A wireless coordination and intelligent identification system based on X-ray flaw detection of strain clamp

Abstract

The present application belongs to the field of image recognition, and particularly relates to a wireless cooperation and intelligent identification system based on X-ray flaw detection of strain clamp. The system comprises a wireless receiver, a channel state monitor, a frequency domain adaptive filter and a physical feature guided identification processor. The channel state monitor extracts the signal-to-noise ratio and multipath time delay spread parameters of the wireless transmission link. The frequency domain adaptive filter calculates the interference energy distribution interval according to the above parameters and dynamically generates a group of notch filters to perform frequency domain adaptive filtering on the image data stream. The physical feature guided identification processor extracts the gray scale distribution features of the filtered image data, calculates and generates a weight map in combination with the linear attenuation coefficient equation of the ray in aluminum and steel materials, inputs the filtered image data and the weight map into an identification network to output a defect identification result. The present application filters out channel interference noise, suppresses the scattered gray scale fluctuation at the material interface, and eliminates the interference of noise and artifacts on defect identification.

Description

Technical Field

[0001] This invention belongs to the field of image recognition, specifically relating to a wireless collaborative and intelligent recognition system based on X-ray flaw detection using tension clamps. Background Technology

[0002] In existing tension clamp X-ray inspection systems, to address the difficulty of wiring, wireless image transmission technology is used to transmit image data streams collected by flat panel detectors to the terminal during field operations. Current technologies typically employ fixed mean or median filtering algorithms for noise reduction of the received image data stream at the receiving end. For images generated at the interface between the aluminum tube and the internal steel core of the tension clamp, the terminal usually uses common data-driven models such as convolutional neural networks to classify and identify porosity and inclusion defects.

[0003] In the complex electromagnetic environment of power transmission lines, wireless transmission links introduce periodic strip noise, the frequency of which overlaps with the inherent texture frequency of the image. Existing fixed filtering algorithms cannot distinguish between the inherent texture of the image and the noise introduced by the channel, resulting in the destruction of the original texture structure of the image when filtering out noise. When rays penetrate the junction of aluminum tube and steel core, they scatter, causing tiny pores to appear as low-contrast areas in the image. General data-driven models lack constraints on the physical attenuation law of rays, easily misjudging grayscale fluctuations caused by scattering as defects. Current technologies cannot simultaneously achieve directional filtering of channel noise and defect feature identification based on physical attenuation law under complex electromagnetic interference. Summary of the Invention

[0004] The purpose of this invention is to provide a wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A wireless collaborative and intelligent identification system for X-ray flaw detection of tension wire clamps includes a wireless receiver, a channel state monitor, a frequency-domain adaptive filter, and a physical feature-guided identification processor. A flat panel detector acquires X-ray images of the tension wire clamps and outputs an image data stream. The wireless receiver is communicatively connected to the flat panel detector to receive the image data stream. The channel state monitor is connected to the wireless receiver to extract the signal-to-noise ratio (SNR) and multipath delay spread parameters of the wireless transmission link. The frequency-domain adaptive filter is connected to both the channel state monitor and the wireless receiver, and calculates the interference of the wireless channel in the frequency domain based on the SNR and the multipath delay spread parameters. The system dynamically generates a notch filter bank within the energy distribution range of the disturbance. The frequency-domain adaptive filter uses the notch filter bank to perform frequency-domain adaptive filtering on the image data stream to output filtered image data. The physical feature-guided recognition processor is connected to the frequency-domain adaptive filter. The physical feature-guided recognition processor extracts the grayscale distribution features of the filtered image data and calculates and generates a ray attenuation physical feature-guided weight map by combining it with the linear attenuation coefficient equation of X-rays in aluminum and steel materials. The physical feature-guided recognition processor inputs the filtered image data and the ray attenuation physical feature-guided weight map into the recognition network to output the porosity and inclusion defect recognition results of the intelligent recognition system.

[0007] Preferably, the channel state monitor includes an orthogonal frequency division multiplexing (OFDM) signal analysis component and a delay spread calculation component. The OFDM signal analysis component performs a fast Fourier transform on the physical layer preamble of the image data stream received by the wireless receiver to obtain multiple subcarrier channel frequency domain response values. The delay spread calculation component performs an inverse fast Fourier transform on the multiple subcarrier channel frequency domain response values ​​to obtain a time-domain power delay distribution spectrum. The delay spread calculation component extracts multipath components exceeding a preset noise threshold from the time-domain power delay distribution spectrum and calculates the multipath delay spread parameter based on the arrival time and power value of the multipath components. The OFDM signal analysis component calculates the signal-to-noise ratio (SNR) based on the sum of the squares of the real and imaginary parts of the multiple subcarrier channel frequency domain response values.

