A method for magnetic resonance signal three-dimensional spectrum feature recognition and analysis

Abstract

The present application relates to the technical field of nuclear magnetic resonance analysis, and particularly relates to a magnetic resonance signal three-dimensional spectrum feature recognition and analysis method.The method comprises the following steps: acquiring a multi-channel time-domain magnetic field signal sequence through a three-dimensional sensing array; performing frame windowing processing and fast Fourier transform on the sequence to construct a complex spectrum matrix; decoupling the matrix in the complex domain by using a statistical independence criterion, eliminating the arrangement ambiguity introduced by the frequency domain transform through sub-band division, component extraction and spatial response consistency judgment, and restoring to obtain independent spectrum component signals corresponding to different radiation source regions; calculating the physical center coordinates of each radiation source region in combination with the array spatial coordinates, and then constructing a three-dimensional spectrum feature fingerprint of space-frequency domain fusion; and finally recognizing the state feature of the to-be-tested subject according to the matching deviation degree of the fingerprint and the standard feature library.The present application solves the problems of multi-source magnetic signal space coupling and frequency domain stripping, and improves the accuracy of weak feature recognition.

Description

Technical Field

[0001] This invention relates to the field of nuclear magnetic resonance analysis technology, and in particular to a method for identifying and analyzing three-dimensional spectral features of magnetic resonance signals. Background Technology

[0002] Magnetic resonance (MR) analysis, as a non-invasive detection method, plays a crucial role in fields such as industrial non-destructive testing, biomedical monitoring, and materials science. In practical applications, the object under test often contains multiple simultaneously operating or excited electromagnetic radiation sources. These source signals are highly coupled spatially, forming a complex aliased magnetic field at the sensor array. Traditional analysis methods often focus on the direct extraction of time-domain waveforms or simple frequency-domain amplitude analysis, making it difficult to effectively separate weak feature components from a high-noise background. Especially when dealing with multi-layered complex structures or biological tissues, the magnetic signals generated by each radiation source undergo phase shifts and energy attenuation during propagation, resulting in the original signals captured by the sensor exhibiting strong nonlinearity and spatial overlap characteristics, significantly limiting the accuracy of identifying the state of specific internal regions.

[0003] Existing spatial decoupling schemes often face severe permutation ambiguity problems when processing broadband magnetic field signals. Since the sub-band signals after frequency domain transformation are typically treated as independent random variables, the signal components extracted at different frequency points often fail to maintain consistency in physical index when performing algorithms such as blind source separation. This random permutation leads to confusion in the frequency axis of the reconstructed signal, preventing the formation of a complete spectral fingerprint. Furthermore, existing technologies often lack deep integration of spatial location information and spectral characteristics, making it difficult to achieve quantitative analysis of the physical offset of radiation sources in complex three-dimensional space. Summary of the Invention

[0004] To overcome the above deficiencies, this invention provides a method for three-dimensional spectral feature identification and analysis of magnetic resonance signals, aiming to achieve accurate separation of multi-source magnetic signals and construct a high-dimensional feature descriptor with spatial positioning capabilities to achieve accurate analysis of magnetic resonance signals.

[0005] This invention provides the following technical solution: a method for three-dimensional spectral feature identification and analysis of magnetic resonance signals, comprising:

[0006] A multi-channel time-domain magnetic field signal sequence is acquired by a three-dimensional sensing array pre-set around the subject under test;

[0007] Fourier transform is performed on the time-domain magnetic field signals of each channel to obtain a complex spectrum matrix containing amplitude and phase information;

[0008] In the complex domain, the statistical independence criterion is used to spatially decouple the complex spectrum matrix and extract the independent spectrum component signals corresponding to different radiation source regions.

[0009] By combining the spatial coordinates of the three-dimensional sensing array, each independent spectral component signal is mapped to the three-dimensional spatial coordinate system to construct a three-dimensional spectral feature fingerprint of each radiation source region.

[0010] The three-dimensional spectral fingerprint is matched with a preset standard feature library, and the state characteristics of the subject under test are identified based on the matching deviation.

[0011] Preferably, the step of acquiring a multi-channel time-domain magnetic field signal sequence includes:

[0012] Based on the preset geometric topology, multiple magnetic sensors are deployed on the outer periphery of the subject under test to construct a three-dimensional sensing array.

[0013] The magnetic signals captured by each magnetic sensor in the three-dimensional sensing array are synchronously triggered and acquired through a multi-channel sampling circuit to obtain multiple raw time-domain sequences.

[0014] By using a noise reduction algorithm, background interference is removed from the multiple original time-domain sequences using a reference environmental signal, resulting in the time-domain magnetic field signal sequence corresponding to each channel.

[0015] Preferably, the step of performing Fourier transform on the time-domain magnetic field signals of each channel includes:

[0016] The time-domain magnetic field signal sequences of each channel are processed by frame division to generate multiple time-domain signal frames with a preset overlap rate;

[0017] A preset window function is used to perform windowing smoothing on each of the time-domain signal frames to suppress spectral leakage characteristics;

[0018] A fast Fourier transform is performed on each of the windowed time-domain signal frames to convert each signal frame from the time domain to the frequency domain, and the corresponding amplitude and phase information are extracted to construct a complex spectrum matrix.

