represents the stiffness variation in different regions of the blade, the density axis represents the local difference in material density, and the attenuation coefficient axis represents the energy loss of ultrasonic waves propagating in the material. By analyzing the numerical changes on each coordinate axis in this multi-modal feature space, abnormal areas inside the blade, such as microcracks, pores, or delamination locations, can be intuitively identified, achieving high-precision detection and visualization of complex structures.

[0042] Through the above technical solution, this application integrates normalized feature parameters into a unified multimodal feature space, realizing the systematic management and analysis of different physical attribute information. By classifying them according to physical attributes such as elastic modulus, density, and attenuation coefficient, and mapping them to a multidimensional Cartesian coordinate system in a preset dimensional order, with each coordinate axis corresponding to a feature parameter, the features of different modal signals can be intuitively represented in a unified space. This ensures the fairness of various features in the fusion analysis and avoids information distortion caused by differences in dimensions or amplitudes; it provides an accurate data foundation for subsequent dynamic weight allocation and information fusion, enhances the sensitivity and detection accuracy of anomalies in the internal structure of the measured medium, and facilitates the visualization of multimodal information.

[0043] This application further proposes a dynamic weight allocation strategy based on the sensitivity distribution of characteristic parameters to specific structural parameters of the measured medium, including: Based on the local correlation between feature parameters and specific structural parameters, the sensitivity distribution function of each feature parameter in the multimodal feature space is determined; the dynamic weight allocation strategy adjusts the weight coefficients in real time according to the sensitivity distribution function, and the allocation of weight coefficients follows the principle that the higher the sensitivity value, the greater the weight. The construction of the sensitivity distribution function is based on the gradient change characteristics of the feature parameters in the feature space.

[0044] Specifically, when dynamically weighting specific structural parameters of the tested medium, it is necessary to analyze the local correlation between each normalized feature parameter and the target structural parameter, i.e., to determine the sensitivity of each feature parameter to changes in the structural parameter in the multimodal feature space. A sensitivity distribution function can be constructed, which is based on the gradient change characteristics of the feature parameter in the feature space and reflects the response amplitude of the feature parameter as the structural parameter changes. The dynamic weighting strategy then adjusts the weight coefficients of each feature parameter in real time according to this sensitivity distribution function, giving higher weights to features with higher sensitivity, thereby enhancing the responsiveness to changes in key structural parameters during information fusion. When a feature parameter shows a significant gradient change in the feature space, its weight is automatically increased to ensure that this feature plays a dominant role in the identification and localization of abnormal regions when forming multimodal feature vectors; while features with lower sensitivity are assigned smaller weights to avoid noise or redundant information interfering with the detection results. This enables accurate characterization of the complex internal structure of the tested medium and improves the detection capability of minute defects or abnormal regions.

[0045] As a preferred embodiment, the solution of this application is implemented as follows: When detecting internal defects in composite material plates, in the multimodal signals acquired by the electromagnetic ultrasonic transducer, certain characteristic parameters, such as the energy distribution of high-frequency components, are highly sensitive to the presence of microcracks, while low-frequency components are more sensitive to changes in overall thickness. By analyzing the local correlation between these characteristic parameters and the internal structural parameters of the plate, a sensitivity distribution function can be constructed. When a significant gradient change occurs in the high-frequency energy distribution in the feature space, a higher weight is automatically assigned to that feature, while a lower weight is assigned to low-sensitivity features. In this way, when information fusion is performed to generate multimodal feature vectors, the signals of key abnormal areas such as microcracks are enhanced, thereby improving the accuracy of the detection conclusion in identifying the location and spatial boundaries of defects, and achieving efficient localization and visualization of minute anomalies inside the composite material plate.

[0046] Through the above technical solution, this application analyzes the local correlation between feature parameters and specific structural parameters of the tested medium, constructs a sensitivity distribution function, and adjusts the weight coefficients in real time, thereby improving the accuracy and reliability of multimodal signal fusion. High-sensitivity features receive a larger weight in the feature space, amplifying their detection effect on abnormal regions during information fusion, thus improving the ability to identify minute defects, cracks, or changes in material properties. The weights of low-sensitivity features are appropriately reduced, minimizing the interference of redundant information on the interpretation results. This dynamic weight allocation strategy not only optimizes the efficiency of feature fusion but also enhances the ability of detection conclusions to spatially locate and identify boundaries of internal structural anomalies in the tested medium.