[0008] Preferably, the frequency domain adaptive filter includes a spectrum energy mapping component and a notch filtering component. The spectrum energy mapping component converts the signal-to-noise ratio into a noise floor power spectral density and converts the multipath delay spread parameter into a frequency selective fading depth parameter. The spectrum energy mapping component calculates the maximum values ​​of the noise floor power spectral density and the frequency selective fading depth parameter within their respective ranges. The two parameters are divided by their corresponding maximum values ​​to complete the unified dimension normalization process. The normalized noise floor power spectral density is used as a flat distribution base across the entire bandwidth. The normalized frequency selective fading depth parameter is used as the peak value at the center frequency to generate a Gaussian fading distribution. The flat distribution and the Gaussian distribution are superimposed to generate the total interference energy spectrum.

[0009] The spectrum energy mapping component adds the noise floor power spectral density to the frequency selective fading depth parameter to generate the interference energy distribution interval. The notch filter execution component locates the set of frequency coordinates corresponding to the interference energy distribution interval in the two-dimensional Fourier transform spectrum of the image data stream. The conversion scaling factor is the ratio of the maximum value of the horizontal spatial frequency of the image to the bandwidth of the wireless channel. The start frequency and end frequency of the interference energy distribution interval are multiplied by this ratio to obtain the corresponding horizontal spatial frequency range. The vertical spatial frequency range is taken as the same numerical range as the horizontal spatial frequency range. All coordinate points in the two-dimensional spectrum that fall within the rectangular interval are marked as the set of frequency coordinates.

[0010] Preferably, the physical feature-guided recognition processor includes a material attenuation modeling component and a weight map generation component. The material attenuation modeling component acquires the aluminum tube thickness and internal steel core diameter parameters of the tension clamp and sets the initial radiation intensity of the X-ray source. The material attenuation modeling component completes imaging geometric calibration through a standard calibration plate, establishes a one-to-one correspondence between pixel coordinates and physical space coordinates, and determines the radiation projection path as a parallel ray perpendicular to the detector plane based on the relative position of the X-ray source and the detector. The material attenuation modeling component constructs the linear attenuation coefficient equation based on the initial radiation intensity, the linear attenuation coefficient of aluminum material corresponding to the aluminum tube thickness, and the linear attenuation coefficient of steel material corresponding to the internal steel core diameter. The weight map generation component calculates the thickness of the radiation penetrating the aluminum tube and steel core based on the physical space coordinates corresponding to the pixel coordinates, substitutes the calculated thickness value into the linear attenuation coefficient equation to solve for the theoretical gray value, and uses the absolute value of the difference between the theoretical gray value and the actual gray value as the initial weight. The weight map generation component normalizes the initial weight to generate the radiation attenuation physical feature-guided weight map.

[0011] Preferably, the physical feature-guided recognition processor includes a feature extraction backbone network and a weight fusion component. The feature extraction backbone network includes multiple convolutional layers and pooling layers connected in series to downsample the filtered image data and output a multi-scale convolutional feature map. The weight fusion component performs bilinear interpolation on the ray attenuation physical feature-guided weight map to obtain a spatial attention matrix with the same size as the multi-scale convolutional feature map. The weight fusion component multiplies the spatial attention matrix element-wise with the multi-scale convolutional feature map to output a weighted feature map. The physical feature-guided recognition processor includes a classification head component. The classification head component performs global average pooling and fully connected mapping on the weighted feature map to output the identification results of pores and inclusion defects.

[0012] Preferably, the wireless receiver includes a frame header parsing component and a data reassembly component. The frame header parsing component extracts the data frame header containing row number identifiers and column number identifiers from the image data stream. The data reassembly component reassembles the out-of-order received data frames into the image data stream in the form of a two-dimensional matrix according to the row number identifiers and column number identifiers. The channel state monitor synchronously reads the channel estimation pilot symbols in the physical layer link control frame when the frame header parsing component extracts the data frame header. The frequency domain adaptive filter directly performs frequency domain adaptive filtering operation on the data in the row buffer queue of the image data stream in the form of the two-dimensional matrix output by the data reassembly component.