[0019] Preferably, the step of spatially decoupling the complex spectrum matrix includes:

[0020] According to the preset frequency band step size, the complex spectrum matrix is ​​divided into multiple independent sub-band matrices to be processed;

[0021] Solve the subband matrix to be processed to obtain the decoupling matrix corresponding to each subband, and extract the candidate independent components under each subband based on the decoupling matrix;

[0022] Determine the spatial response consistency of the candidate independent components, and perform frequency alignment and sorting normalization processing on the candidate independent components under different sub-bands to eliminate the ambiguity of arrangement introduced by frequency domain transformation.

[0023] The signal is reconstructed based on the aligned components of each sub-band, and the independent spectral component signals corresponding to different target radiation source regions are restored.

[0024] Preferably, the solution to the subband matrix to be processed is achieved by an independent component analysis algorithm, the steps of which include:

[0025] Centering and whitening preprocessing are performed on each of the sub-band matrices to be processed in order to eliminate the correlation between the signals of each channel and unify the variance;

[0026] The objective function is to maximize the non-Gaussianity of the feature signal, and the preprocessed matrix is ​​iteratively optimized in the complex domain.

[0027] When the objective function reaches the convergence condition, the complex-valued separation matrix corresponding to each sub-band is output as the decoupling matrix.

[0028] Preferably, the step of determining the spatial response consistency of the candidate independent components includes:

[0029] Obtain the spatial guiding vector corresponding to each candidate independent component under each sub-band;

[0030] Calculate the correlation coefficient of the spatial guidance vector between adjacent sub-bands or between preset reference sub-bands;

[0031] Based on the magnitude of the correlation coefficient, the candidate independent components are clustered and sorted, and independent components belonging to the same radiation source region are merged into the same index sequence to complete frequency alignment.

[0032] Preferably, the steps for constructing the three-dimensional spectral feature fingerprint of each radiation source region include:

[0033] Extract the characteristic parameters of each independent spectral component signal in the corresponding sub-band, the characteristic parameters including center frequency intensity, frequency band energy distribution and spectral entropy value;

[0034] Based on the response weights of each channel in the three-dimensional sensing array to specific independent components, the physical center coordinates of each radiation source region in the three-dimensional coordinate system are calculated using a spatial inversion positioning algorithm.

[0035] The physical center coordinates are associated with the corresponding feature parameters in a multi-dimensional attribute relationship to generate a spatial-frequency domain fusion descriptor for each radiation source region, which serves as the three-dimensional spectral feature fingerprint.

[0036] Preferably, the step of identifying the state characteristics of the subject under test based on the matching deviation includes:

[0037] Input the three-dimensional spectral feature fingerprints corresponding to each radiation source region into a preset classification and discrimination model;

[0038] Calculate the cosine similarity between the three-dimensional spectral feature fingerprint and the corresponding benchmark fingerprint in the standard feature library to obtain the quantized matching deviation.

[0039] Determine whether the matching deviation exceeds a preset abnormal threshold range;

[0040] If the deviation exceeds the limit, the physical state offset category of the corresponding radiation source region is identified based on the distribution characteristics of the matching deviation, and the corresponding state feature analysis report is output.

[0041] Secondly, the present invention provides the following technical solution: a three-dimensional spectral feature recognition and analysis system for magnetic resonance signals, used to perform the three-dimensional spectral feature recognition and analysis method for magnetic resonance signals as described above, comprising:

[0042] The magnetic field signal acquisition module is used to acquire multi-channel time-domain magnetic field signal sequences through a three-dimensional sensing array preset around the subject under test;

[0043] The magnetic field signal transformation module is used to perform Fourier transform on the time-domain magnetic field signals of each channel to obtain a complex spectrum matrix containing amplitude and phase information;

[0044] The magnetic field spectrum decoupling module is used to spatially decouple the complex spectrum matrix in the complex domain using statistical independence criteria, and to extract independent spectral component signals corresponding to different radiation source regions.

[0045] The spectrum fingerprint construction module is used to combine the spatial coordinates of the three-dimensional sensing array to map each independent spectrum component signal to the three-dimensional spatial coordinate system and construct the three-dimensional spectrum feature fingerprint of each radiation source region.

[0046] The spectrum deviation identification module is used to match the three-dimensional spectrum feature fingerprint with a preset standard feature library, and identify the state features of the subject under test based on the degree of matching deviation.

[0047] The present invention has the following beneficial effects:

[0048] 1. This invention effectively solves the problem of permutation ambiguity in frequency domain decoupling by using independent component analysis and spatial response consistency alignment in the complex domain. It can extract independent and physically continuous spectral components from strong background noise, greatly improving the recognition accuracy of weak electromagnetic features in complex environments.