[0047] This application further proposes methods for information fusion to form multimodal feature vectors, including: Information fusion is based on a weighted linear combination that multiplies the feature parameters with their corresponding weight coefficients and then sums them to generate a multimodal feature vector that represents the comprehensive characteristics of multimodal signals.

[0048] Specifically, in the process of information fusion to form a multimodal feature vector, each extracted feature parameter is first multiplied by its corresponding dynamic weight coefficient. The weight coefficient reflects the sensitivity and importance of the feature parameter to specific structural parameters in the multimodal feature space. Through a weighted linear combination, all weighted feature parameters are accumulated to generate a feature vector that comprehensively characterizes the multimodal signal properties of the measured medium. This multimodal feature vector not only retains the key physical information of each modal signal but also highlights sensitive features and suppresses noise and interference from low-sensitivity features through weight adjustment, thereby improving the accuracy and stability of subsequent interpretation and analysis. This feature vector can accurately represent the overall characteristics of the internal structural state of the medium in a multidimensional feature space, providing a reliable data foundation for anomaly region identification and spatial boundary determination.

[0049] As a preferred embodiment, the solution of this application is implemented as follows: For example, when detecting internal defects in a composite material plate, signals from different propagation modes obtained from an electromagnetic ultrasonic transducer, after feature extraction and weight allocation, yield a set of feature parameters reflecting the elastic modulus, density, and attenuation coefficient, along with their corresponding weights. During information fusion, these feature parameters are multiplied by their respective weight coefficients and then summed to generate a multimodal feature vector. This feature vector integrates key physical information from each modal signal, such as changes in spectral energy distribution caused by local cracks in the plate, anomalies in the signal envelope peak, and instantaneous phase shifts, thereby comprehensively and accurately reflecting the internal structural state of the composite material plate.

[0050] Through the above technical solution, this application generates multimodal feature vectors by weighted linear combination of each feature parameter and its corresponding weight coefficient, which effectively integrates key information in ultrasonic signals of different propagation modes, improving the completeness and accuracy of signal characterization. This method not only highlights the features most sensitive to changes in the internal structure of the measured medium, improving the sensitivity of anomaly detection, but also suppresses the influence of noise and redundant information, thereby enhancing the reliability and stability of the detection results.

[0051] This application further proposes a comprehensive interpretation and analysis of multimodal feature vectors to generate detection conclusions, including: The multimodal feature vectors are input into the judgment rule base, which contains anomaly discrimination threshold ranges and boundary recognition logic constructed based on historical detection data. The comprehensive interpretation and analysis performs multi-level comparison operations; By comparing the Euclidean distance between the feature vector and the reference vector in the normal state, abnormal regions can be identified. The spatial boundary of the abnormal region is determined by the gradient direction change of the feature vector in the feature space.

[0052] Specifically, in the process of comprehensively analyzing multimodal feature vectors to generate detection conclusions, each multimodal feature vector is fed into an interpretation rule base. This rule base pre-contains anomaly discrimination threshold ranges, noise estimation parameters, and boundary recognition logic (such as spatial connectivity and minimum size constraints) obtained from a large amount of historical detection data. Subsequently, a multi-level comparison process is executed: in the coarse screening stage, the Euclidean distance between the current feature vector and the normal state reference vector is calculated and compared with the threshold range in the rule base (adaptive thresholds can be used to take into account background noise); in the fine kernel stage, multi-model cross-validation is performed on candidate samples (such as independent criteria for different modalities, comparison with historical similar sample databases, and time series consistency tests), and the comprehensive confidence level is calculated. The degree score (combining distance excess, modal consistency, and historical priors) is used in the boundary extraction stage. In the multimodal feature space, the feature gradient vector field is calculated with the candidate point as the center (the rate of change of each coordinate direction is obtained by neighborhood difference or local derivative estimation). Based on the gradient direction and magnitude, the contour line of the anomaly is determined along the contour line or gradient ascent path of the gradient change. The boundary in the feature space is transformed back to the physical coordinate system of the measured medium through a pre-established mapping relationship (such as spatial interpolation / back projection or array geometry-based localization algorithm) to obtain the coordinate set of the anomaly region and the boundary polygon. The judgment rule base performs post-processing on the boundary (morphological filtering, minimum size filtering, and neighbor candidate merging) and adds confidence labels and feature descriptions to the output.