[0013] Preferably, the time delay spread calculation component includes a Hanning window function applicator and a moment calculator. The Hanning window function applicator multiplies a fixed-length Hanning window function with the time-domain power delay distribution spectrum in the time domain to truncate the multipath components within a preset time window. The moment calculator performs first-order moment calculation on the truncated time-domain power delay distribution spectrum to obtain the average multipath arrival time. The moment calculator performs second-order central moment calculation on the truncated time-domain power delay distribution spectrum to obtain the multipath delay variance. The time delay spread calculation component outputs the square root of the multipath delay variance as the multipath delay spread parameter to the frequency domain adaptive filter.

[0014] Preferably, the notch filter execution component includes a pole arbiter and a quality factor calculator. The pole arbiter sets conjugate pole pairs inside the unit circle of the complex frequency plane according to the center frequency in the frequency coordinate set. The quality factor calculator calculates the bandwidth of the interference energy distribution range. The quality factor calculator uses the ratio of the center frequency to the bandwidth as the quality factor of the second-order IIR notch filter. The pole arbiter adjusts the polar radius distance of the conjugate pole pairs from the unit circle of the complex frequency plane according to the quality factor to generate the second-order IIR notch filter with a specified notch depth. The notch filter execution component combines the multiple second-order IIR notch filters in a cascaded manner to form the notch filter bank.

[0015] Preferably, the weight map generation component includes a boundary fitter and a smooth transition processor. The boundary fitter delineates the aluminum tube region, the steel core region, and the aluminum-steel interface region in the filtered image data based on the aluminum tube thickness and the internal steel core diameter parameters. The boundary fitter adds a geometric transition compensation coefficient to the pixel coordinates within the aluminum-steel interface region to update the linear attenuation coefficient equation. The smooth transition processor performs spatial convolution operations on the initial weights at the edge pixel coordinates of the aluminum-steel interface region using a Gaussian kernel function. The smooth transition processor inputs the initial weights after the convolution operation into the normalization process to generate a ray attenuation physical feature guided weight map with smooth edges.

[0016] Preferably, the classification head component includes a receptive field pyramid pooler and a focus loss calculator, and the classification head component operates in offline training mode and online flaw detection mode;

[0017] In online flaw detection mode, the receptive field pyramid pooler performs adaptive average pooling at different scales on the weighted feature map to obtain multi-scale pooling feature vectors. The receptive field pyramid pooler concatenates the multi-scale pooling feature vectors by channel dimension and inputs them into a fully connected mapping layer to output the porosity and inclusion defect identification results.

[0018] In offline training mode, the focus loss calculator obtains the label matrix of the porosity and inclusion defect identification results. The focus loss calculator calculates the cross-entropy loss based on the label matrix and the output vector of the fully connected mapping layer. The focus loss calculator generates a loss value based on the modulation factor of the exponential term of the cross-entropy loss and backpropagates it to the feature extraction backbone network to update the convolution kernel parameters of the feature extraction backbone network.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] 1. The signal-to-noise ratio (SNR) and multipath delay spread parameters of the wireless transmission link are extracted by a channel state monitor. A frequency-domain adaptive filter calculates the interference energy distribution range based on these parameters and dynamically generates a notch filter bank for frequency-domain adaptive filtering of the image data stream. A physical feature-guided recognition processor calculates and generates a ray attenuation physical feature-guided weight map by combining the linear attenuation coefficient equation of rays in aluminum and steel materials, and inputs it along with the filtered image data into the recognition network. This approach introduces the physical layer parameters of wireless communication and the physical laws of ray imaging into the image processing flow. In the frequency domain, channel interference overlapping with the inherent frequency of the image is eliminated. Simultaneously, the physical attenuation characteristics are used to amplify abnormally low grayscale areas that conform to the attenuation law, suppressing grayscale fluctuations caused by edge scattering and eliminating interference from channel noise and scattering artifacts on defect identification.