[0049] 2. This invention deeply couples the three-dimensional coordinates obtained from spatial inversion positioning with frequency domain feature parameters, forming a feature fingerprint that not only has frequency specificity but also precise spatial orientation, realizing a quantitative and visual assessment of the physical state shift of the radiation source region, and providing a reliable basis for accurate analysis. Attached Figure Description

[0050] Figure 1 A flowchart illustrating a method for identifying and analyzing three-dimensional spectral features of magnetic resonance signals provided in an embodiment of the present invention;

[0051] Figure 2 A structural diagram of a three-dimensional spectral feature recognition and analysis system for magnetic resonance signals provided in an embodiment of the present invention;

[0052] Figure 3 This is a comparative diagram showing the spatial response consistency alignment before and after, provided in an embodiment of the present invention; wherein, Figure 3 (a) shows the state where there is ambiguity in the arrangement of each subband after decoupling. Figure 3 (b) shows the state after frequency alignment using spatial guide vectors. Detailed Implementation

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example 1

[0055] In the first embodiment of the present invention, a method for three-dimensional spectral feature identification and analysis of magnetic resonance signals is provided, such as... Figure 1 As shown, it includes the following steps:

[0056] S1. Acquire a multi-channel time-domain magnetic field signal sequence through a three-dimensional sensing array pre-set around the subject under test;

[0057] Preferably, the step of acquiring a multi-channel time-domain magnetic field signal sequence includes:

[0058] Based on the preset geometric topology, multiple magnetic sensors are deployed on the outer periphery of the subject under test to construct a three-dimensional sensing array.

[0059] The magnetic signals captured by each magnetic sensor in the three-dimensional sensing array are synchronously triggered and acquired through a multi-channel sampling circuit to obtain multiple raw time-domain sequences.

[0060] By using a noise reduction algorithm, background interference is removed from the multiple original time-domain sequences using a reference environmental signal, resulting in the time-domain magnetic field signal sequence corresponding to each channel.

[0061] Specifically, the geometric topology of the sensors is pre-defined based on the geometric shape of the subject under test. In this embodiment, a cylindrical array layout is used, with several magnetic sensors evenly distributed within the space surrounding the subject under test. The magnetic sensors are tunneling magnetoresistive sensors with high magnetic field resolution. The three-dimensional coordinates of each magnetic sensor in space are... Pre-stored in the data processing unit, where This is the index number of the sensor.

[0062] A multi-channel synchronous sampling circuit is used, with each sensing channel sharing the same clock trigger signal. When the system initiates the acquisition command, the analog magnetic field strength signals captured by each magnetic sensor are amplified by a preamplifier circuit and then enter a multiplexer. Each channel operates at a preset sampling frequency. Perform equal-interval sampling. The resulting multi-path raw time-domain sequence is represented as follows: ,in This is the sampling time point. Synchronous triggering ensures that the phase deviation of each channel sequence on the time axis is less than a preset allowable range.

[0063] Before performing excitation acquisition or in a specific electromagnetic shielding environment, environmental reference signals should be acquired in advance. The noise reduction algorithm processes multiple original time-domain sequences using the following formula:

[0064] ;

[0065] in, This is the time-domain magnetic field signal sequence for each channel obtained after removing interference. For recording the environmental reference background signal, This is a gain compensation coefficient that is adjusted in real time according to environmental fluctuations. The power frequency interference component is extracted using an adaptive filter. In this process, the algorithm first cancels the static geomagnetic field component through differential operations, and then uses a notch filter to remove periodic noise at specific frequencies. After background interference removal, the weak alternating magnetic field signal related to the state of the subject under test in each channel can be extracted.

[0066] S2. Perform Fourier transform on the time-domain magnetic field signals of each channel to obtain a complex spectrum matrix containing amplitude and phase information;

[0067] Preferably, the step of performing Fourier transform on the time-domain magnetic field signals of each channel includes:

[0068] The time-domain magnetic field signal sequences of each channel are processed by frame division to generate multiple time-domain signal frames with a preset overlap rate;

[0069] A preset window function is used to perform windowing smoothing on each of the time-domain signal frames to suppress spectral leakage characteristics;

[0070] A fast Fourier transform is performed on each of the windowed time-domain signal frames to convert each signal frame from the time domain to the frequency domain, and the corresponding amplitude and phase information are extracted to construct a complex spectrum matrix.

[0071] Specifically, the time-domain magnetic field signal sequences of each channel after interference removal The signal is divided into several consecutive time segments, known as time-domain signal frames. To ensure the continuity of signal characteristics under non-stationary conditions, a preset overlap rate is set between adjacent signal frames. If the length of a single frame is If there are 1 sampling point, then the starting point offset between two adjacent frames is 1 / 2. By using frame segmentation, the continuous magnetic field signal stream is transformed into a discrete frame sequence. ,in For frame index, For the intra-frame sampling point index, satisfying .