[0053] As a preferred embodiment, the solution of this application is implemented as follows: For example, when performing electromagnetic ultrasonic testing on a metal plate, the acquired multimodal feature vectors are input into an interpretation rule base. This rule base stores the normal state feature vectors and their abnormal thresholds for metal plates of different thicknesses and materials from historical testing. Through multi-level comparison operations, the Euclidean distance between the current feature vector and the normal reference vector is first calculated. If the distance exceeds the threshold, an abnormal region is initially determined to exist. Subsequently, the spatial boundary of the abnormal region is determined by combining the gradient direction change of the feature vector in the multimodal feature space and along the region with a high gradient change rate. For example, if during the testing process it is found that the amplitude and frequency characteristics of the feature vector in a certain region deviate significantly from those of a normal plate, and the gradient change is concentrated in the upper left corner of the plate center, then the coordinates and boundary contour of that region are finally output, indicating that there may be internal cracks or pores.

[0054] Through the above technical solution, this application inputs the feature vector into the judgment rule base, and performs multi-level comparison operations based on the anomaly discrimination threshold range and boundary recognition logic constructed from historical detection data to distinguish between normal and abnormal states. When the Euclidean distance between the feature vector and the normal reference vector exceeds a preset threshold, a potential abnormal region is identified; combined with the gradient direction change of the feature vector in the feature space, the spatial boundary of the abnormal region is determined.

[0055] This application further proposes that when outputting visual representation information corresponding to the coordinate data and spatial boundary data of the abnormal region based on the detection conclusion, the following should be included: Map the coordinate data of abnormal areas and spatial boundary data in the detection results to the three-dimensional geometric model of the measured medium; The coordinate data of the abnormal region is based on the overlay of highlighted color blocks on the surface of the 3D geometric model, and the spatial boundary data is based on the contour line annotation; a visual representation information of the text annotation layer containing the location, size and feature parameter association information of the abnormal region is generated.

[0056] Specifically, in the process of outputting visual representation information based on the detection results, the coordinate data and spatial boundary data of the identified abnormal areas are precisely mapped onto the three-dimensional geometric model of the tested medium to ensure accurate spatial representation of the detection results. Abnormal areas are superimposed on the surface of the three-dimensional model using highlighted color blocks to visually display the specific location of defects or anomalies. Simultaneously, contour lines are used to indicate spatial boundaries, clearly outlining the shape and extent of the abnormal areas. The generated visual representation information also includes a text annotation layer, recording the location, size, and correspondence with feature parameters of each abnormal area, enabling users to intuitively understand the specific physical characteristics and severity of each anomaly.

[0057] As a preferred embodiment, the solution of this application is implemented as follows: For example, when inspecting an industrial composite material plate for internal defects, the coordinate data and spatial boundary data of the abnormal area obtained through multimodal ultrasonic signal analysis are mapped onto the three-dimensional geometric model of the composite material plate. On the three-dimensional model, the identified defect areas are clearly displayed through highlighted color blocks, allowing users to intuitively determine the specific location of the defects. Simultaneously, contour lines are used to depict the spatial boundaries of the defects, facilitating observation of their shape and extent. To further enhance the information visualization effect, a text annotation layer is generated on the model surface, recording in detail the spatial coordinates, size, and correlation information with characteristic parameters (such as changes in elastic modulus, density differences, and abnormal attenuation coefficients) of each defect area. In this way, operators can not only quickly locate and assess internal defects in the material but also conduct a comprehensive analysis of the nature of the defects and potential risks.

[0058] Through the above technical solution, this application maps the coordinate data and spatial boundary data of the abnormal areas obtained from the detection results to the three-dimensional geometric model of the tested medium, presenting the abnormal distribution within the medium. Highlighted color blocks are used to cover the abnormal areas, making them clearly visible on the surface of the three-dimensional model, while contour lines accurately depict the spatial boundaries of the abnormal areas, facilitating observation of their shape and extent. The generated text annotation layer records the specific location, size, and correlation information with feature parameters of each abnormal area, such as changes in elastic modulus, density differences, and abnormal attenuation coefficients, improving the intuitiveness and comprehensibility of the detection results.

[0059] Based on the other preferred method described above, see [link / reference]. Figure 2 As shown, this embodiment provides a multimodal signal fusion detection system for an electromagnetic ultrasonic transducer, used to apply the above-mentioned multimodal signal fusion detection method for an electromagnetic ultrasonic transducer, including: The acquisition unit is configured to excite and receive ultrasonic signals with multiple propagation modes in the measured medium based on an electromagnetic ultrasonic transducer, acquire multimodal raw signal data containing different physical characteristics, and perform time synchronization processing and data preprocessing on the multimodal raw signals. The processing unit is configured to extract feature parameters of the internal state of the measured medium based on the multimodal raw signal data after data preprocessing; and to normalize the feature parameters according to physical properties and integrate them into a unified multimodal feature space. The generation unit is configured to set a dynamic weight allocation strategy based on the sensitivity distribution of feature parameters to specific structural parameters of the measured medium in the multimodal feature space, and perform information fusion to form a multimodal feature vector; perform comprehensive interpretation and analysis on the multimodal feature vector to generate a detection conclusion; the detection conclusion is the coordinate data of the abnormal area of ​​the internal structure of the measured medium and the corresponding spatial boundary data. The output unit is configured to output visual representation information corresponding to the coordinate data of the abnormal area and the spatial boundary data based on the detection conclusion.