[0021] 2. The time-domain power delay distribution spectrum was obtained through orthogonal frequency division multiplexing signal analysis and inverse fast Fourier transform. Multipath components were extracted by applying the Hanning window function and calculating moments, improving the accuracy of multipath delay extension parameter calculation. When generating the notch filter bank, a second-order infinite impulse response notch filter with a specified notch depth was constructed by setting conjugate pole pairs in the complex frequency plane and adjusting the pole distance according to the ratio of center frequency to bandwidth. When generating the weight map, the aluminum tube region, steel core region, and aluminum-steel interface region were defined, and a geometric transition compensation coefficient was added to the interface region. Spatial convolution using a Gaussian kernel function eliminated abrupt weight changes at material boundaries. In the recognition network, a spatial attention matrix was generated through bilinear interpolation and multiplied element-wise with the multi-scale convolutional feature map. Network parameters were updated by combining receptive field pyramid pooling and focus loss calculation, improving the completeness of feature representation of micropores and inclusion defects in the weighted feature map. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the wireless collaborative and intelligent identification system for X-ray flaw detection based on tension wire clamps provided in an embodiment of the present invention;

[0023] Figure 2 A flowchart for extracting channel state parameters provided in an embodiment of the present invention;

[0024] Figure 3 This is a flowchart of frequency domain adaptive filtering provided in an embodiment of the present invention;

[0025] Figure 4 A flowchart for generating a weighted graph provided in an embodiment of the present invention;

[0026] Figure 5 This is a flowchart of feature extraction and fusion provided in an embodiment of the present invention;

[0027] Figure 6 A flowchart illustrating defect classification and network update provided in this embodiment of the invention. Detailed Implementation

[0028] refer to Figure 1 This embodiment provides a wireless collaborative and intelligent identification system for X-ray flaw detection of tension clamps deployed in a power transmission line field for X-ray flaw detection operations. The system includes a wireless receiver, a channel status monitor, a frequency domain adaptive filter, and a physical feature-guided identification processor. A flat panel detector works in conjunction with an X-ray emission source to complete X-ray transmission imaging of the tension clamps. The flat panel detector converts the acquired X-ray images of the tension clamps into a digital image data stream and transmits it to the wireless receiver via a wireless communication link. The flat panel detector uses an X-ray imaging detector based on an amorphous silicon photodiode array. The acquired image data stream is a 16-bit grayscale digital image sequence. Each frame corresponds to the transmission image of the tension clamp obtained from a single X-ray exposure. The pixel resolution of each image frame is 3072×2048, and the grayscale value of each pixel has a linear mapping relationship with the intensity of the X-rays reaching the detector plane after penetrating the tension clamp.

[0029] The wireless receiver employs an orthogonal frequency division multiplexing (OFDM) wireless communication receiver operating in the 2.4GHz ISM band. It establishes a point-to-point communication link with the built-in wireless transmitter module of the flat panel detector. The receiver's RF front-end performs down-conversion and analog-to-digital conversion of the wireless signal, then outputs the baseband signal to the baseband processing unit. The baseband processing unit demodulates and decodes the signal to reconstruct the image data stream transmitted by the flat panel detector. The wireless receiver includes a frame header parsing component and a data reassembly component. The frame header parsing component parses the demodulated and decoded data frames, extracting fields such as row and column identifiers, frame count identifiers, and cyclic redundancy check (CRC) codes from the frame header. The frame header parsing component verifies the integrity of the data frames using the CRC code, discarding erroneous data frames that fail the check and outputting the verified data frames to the data reassembly component. The data reconstruction component has a pre-defined two-dimensional matrix buffer space corresponding to the image frame resolution. Based on the row and column identifiers in the data frame header, out-of-order received data frames are filled into the corresponding positions in the two-dimensional matrix buffer space. Once all data frames corresponding to all row numbers have been received, they are reconstructed into a complete two-dimensional matrix image data stream. While the frame header parsing component extracts the data frame header, the channel state monitor simultaneously reads the channel estimation pilot symbols in the physical layer link control frame, completing the real-time extraction of channel state parameters and ensuring the time synchronization between the channel state parameters and the corresponding image data stream. The frequency domain adaptive filter is connected to the row buffer queue of the data reconstruction component. During the output of the two-dimensional matrix image data stream by the data reconstruction component, frequency domain adaptive filtering is directly performed on each row of data in the row buffer queue, achieving pipelined processing of the image data and reducing processing latency.

[0030] The channel state monitor (SVM) is connected to the baseband processing unit of the wireless receiver. It acquires the physical layer baseband signal of the wireless transmission link from the baseband processing unit. The SVM analyzes the channel estimation pilot symbols in the baseband signal to extract the current signal-to-noise ratio (SNR) and multipath delay spread (MDP) parameters of the wireless transmission link. The SNR parameter represents the ratio of signal power to noise power in the wireless transmission link, reflecting the overall noise level of the link. The MDP parameter represents the time dispersion of arrival of different multipath components in the wireless transmission link, reflecting the frequency-selective fading characteristics of the link. The SVM outputs the extracted SNR and MDP parameters to a frequency-domain adaptive filter in real time.