[0072] For each time-domain signal frame Apply a preset window function This is to reduce spectral leakage caused by finite-length truncation. The preferred window function here is the Hamming window, whose mathematical expression is as follows:

[0073] ;

[0074] By applying a windowing process, the data at both ends of the signal frame is smoothly attenuated to zero, thereby reducing the edge discontinuities caused by the periodicity assumption when performing the Discrete Fourier Transform and improving the amplitude accuracy of the spectrum analysis.

[0075] For each time-domain signal frame after windowing Perform a Fast Fourier Transform to obtain the corresponding frequency domain response. :

[0076] ;

[0077] in, For frequency index, For sensor index number, For frame index, For intra-frame sampling point index, It is a natural constant. The imaginary unit, This represents the number of sampling points per frame. For each windowed time-domain signal frame, the following steps are performed. The frequency domain sequence obtained after point fast Fourier transform Each frequency point data point exists in complex form and can be represented as:

[0078] ;

[0079] in, Let be the real part of the transformation result. The imaginary part of the transformation result. It is the imaginary unit.

[0080] Extract the amplitude at each frequency point from the transformation result. With phase And thus construct a complex spectrum matrix. Each element in this matrix All are represented in complex form:

[0081] ;

[0082] The resulting complex spectrum matrix It includes the full-band electromagnetic distribution characteristics of each channel at different time frames.

[0083] S3. In the complex domain, the complex spectrum matrix is ​​spatially decoupled using the statistical independence criterion to extract the independent spectrum component signals corresponding to different radiation source regions.

[0084] Preferably, the step of spatially decoupling the complex spectrum matrix includes:

[0085] According to the preset frequency band step size, the complex spectrum matrix is ​​divided into multiple independent sub-band matrices to be processed;

[0086] Solve the subband matrix to be processed to obtain the decoupling matrix corresponding to each subband, and extract the candidate independent components under each subband based on the decoupling matrix;

[0087] Determine the spatial response consistency of the candidate independent components, and perform frequency alignment and sorting normalization processing on the candidate independent components under different sub-bands to eliminate the ambiguity of arrangement introduced by frequency domain transformation.

[0088] The signal is reconstructed based on the aligned components of each sub-band, and the independent spectral component signals corresponding to different target radiation source regions are restored.

[0089] Specifically, based on a preset frequency band step size The complex spectrum matrix will be constructed The frequency range is segmented along the frequency axis. Each frequency range corresponds to a sub-band matrix to be processed. ,in The center frequency of this sub-band is defined. By dividing the signal into sub-bands, the broadband magnetic field signal is transformed into multiple narrowband signal processing tasks. The aim is to reduce the computational dimensionality of spatial decoupling by utilizing the linear instantaneous mixing model satisfied by the narrowband signal in the frequency domain.

[0090] Extract the matrix for each subband to be processed. The corresponding decoupling matrix is ​​obtained by optimizing the solution using the statistical independence criterion. By performing a linear transformation on the original observation vector within the sub-band using the decoupling matrix, mutually independent candidate components at the sub-band frequency are extracted. The calculation process is as follows:

[0091] ;

[0092] Extracted It includes the contributions of multiple potential radiation sources in this specific subband.

[0093] Since the decoupling process for each sub-band is performed independently, the order of components extracted from different sub-bands is random, leading to permutation ambiguity. Therefore, the system extracts the spatial pointing features, i.e., the spatial response vectors, corresponding to each candidate independent component. The correlation coefficients of the spatial response vectors of components between adjacent sub-bands are calculated, and based on these correlation coefficients... The components of different subbands are rearranged and aligned to ensure that cross-band components with the same index all point to the same physical radiation source. Subsequently, the aligned components are subjected to amplitude normalization.

[0094] Based on the frequency-aligned index sequence, the candidate independent components scattered across each sub-band are spectrally stitched together. Through inverse mapping, the narrowband components of each sub-band are reconstructed into independent spectral component signals within the complete observation frequency band.

[0095] Preferably, the solution to the subband matrix to be processed is achieved by an independent component analysis algorithm, the steps of which include:

[0096] Centering and whitening preprocessing are performed on each of the sub-band matrices to be processed in order to eliminate the correlation between the signals of each channel and unify the variance;

[0097] The objective function is to maximize the non-Gaussianity of the feature signal, and the preprocessed matrix is ​​iteratively optimized in the complex domain.

[0098] When the objective function reaches the convergence condition, the complex-valued separation matrix corresponding to each sub-band is output as the decoupling matrix.

[0099] Specifically, for each sub-band matrix to be processed First, the signal is centered by subtracting the mean of each channel signal to make it a zero-mean vector. Then, a whitening transformation is performed on the centered data. The covariance matrix of the signal is obtained using eigenvalue decomposition or singular value decomposition methods, and a whitening transformation matrix is ​​constructed. Data after whitening satisfy:

[0100] ;

[0101] This process eliminates the second-order correlation between the channels of the observed signal and unifies the variance of each component to a unit variance, thereby simplifying the search space for subsequent decoupling matrix optimization.