[0060] In summary, using an electromagnetic ultrasonic transducer to acquire ultrasonic signals with multiple propagation modes avoids the dependence on coupling agents found in traditional piezoelectric transducers. This allows for stable operation under complex conditions such as high temperatures, strong corrosion, or rough surfaces, improving the environmental adaptability and reliability of the detection. By performing time-synchronized processing and data preprocessing on the multimodal raw signals, and extracting feature parameters from three dimensions—time domain, frequency domain, and instantaneous characteristics—a comprehensive characterization of the internal physical state of the measured medium can be achieved, avoiding the information loss problem caused by noise interference in single-mode signals. Normalizing the feature parameters of different physical properties and integrating them into a unified multimodal feature space ensures the comparability between features of different dimensions, facilitating subsequent weight allocation and fusion calculations, thereby improving the scientific rigor and accuracy of the detection data processing. A dynamic weight allocation strategy based on the sensitivity distribution of feature parameters to structural parameters allows the weight coefficients to be adjusted in real time according to detection conditions and local characteristics, ensuring high sensitivity to abnormal areas and improving the accuracy of defect identification and location. Information fusion is achieved through weighted linear combination to obtain a multimodal feature vector with comprehensive characteristics. This vector is then combined with a rule base built on historical data to perform multi-level comparisons, reducing the false alarm rate and improving the accuracy of anomaly region identification and spatial boundary determination. The detection results are visualized using a 3D geometric model, combined with highlighted color blocks, contour lines, and text annotation layers. This visualization displays the location, extent, and relevant feature parameters of the anomaly regions, providing a clear basis for subsequent structural evaluation and decision-making.

[0061] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for multimodal signal fusion detection of an electromagnetic ultrasonic transducer, characterized in that, include: Based on the electromagnetic ultrasonic transducer, ultrasonic signals with multiple propagation modes are excited and received in the medium under test to obtain multimodal raw signal data containing different physical properties. The original multimodal signal is subjected to time synchronization processing and data preprocessing. Based on the multimodal raw signal data after the data preprocessing operation, feature parameters of the internal state of the measured medium are extracted; and the feature parameters are normalized according to physical properties and integrated into a unified multimodal feature space. In the multimodal feature space, a dynamic weight allocation strategy is set according to the sensitivity distribution of the feature parameters to specific structural parameters of the tested medium, and information fusion is performed to form a multimodal feature vector; the multimodal feature vector is comprehensively interpreted and analyzed to generate a detection conclusion; the detection conclusion is the coordinate data of the abnormal area of ​​the internal structure of the tested medium and the corresponding spatial boundary data. Based on the detection results, output visual representation information corresponding to the coordinate data and spatial boundary data of the abnormal region.

2. The multi-mode signal fusion detection method for an electromagnetic ultrasonic transducer according to claim 1, characterized in that, When performing time synchronization processing and data preprocessing operations on the multimodal raw signal, the following are included: The time synchronization process involves aligning the sampling points of each modal ultrasound signal in phase based on a reference clock signal; the data preprocessing operations include performing bandpass filtering, DC component removal, and amplitude normalization. The time synchronization process and data preprocessing operation are executed sequentially, and the data preprocessing operation is limited to being performed after the time synchronization process is completed.

3. The multi-mode signal fusion detection method for an electromagnetic ultrasonic transducer according to claim 2, characterized in that, When extracting feature parameters of the internal state of the measured medium from the multimodal raw signal data after the data preprocessing operation, the following are included: Based on ultrasonic signals of different propagation modes, time-domain characteristic parameters, frequency-domain characteristic parameters, and instantaneous characteristic parameters directly related to the internal structure of the medium are extracted respectively. The time-domain characteristic parameters cover the peak amplitude and zero-crossing rate of the signal envelope, the frequency-domain characteristic parameters cover the energy distribution and spectral centroid of the dominant frequency component, and the instantaneous characteristic parameters cover the instantaneous phase change rate and the analytical signal envelope derived from the Hilbert transform.