[0031] The frequency-domain adaptive filter simultaneously receives image data streams from the wireless receiver and signal-to-noise ratio (SNR) and multipath delay spread parameters from the channel state monitor. Since the noise floor power spectral density (NDPS) and frequency-selective fading depth (FDR) parameters have different dimensions and cannot be directly superimposed, the frequency-domain adaptive filter first normalizes these two parameters: it calculates the maximum values ​​of the NDPS and FDR parameters within their respective preset ranges, and divides each parameter by its corresponding maximum value to obtain a normalized scalar with a range of 0-1. The normalized NDPS is used as a flat distribution base across the entire bandwidth, and the normalized FDR parameter is used as the peak value at the center frequency to generate a Gaussian fading distribution. The flat distribution and the Gaussian distribution are superimposed to generate the total interference energy spectrum of the wireless channel in the frequency domain. The frequency-domain adaptive filter has a preset interference energy threshold. Continuous frequency bands in the total interference energy spectrum whose energy values ​​exceed the interference energy threshold are marked as complete interference energy distribution intervals. These intervals include the start frequency, end frequency, center frequency, and bandwidth parameters.

[0032] To achieve accurate mapping from link frequency domain parameters to image spatial frequency coordinates, the frequency domain adaptive filter locates the set of frequency coordinates corresponding to the interference energy distribution interval in the image's two-dimensional Fourier transform spectrum according to the following rules: The ratio of the maximum horizontal spatial frequency value to the wireless channel system bandwidth is calculated as a conversion scaling factor. The start and end frequencies of the interference energy distribution interval are multiplied by this conversion scaling factor to obtain the corresponding horizontal spatial frequency range. The vertical spatial frequency range is taken from the same numerical interval as the horizontal spatial frequency range, and all coordinate points in the two-dimensional spectrum falling within this rectangular interval are marked as the target frequency coordinate set. Based on parameters such as the center frequency and bandwidth of the interference energy distribution interval, the frequency domain adaptive filter dynamically generates a corresponding number and parameters of notch filter banks. Each notch filter unit corresponds to a sub-band covered by the target frequency coordinate set, used to filter out channel interference within that sub-band.

[0033] The frequency-domain adaptive filter first performs a two-dimensional Fourier transform on each frame of the image data stream, converting the image from the spatial domain to the frequency domain, obtaining the two-dimensional spectral data of the image. Then, it uses a generated notch filter bank to filter out the spectral components corresponding to the interference energy distribution range in the two-dimensional spectral data. After filtering, it performs an inverse two-dimensional Fourier transform on the processed two-dimensional spectral data, converting the data back from the frequency domain to the spatial domain, obtaining the filtered image data. The frequency-domain adaptive filter outputs the filtered image data to the physical feature-guided recognition processor.

[0034] After receiving filtered image data, the physical feature-guided recognition processor first extracts grayscale distribution features from the filtered image data to obtain the actual grayscale value corresponding to each pixel coordinate in each frame of the image, as well as features such as the grayscale histogram distribution and grayscale gradient distribution of the entire image. The physical feature-guided recognition processor pre-stores the linear attenuation coefficients of X-rays in aluminum and steel materials. Combined with the structural parameters of the tension clamp, it constructs an equation for the linear attenuation coefficient of X-rays in the tension clamp. The expression for the linear attenuation coefficient equation is as follows:

[0035]

[0036] in, Coordinates in the image The intensity of the X-rays that reach the detector after transmission at the corresponding location. The initial radiation intensity of the X-ray source is denoted as . The linear attenuation coefficient of X-rays in aluminum is given. coordinates The X-ray penetration thickness of the corresponding tension clamp aluminum tube at that location. The linear attenuation coefficient of X-rays in steel is given by [the parameter]. coordinates The ray penetration thickness of the steel core inside the tension clamp at the corresponding location.