[0102] Preprocessed data Based on this, an objective function is established with the criterion of maximizing non-Gaussianity. Since the magnetic field signals of each channel originate from different radiation sources, according to the central limit theorem, the Gaussianity of the mixed signal is higher than that of any single independent source signal. Therefore, by finding the mapping direction that maximizes the non-Gaussianity of the output sequence, the source signals can be decoupled. In complex domain operations, the approximate value of negative entropy is used as the quantification criterion for non-Gaussianity, and the following formula is used for iterative optimization:

[0103] ;

[0104] in, Let be the separation vector to be solved. Its conjugate transpose. For mathematical expectation, For the preset nonlinear comparison function, The variables are standardized Gaussian complex variables. Gradient descent is used to continuously refine the values. The pointer continues until the objective function reaches its maximum.

[0105] After each iteration, the change in the separation vector between two adjacent iterations is calculated. The algorithm is considered to have reached convergence when the change is less than a preset allowable bias, or when the rate of change of the objective function tends to stabilize. At this point, the separation vectors corresponding to each independent component are orthogonalized and assembled, and the complex-valued separation matrix corresponding to that sub-band is output. The resulting decoupling matrix is ​​the linear transformation operator that maps from the mixed observation space to the independent feature space at that sub-band frequency.

[0106] Preferably, the step of determining the spatial response consistency of the candidate independent components includes:

[0107] Obtain the spatial guiding vector corresponding to each candidate independent component under each sub-band;

[0108] Calculate the correlation coefficient of the spatial guidance vector between adjacent sub-bands or between preset reference sub-bands;

[0109] Based on the magnitude of the correlation coefficient, the candidate independent components are clustered and sorted, and independent components belonging to the same radiation source region are merged into the same index sequence to complete frequency alignment.

[0110] Specifically, after completing the initial decoupling of each subband, the spatial guiding vector corresponding to each candidate independent component under each subband is extracted. ,in For sub-band index, This is the index of the independent components within this sub-band. The spatial guiding vector is calculated by using the decoupling matrix. The inverse matrix is ​​obtained, and each component represents the complex gain weight of a specific independent radiation source on each channel of the three-dimensional sensing array.

[0111] To measure the attribution relationship of components between different subbands, adjacent subbands are calculated. and The correlation coefficients of the spatial guided vectors of each component The normalized inner product is used as the correlation criterion, and its mathematical expression is:

[0112] ;

[0113] in, Indicates the first Sub-band The component and the first Sub-band The spatial similarity between the components. The resulting correlation coefficient ranges from [0, 1], and the closer the value is to 1, the higher the probability that the two components come from the same physical radiation source region.

[0114] Based on the correlation coefficient matrix, a clustering algorithm is used to reorder the candidate independent components under each sub-band. Using a preset reference sub-band as a benchmark, the component pairs with the highest correlation coefficients in adjacent sub-bands are sequentially identified and assigned to the same physical index sequence. Below this, components with correlation coefficients below a preset threshold are identified as isolated noise components and removed. This sorting and normalization process eliminates the randomization of logical order caused by frequency domain independent decoupling, ensuring that spectral components representing the same radiation source in each sub-band are arranged in a consistent row vector.

[0115] S4. Combining the spatial coordinates of the three-dimensional sensing array, map each independent spectral component signal to the three-dimensional spatial coordinate system to construct a three-dimensional spectral feature fingerprint of each radiation source region.

[0116] Preferably, the steps for constructing the three-dimensional spectral feature fingerprint of each radiation source region include:

[0117] Extract the characteristic parameters of each independent spectral component signal in the corresponding sub-band, the characteristic parameters including center frequency intensity, frequency band energy distribution and spectral entropy value;

[0118] Based on the response weights of each channel in the three-dimensional sensing array to specific independent components, the physical center coordinates of each radiation source region in the three-dimensional coordinate system are calculated using a spatial inversion positioning algorithm.

[0119] The physical center coordinates are associated with the corresponding feature parameters in a multi-dimensional attribute relationship to generate a spatial-frequency domain fusion descriptor for each radiation source region, which serves as the three-dimensional spectral feature fingerprint.

[0120] Specifically, for each independent spectral component signal obtained after reconstruction, feature quantization is performed in the corresponding sub-band frequency domain. First, the center frequency intensity is extracted, that is, the peak amplitude and corresponding frequency position of the independent component in the spectral envelope are identified; second, the frequency band energy distribution is calculated by integrating the power spectral density within a specific frequency band to obtain the energy distribution curve as a function of frequency; finally, the spectral entropy value is calculated, and the information entropy is calculated using the normalized probability distribution of energy at each frequency point to characterize the complexity and randomness of the radiation source signal. The above parameters together constitute the frequency domain feature vector of the independent component.

[0121] The spatial location of the radiation source is determined by utilizing the response weights of each magnetic sensor channel in a three-dimensional sensing array to specific independent components. Since spatial crosstalk has been eliminated for the independent components, the strength of each sensing channel's response depends only on the physical distance and direction vector between the radiation source and the sensor. A magnetic dipole model inversion algorithm is then used, combined with the spatial coordinates of each sensor. Inverse fitting is performed to calculate the physical center coordinates of each radiation source region in the three-dimensional coordinate system. .