4. The multimodal signal fusion detection method for an electromagnetic ultrasonic transducer according to claim 3, characterized in that, When normalizing the feature parameters according to physical properties, the following steps are included: Based on the physical dimensions and dynamic range of the characteristic parameters, dimensionless normalization is performed using a linear transformation based on the minimum-maximum scale.

5. The multi-mode signal fusion detection method for an electromagnetic ultrasonic transducer according to claim 4, characterized in that, When integrating into a unified multimodal feature space, it includes: The normalized feature parameters are divided into elastic modulus-related parameters, density-related parameters, and attenuation coefficient-related parameters according to the physical property category; the divided feature parameters are integrated into the multimodal feature space composed of a multidimensional Cartesian coordinate system according to the preset physical property dimension order. Each coordinate axis of the multimodal feature space corresponds to a normalized feature parameter.

6. The multi-mode signal fusion detection method for an electromagnetic ultrasonic transducer according to claim 5, characterized in that, When setting a dynamic weight allocation strategy based on the sensitivity distribution of the characteristic parameters to specific structural parameters of the measured medium, the strategy includes: Based on the local correlation between the feature parameters and specific structural parameters, the sensitivity distribution function of each feature parameter in the multimodal feature space is determined; the dynamic weight allocation strategy adjusts the weight coefficients in real time according to the sensitivity distribution function, and the allocation of the weight coefficients follows the principle that the higher the sensitivity value, the greater the weight. The sensitivity distribution function is constructed based on the gradient change characteristics of the feature parameters in the feature space.

7. The multimodal signal fusion detection method for an electromagnetic ultrasonic transducer according to claim 6, characterized in that, When performing information fusion to form multimodal feature vectors, the following are included: The information fusion is based on a weighted linear combination, which multiplies the feature parameters with their corresponding weight coefficients and then sums them to generate a multimodal feature vector that represents the comprehensive characteristics of the multimodal signal.

8. The multi-mode signal fusion detection method for an electromagnetic ultrasonic transducer according to claim 7, characterized in that, When performing comprehensive interpretation and analysis on the multimodal feature vectors to generate detection conclusions, the following are included: The multimodal feature vector is input into the judgment rule base, which includes anomaly discrimination threshold range and boundary recognition logic constructed based on historical detection data; The comprehensive interpretation and analysis performs multi-level comparison operations; By comparing the Euclidean distance between the feature vector and the reference vector in the normal state, abnormal regions can be identified. The spatial boundary of the abnormal region is determined by the gradient direction change of the feature vector in the feature space.

9. The multimodal signal fusion detection method for an electromagnetic ultrasonic transducer according to claim 8, characterized in that, When outputting visual representation information corresponding to the coordinate data and spatial boundary data of the abnormal region based on the detection conclusion, it includes: Map the coordinate data of the abnormal area and the spatial boundary data in the detection conclusion to the three-dimensional geometric model of the medium under test; The coordinate data of the abnormal region is based on the superposition of bright color blocks on the surface of the three-dimensional geometric model, and the spatial boundary data is based on the outline annotation; the visualization representation information of the text annotation layer containing the location, size and feature parameter association information of the abnormal region is generated.

10. A multi-mode signal fusion detection system for an electromagnetic ultrasonic transducer, used in applying the multi-mode signal fusion detection method for an electromagnetic ultrasonic transducer as described in any one of claims 1-9, characterized in that, include: The acquisition unit is configured to excite and receive ultrasonic signals with multiple propagation modes in the medium under test based on an electromagnetic ultrasonic transducer, acquire multimodal raw signal data containing different physical characteristics, and perform time synchronization processing and data preprocessing operations on the multimodal raw signals. The processing unit is configured to extract feature parameters of the internal state of the measured medium based on the multimodal raw signal data after the data preprocessing operation; and to normalize the feature parameters according to physical properties and integrate them into a unified multimodal feature space. The generation unit is configured to, in the multimodal feature space, set a dynamic weight allocation strategy based on the sensitivity distribution of the feature parameters to specific structural parameters of the tested medium, and perform information fusion to form a multimodal feature vector; perform comprehensive interpretation and analysis on the multimodal feature vector to generate a detection conclusion; the detection conclusion is the coordinate data of the abnormal area of ​​the internal structure of the tested medium and the corresponding spatial boundary data; The output unit is configured to output visual representation information corresponding to the coordinate data of the abnormal region and the spatial boundary data based on the detection conclusion.

Images