[0037] The physical feature-guided recognition processor converts the actual grayscale value of each pixel coordinate in the filtered image data into the corresponding actual X-ray intensity value. Combining this with the linear attenuation coefficient equation, it calculates the theoretical X-ray intensity and theoretical grayscale value corresponding to each pixel coordinate. Then, based on the difference between the theoretical and actual grayscale values, it calculates the weight value for each pixel coordinate, ultimately generating a X-ray attenuation physical feature-guided weight map with the same size as the filtered image data. In the X-ray attenuation physical feature-guided weight map, the weight value of each pixel represents the degree of deviation between the actual grayscale value at that coordinate and the theoretical grayscale value that conforms to the physical attenuation law of X-rays. The greater the deviation, the higher the weight value, and the higher the probability of a defect at that coordinate. The physical feature-guided recognition processor synchronously inputs the filtered image data and the X-ray attenuation physical feature-guided weight map into the recognition network. The recognition network first uses the X-ray attenuation physical feature-guided weight map to guide the feature extraction process of the filtered image data, amplifying the feature response in areas with higher weight values ​​and suppressing the feature response in areas with lower weight values. Then, it performs classification and recognition processing on the guided features, finally outputting the identification results of pores and inclusion defects in the tension clamp image. The identification results include information such as the defect type, its coordinate location, and its contour range.

[0038] Table 1. Frame Structure Parameters of X-ray Image Data Stream for Tension Clamps

[0039] Image frame pixel resolution 3072×2048 The number of horizontal and vertical pixels in a single frame of an X-ray image Pixel grayscale bit depth 16bit The number of bits used to quantize the grayscale value of each pixel corresponds to a grayscale value range of 0-65535. Data frame payload length 12288 bytes The payload data length of a single data frame, corresponding to the number of bytes per row of pixels. Frame header length 16 bytes The header bytes of a single data frame are of length and include fields such as row number identifier, column number identifier, and checksum. Total number of data frames per frame 2048 The number of split data frames corresponding to a single complete image frame is consistent with the number of vertical pixels in the image. Wireless transmission modulation method QPSK / 16QAM Adaptive The modulation scheme of the wireless link adaptively switches according to the channel state.

[0040] Specifically, the frame structure parameters defined in Table 1 are used to standardize the encapsulation format of the image data stream transmitted by the flat panel detector. The wireless receiver uses these frame structure parameters to parse and reconstruct the image data stream, ensuring the integrity and reconstructability of the image data during wireless transmission. The row and column identifiers in the frame header are used to mark the position of the corresponding data frame in the complete image. The wireless receiver can reassemble out-of-order received data frames according to the row and column identifiers to reconstruct a complete two-dimensional image matrix.

[0041] refer to Figure 2 and Figure 3In this embodiment, a wireless receiver wirelessly receives the X-ray image data stream of the tension clamp. A channel state monitor extracts the signal-to-noise ratio and multipath delay spread parameters of the wireless transmission link in real time. A frequency-domain adaptive filter dynamically generates a notch filter bank based on channel parameters and performs frequency-domain adaptive filtering of the image data stream. A physical feature-guided recognition processor generates a X-ray attenuation physical feature-guided weight map based on the X-ray linear attenuation coefficient equation, and the recognition network is guided by the weight map to identify pores and inclusions. The technical solution of this embodiment introduces the physical layer parameters of wireless communication and the physical laws of X-ray imaging into the image processing flow, eliminates interference components introduced by the channel in the frequency domain, and guides the defect recognition process using physical attenuation features, thus eliminating the interference of channel noise and scattering artifacts on defect recognition.

[0042] In a preferred embodiment, the channel state monitor includes an orthogonal frequency division multiplexing (OFDM) signal parsing component and a delay spread calculation component. The OFDM signal parsing component is connected to the baseband processing unit of the wireless receiver, acquires the physical layer baseband signal corresponding to the image data stream received by the wireless receiver, and processes the channel estimation pilot symbols contained in the physical layer preamble of the baseband signal. The physical layer preamble is located at the beginning of each data frame and contains a fixed-length channel estimation pilot symbol. The pilot symbol is generated using a known pseudo-random sequence and is used by the receiver to complete the channel estimation processing.

[0043] The Orthogonal Frequency Division Multiplexing (OFDM) signal analysis component first performs synchronization processing on the time-domain baseband signal of the physical layer preamble to extract the complete channel estimation pilot symbol time-domain sequence. Then, it performs a Fast Fourier Transform (FFT) on the pilot symbol time-domain sequence to convert it to the frequency domain, obtaining the frequency-domain received symbol corresponding to each subcarrier. The OFDM signal analysis component divides the frequency-domain received symbol of each subcarrier with the locally stored known pilot symbols to obtain the channel frequency-domain response value for each subcarrier. The channel frequency-domain response value is in complex form, containing real and imaginary parts, representing the amplitude and phase responses of the subcarrier channel, respectively. The OFDM system used in the wireless communication link includes 128 effective subcarriers with a subcarrier spacing of 15kHz. The OFDM signal analysis component outputs the channel frequency-domain response values ​​corresponding to the 128 effective subcarriers.