[0122] The calculated physical center coordinates are correlated with the extracted feature parameters (center frequency intensity, energy distribution, and spectral entropy value) using multi-dimensional attributes. Through data mapping logic, geospatial information and signal physical characteristics are encapsulated into a unified feature vector, thus generating a spatial-frequency domain fusion descriptor.

[0123] S5. Match the three-dimensional spectral feature fingerprint with a preset standard feature library, and identify the state features of the subject under test based on the matching deviation.

[0124] Preferably, the step of identifying the state characteristics of the subject under test based on the matching deviation includes:

[0125] Input the three-dimensional spectral feature fingerprints corresponding to each radiation source region into a preset classification and discrimination model;

[0126] Calculate the cosine similarity between the three-dimensional spectral feature fingerprint and the corresponding benchmark fingerprint in the standard feature library to obtain the quantized matching deviation.

[0127] Determine whether the matching deviation exceeds a preset abnormal threshold range;

[0128] If the deviation exceeds the limit, the physical state offset category of the corresponding radiation source region is identified based on the distribution characteristics of the matching deviation, and the corresponding state feature analysis report is output.

[0129] Specifically, the constructed three-dimensional spectral feature fingerprints (i.e., spatial-frequency domain fusion descriptors) of each radiation source region are input into a pre-defined classification and discrimination model. This model is built based on a supervised learning algorithm and internally stores a large amount of standard feature data under known states. The model first normalizes the input feature fingerprints to ensure that all dimensions of its attributes (coordinates, intensity, energy distribution, entropy) are consistent in numerical magnitude, facilitating subsequent similarity measurements.

[0130] The baseline fingerprint closest to the spatial coordinates of the current radiation source is retrieved from the standard feature library. The cosine similarity algorithm is then used to calculate the three-dimensional spectral feature fingerprint vector of the target source. Compared with the baseline fingerprint vector cosine value of the angle between Based on the obtained similarity values, the formula is used to... Calculate the quantized matching deviation This deviation reflects the geometric distance between the current radiation source's spectral characteristics and the standard healthy or qualified state.

[0131] The calculated matching deviation Compared with the system's preset abnormal threshold range Perform real-time comparison. If If so, the radiation source area is determined to be in a standard preset state; if If the system detects an anomaly in the region, it determines that the region is in an abnormal state. At this point, the system further analyzes the distribution characteristics of the matching deviation across various dimensions (such as frequency offset direction and energy attenuation level), and identifies the specific category of the physical state deviation through logistic regression or decision tree discrimination. Based on the identified deviation category and its corresponding physical center coordinates, the system generates a structured state feature analysis report. This report details the spatial location information of the anomalous radiation source, its spectral deviation values, and the associated state feature descriptions.

[0132] Example 2

[0133] This invention also provides a system for identifying and analyzing three-dimensional spectral features of magnetic resonance signals, such as... Figure 2 As shown, it includes:

[0134] The magnetic field signal acquisition module is used to acquire multi-channel time-domain magnetic field signal sequences through a three-dimensional sensing array preset around the subject under test;

[0135] The magnetic field signal transformation module is used to perform Fourier transform on the time-domain magnetic field signals of each channel to obtain a complex spectrum matrix containing amplitude and phase information;

[0136] The magnetic field spectrum decoupling module is used to spatially decouple the complex spectrum matrix in the complex domain using statistical independence criteria, and to extract independent spectral component signals corresponding to different radiation source regions.

[0137] The spectrum fingerprint construction module is used to combine the spatial coordinates of the three-dimensional sensing array to map each independent spectrum component signal to the three-dimensional spatial coordinate system and construct the three-dimensional spectrum feature fingerprint of each radiation source region.

[0138] The spectrum deviation identification module is used to match the three-dimensional spectrum feature fingerprint with a preset standard feature library, and identify the state features of the subject under test based on the degree of matching deviation.

[0139] The system described in this invention is applied to the non-destructive testing of internal metal parts in large industrial equipment. Addressing the problem that deep defects within complex mechanical structures are difficult to remove using conventional methods, this embodiment provides a specific application approach:

[0140] A three-dimensional sensing array consisting of 32 tunnel magnetoresistive sensors is deployed around the outer periphery of the multi-layer welded structure according to its geometric contour. During equipment operation or the application of an external low-frequency excitation magnetic field, the sensing array synchronously acquires the original magnetic field signals of each detection point. By using a noise reduction algorithm to eliminate the constant geomagnetic background and surrounding electromagnetic power frequency interference, a multi-channel time-domain magnetic field signal sequence reflecting the distribution of induced current inside the metal is obtained.

[0141] The magnetic field signal transformation module performs windowing and framing processing on the acquired signals from each channel, and then executes a 2048-point Fast Fourier Transform. Since defects at different depths within the metal exhibit slight differences in the frequency and phase of their magnetic response under the excitation field, the system extracts the real and imaginary parts of each frequency point to construct a complex spectrum matrix. This matrix fully preserves the phase information representing the defect location and the amplitude information representing the defect size.

[0142] To address the issue of superimposed magnetic signals from multiple components in complex metal parts, the module performs independent component analysis in the complex domain, dividing the spectrum into multiple narrow sub-bands. It leverages the linear instantaneous mixing characteristics of signals within each sub-band to reduce computational dimensionality. Using the principle of maximizing non-Gaussianity, it extracts weak defect signals masked by strong background metal matrix signals. The extraction result is shown below. Figure 3 As shown. By judging the consistency of spatial response, it is ensured that the independent components extracted at different frequencies all point to the same physical defect point, thus eliminating arrangement ambiguity.

[0143] The system utilizes a spatial inversion localization algorithm to calculate the three-dimensional physical center coordinates of the defect within the metal part based on the response weights of each sensor channel to the extracted defect signal. Subsequently, the center frequency energy and spectral entropy value of this independent component are extracted, and the coordinate information is fused with the spectral characteristics of the defect to generate a spatial-frequency domain fusion descriptor for the defect. This constructs a defect fingerprint for the specific metal part, clarifying the specific depth, coordinates, and damage extent of the defect.

[0144] The module compares the generated defect fingerprint with a standard feature library (which stores reference magnetic response data for defect-free and qualified parts). It calculates the cosine similarity between the current fingerprint vector and the reference fingerprint vector; if the match deviation exceeds a certain threshold... If the abnormal threshold is reached, the system identifies it as internal fatigue cracks or material segregation. The system outputs a non-destructive testing report, indicating the specific part number where the defect occurred, its three-dimensional coordinates, and the damage level, providing a quantitative basis for preventive maintenance of the equipment.

[0145] In this embodiment, such as Figure 3 As shown, during the decoupling process, the magnetic field spectrum decoupling module, due to the different sub-bands ( to The decoupling matrix is ​​generated independently, leading to a random distribution of the independent component indices of the same physical defect in different subbands, resulting in... Figure 3 (a) illustrates the ambiguity issue in the arrangement. Direct extraction will lead to confusion in the spectral features along the frequency axis. Further reference... Figure 3(b) This invention extracts the spatial guiding vectors of each candidate independent component (reflected in the sensor channel response energy patterns in the figure) and calculates the cross-correlation coefficients between vectors in adjacent sub-bands. Experimental observations revealed that the spatial response patterns representing the same physical defect (e.g., defects A, B, and C) exhibit high morphological consistency across a wide frequency band. Using this consistency criterion, the system rearranges the components of each sub-band: signals pointing to the peak response regions of channels 4-6 are grouped into the defect A sequence, signals pointing to the peak response regions of channels 10-12 are grouped into the defect B sequence, and signals pointing to the bimodal response regions of channels 1 and 16 are grouped into the defect C sequence. The aligned results show that components with the same index in different sub-bands all point to the same physical radiation source in space. By eliminating arrangement ambiguities, the continuity and integrity of the weak defect signals extracted from strong background noise in the frequency domain are ensured, providing a reliable data foundation for the subsequent construction of accurate defect fingerprints.

[0146] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying and analyzing three-dimensional spectral features of magnetic resonance signals, characterized in that, include: A multi-channel time-domain magnetic field signal sequence is acquired by a three-dimensional sensing array pre-set around the subject under test; Fourier transform is performed on the time-domain magnetic field signals of each channel to obtain a complex spectrum matrix containing amplitude and phase information; In the complex domain, the statistical independence criterion is used to spatially decouple the complex spectrum matrix and extract the independent spectrum component signals corresponding to different radiation source regions. By combining the spatial coordinates of the three-dimensional sensing array, each independent spectral component signal is mapped to the three-dimensional spatial coordinate system to construct a three-dimensional spectral feature fingerprint of each radiation source region. The three-dimensional spectral fingerprint is matched with a preset standard feature library, and the state characteristics of the subject under test are identified based on the matching deviation.

2. The method for three-dimensional spectral feature identification and analysis of magnetic resonance signals according to claim 1, characterized in that, The steps for obtaining a multi-channel time-domain magnetic field signal sequence include: Based on the preset geometric topology, multiple magnetic sensors are deployed on the outer periphery of the subject under test to construct a three-dimensional sensing array. The magnetic signals captured by each magnetic sensor in the three-dimensional sensing array are synchronously triggered and acquired through a multi-channel sampling circuit to obtain multiple raw time-domain sequences. By using a noise reduction algorithm, background interference is removed from the multiple original time-domain sequences using a reference environmental signal, resulting in the time-domain magnetic field signal sequence corresponding to each channel.