[0044] The delay spread calculation component is connected to the orthogonal frequency division multiplexing (OFDM) signal analysis component to obtain the channel frequency domain response values ​​corresponding to 128 effective subcarriers. The channel frequency domain response value sequence is zero-padding to extend the sequence length to 512 points. Then, an inverse fast Fourier transform is performed on the zero-padding channel frequency domain response value sequence to convert the frequency domain sequence into a time domain sequence, obtaining the channel's time domain impulse response sequence. The corresponding calculation expression is:

[0045]

[0046] in, The time-domain impulse response sequence of the channel. For the arrival time of the multipath components, For the first Channel frequency domain response values ​​corresponding to each subcarrier For the first The center frequency of each subcarrier The number of points in the inverse fast Fourier transform. This is the inverse fast Fourier transform operation.

[0047] The time delay spread calculation component squares the modulus of each sampling point in the time-domain impulse response sequence to obtain the time-domain power delay distribution spectrum. The corresponding calculation expression is:

[0048]

[0049] in, Arrival time The power value of the corresponding time-domain power delay distribution spectrum.

[0050] The time delay spread calculation component has a preset noise threshold, which is set based on the noise floor power value of the time-domain power delay distribution spectrum. The noise floor power value is the average power value of the time interval in the time-domain power delay distribution spectrum where no multipath components arrive. The time delay spread calculation component extracts the multipath components corresponding to the sampling points in the time-domain power delay distribution spectrum whose power values ​​exceed the preset noise threshold, as valid multipath components, and records the arrival time and corresponding power value of each valid multipath component. Based on the arrival time and power value of all valid multipath components, the time delay spread calculation component calculates the multipath time delay spread parameters.

[0051] The Orthogonal Frequency Division Multiplexing (OFDM) signal analysis component sums the squares of the real and imaginary parts of the channel frequency domain response value for each subcarrier to obtain the square of the channel amplitude gain for that subcarrier. Then, it averages the squares of the channel amplitude gains for all effective subcarriers to obtain the average signal power. The OFDM signal analysis component calculates the power of the received signal corresponding to the idle subcarrier in the physical layer preamble to obtain the average noise power. Idle subcarriers are subcarriers that do not carry data or pilot symbols; their received signals only contain channel noise. The OFDM signal analysis component uses the ratio of the average signal power to the average noise power as the signal-to-noise ratio (SNR) of the wireless transmission link. The corresponding expression is:

[0052]

[0053] in, The signal-to-noise ratio of the wireless transmission link. To perform operations on the real part of a complex number, To perform operations on the imaginary part of a complex number, This represents the average noise power value corresponding to the idle subcarrier. This represents the number of effective subcarriers.

[0054] The delay spread calculation component includes a Hanning window function applicator and a moment calculator. The Hanning window function applicator has a preset Hanning window function of fixed length. The time window length of the Hanning window function is set according to the preset maximum multipath delay of the wireless communication link, and is set to 20μs, corresponding to a 512-point time-domain sampling sequence. The Hanning window function applicator performs a time-domain multiplication operation on the Hanning window function and the time-domain power delay distribution spectrum, and performs windowing processing on the time-domain power delay distribution spectrum to extract the multipath components within the preset time window, suppressing the influence of noise and interference components outside the time window on the multipath component extraction. The expression of the Hanning window function is:

[0055]

[0056] in, The first of the Hanning window functions The values ​​of each sampling point The length of the Hanning window function corresponds to the number of sampling points within the preset time window.

[0057] The moment calculator connects to the Hanning window function applicator to obtain the windowed time-domain power delay distribution spectrum. It then normalizes the spectrum by summing the power values ​​of all effective multipath components to 1, obtaining the power probability distribution of each multipath component. The moment calculator performs first-order moment calculations on the normalized time-domain power delay distribution spectrum to obtain the average multipath arrival time, which is the weighted average of the arrival times of all effective multipath components, weighted by their power probabilities. The moment calculator then performs second-order central moment calculations on the normalized time-domain power delay distribution spectrum, calculating the weighted variance of the arrival times of all effective multipath components centered on the average multipath arrival time, obtaining the multipath delay variance. The delay spread calculation component uses the square root of the multipath delay variance as the multipath delay spread parameter; the corresponding expression is:

[0058]

[0059]

[0060] in, The average multipath arrival time, For multipath delay spread parameters, The number of effective multipath components, For the first The normalized power values ​​of the effective multipath components, For the first The arrival time of each effective multipath component is calculated. The delay spread calculation component outputs the calculated multipath delay spread parameters to the frequency domain adaptive filter.