3. The method for three-dimensional spectral feature identification and analysis of magnetic resonance signals according to claim 1, characterized in that, The steps for performing Fourier transform on the time-domain magnetic field signals of each channel include: The time-domain magnetic field signal sequences of each channel are processed by frame division to generate multiple time-domain signal frames with a preset overlap rate; A preset window function is used to perform windowing smoothing on each of the time-domain signal frames to suppress spectral leakage characteristics; A fast Fourier transform is performed on each of the windowed time-domain signal frames to convert each signal frame from the time domain to the frequency domain, and the corresponding amplitude and phase information are extracted to construct a complex spectrum matrix.

4. The method for three-dimensional spectral feature identification and analysis of magnetic resonance signals according to claim 1, characterized in that, The steps for spatial decoupling the complex spectrum matrix include: According to the preset frequency band step size, the complex spectrum matrix is ​​divided into multiple independent sub-band matrices to be processed; Solve the subband matrix to be processed to obtain the decoupling matrix corresponding to each subband, and extract the candidate independent components under each subband based on the decoupling matrix; Determine the spatial response consistency of the candidate independent components, and perform frequency alignment and sorting normalization processing on the candidate independent components under different sub-bands to eliminate the ambiguity of arrangement introduced by frequency domain transformation. The signal is reconstructed based on the aligned components of each sub-band, and the independent spectral component signals corresponding to different target radiation source regions are restored.

5. The method for three-dimensional spectral feature identification and analysis of magnetic resonance signals according to claim 4, characterized in that, Solving for the subband matrix to be processed is achieved through independent component analysis, the steps of which include: Centering and whitening preprocessing are performed on each of the sub-band matrices to be processed in order to eliminate the correlation between the signals of each channel and unify the variance; The objective function is to maximize the non-Gaussianity of the feature signal, and the preprocessed matrix is ​​iteratively optimized in the complex domain. When the objective function reaches the convergence condition, the complex-valued separation matrix corresponding to each sub-band is output as the decoupling matrix.

6. The method for three-dimensional spectral feature identification and analysis of magnetic resonance signals according to claim 4, characterized in that, The steps for determining the spatial response consistency of the candidate independent components include: Obtain the spatial guiding vector corresponding to each candidate independent component under each sub-band; Calculate the correlation coefficient of the spatial guidance vector between adjacent sub-bands or between preset reference sub-bands; Based on the magnitude of the correlation coefficient, the candidate independent components are clustered and sorted, and independent components belonging to the same radiation source region are merged into the same index sequence to complete frequency alignment.

7. The method for three-dimensional spectral feature identification and analysis of magnetic resonance signals according to claim 1, characterized in that, The steps for constructing the three-dimensional spectral fingerprint of each radiation source region include: Extract the characteristic parameters of each independent spectral component signal in the corresponding sub-band, the characteristic parameters including center frequency intensity, frequency band energy distribution and spectral entropy value; Based on the response weights of each channel in the three-dimensional sensing array to specific independent components, the physical center coordinates of each radiation source region in the three-dimensional coordinate system are calculated using a spatial inversion positioning algorithm. The physical center coordinates are associated with the corresponding feature parameters in a multi-dimensional attribute relationship to generate a spatial-frequency domain fusion descriptor for each radiation source region, which serves as the three-dimensional spectral feature fingerprint.

8. The method for three-dimensional spectral feature identification and analysis of magnetic resonance signals according to claim 1, characterized in that, The steps for identifying the state characteristics of the subject under test based on the matching deviation include: Input the three-dimensional spectral feature fingerprints corresponding to each radiation source region into a preset classification and discrimination model; Calculate the cosine similarity between the three-dimensional spectral feature fingerprint and the corresponding benchmark fingerprint in the standard feature library to obtain the quantized matching deviation. Determine whether the matching deviation exceeds a preset abnormal threshold range; If the deviation exceeds the limit, the physical state offset category of the corresponding radiation source region is identified based on the distribution characteristics of the matching deviation, and the corresponding state feature analysis report is output.

9. A system for three-dimensional spectral feature recognition and analysis of magnetic resonance signals, characterized in that, The method for performing the three-dimensional spectral feature identification and analysis of magnetic resonance signals according to any one of claims 1-8 includes: The magnetic field signal acquisition module is used to acquire multi-channel time-domain magnetic field signal sequences through a three-dimensional sensing array preset around the subject under test; The magnetic field signal transformation module is used to perform Fourier transform on the time-domain magnetic field signals of each channel to obtain a complex spectrum matrix containing amplitude and phase information; The magnetic field spectrum decoupling module is used to spatially decouple the complex spectrum matrix in the complex domain using statistical independence criteria, and to extract independent spectral component signals corresponding to different radiation source regions. The spectrum fingerprint construction module is used to combine the spatial coordinates of the three-dimensional sensing array to map each independent spectrum component signal to the three-dimensional spatial coordinate system and construct the three-dimensional spectrum feature fingerprint of each radiation source region. The spectrum deviation identification module is used to match the three-dimensional spectrum feature fingerprint with a preset standard feature library, and identify the state features of the subject under test based on the degree of matching deviation.

Images