[0061] Table 2. Calculation parameters for subcarrier channel frequency domain response and signal-to-noise ratio.

[0062] Number of effective subcarriers 128 The total number of subcarriers carrying pilots and data in an orthogonal frequency division multiplexing (OFDM) system Subcarrier spacing 15kHz Frequency spacing between adjacent subcarriers Pilot symbol sequence length 128 points Time-domain sequence length of channel estimation pilot symbols Fast Fourier Transform Points 128 points Number of Fast Fourier Transform Operation Points for Pilot Symbol Time-Domain Sequences Inverse Fast Fourier Transform Points 512 points Number of operation points for the inverse fast Fourier transform of the channel frequency domain response sequence Hanning window function length 512 points The number of sampling points of the Hanning window used in windowing processing of the time-domain power delay distribution spectrum Noise threshold setting factor 3 The noise threshold is 3 times the average power of the noise floor.

[0063] Specifically, the parameters defined in Table 2 are used to standardize the computation process of the orthogonal frequency division multiplexing (OFDM) signal analysis component and the delay spread calculation component, ensuring the accuracy of the channel frequency domain response, signal-to-noise ratio (SNR), and multipath delay spread parameters. The settings for the number of effective subcarriers and subcarrier spacing match the wireless channel characteristics of the 2.4 GHz ISM band. The expanded setting of the inverse fast Fourier transform points improves the temporal resolution of the time-domain power delay distribution spectrum. The Hanning window function effectively suppresses the impact of spectral leakage on multipath component extraction. The noise threshold setting coefficient balances the integrity of multipath component extraction with noise suppression capability.

[0064] In this embodiment, the orthogonal frequency division multiplexing (OFDM) signal analysis component performs a fast Fourier transform on the physical layer preamble to obtain the subcarrier channel frequency domain response value, and then calculates the signal-to-noise ratio (SNR) of the wireless transmission link. The delay spread calculation component performs an inverse fast Fourier transform on the channel frequency domain response value to obtain the time-domain power delay distribution spectrum. Combined with a Hanning window function applicator, the effective multipath components are truncated. A moment calculator is used to calculate the first-order moment and the second-order central moment to obtain the multipath delay spread parameters. This embodiment improves the accuracy of multipath delay spread parameter and SNR calculations, providing precise channel state parameter inputs for the dynamic filtering processing of frequency domain adaptive filters.

[0065] In another preferred embodiment, the frequency-domain adaptive filter includes a spectrum energy mapping component and a notch filtering component. The spectrum energy mapping component is connected to a channel state monitor to acquire the signal-to-noise ratio (SNR) and multipath delay spread parameters output by the channel state monitor. The spectrum energy mapping component first converts the SNR into a noise floor power spectral density (HPS), which characterizes the noise power distribution density of the wireless channel in the frequency domain, with units of W / Hz. Based on the linear ratio of the SNR, combined with the system bandwidth of the wireless communication link and the total power of the received signal, the spectrum energy mapping component calculates the average noise floor power over the entire system bandwidth. Then, it divides the average noise floor power by the system bandwidth to obtain a flat noise floor power spectral density. The noise floor power spectral density is a constant value over the entire system bandwidth, corresponding to the power spectral distribution characteristics of additive white Gaussian noise. The corresponding calculation expression is:

[0066]

[0067] in, The power spectral density of the noise floor. This represents the total average power of the signal received by the wireless receiver. This refers to the system bandwidth of the wireless communication link.

[0068] The spectrum energy mapping component converts the multipath delay spread parameter into a frequency-selective fading depth parameter. This parameter characterizes the correlation bandwidth and fading depth of the frequency-selective fading caused by multipath effects in the wireless channel. According to wireless communication principles, the channel's correlation bandwidth is inversely proportional to the multipath delay spread parameter. The expression for the correlation bandwidth is:

[0069]

[0070] in, The relevant bandwidth of the channel, This is the multipath delay spread parameter. The spectral energy mapping component uses the reciprocal of the relevant bandwidth as the frequency-selective fading depth parameter. The larger the frequency-selective fading depth parameter, the stronger the frequency selectivity of the corresponding channel, and the more severe the fading fluctuations in the frequency domain. The corresponding calculation expression is:

[0071]

[0072] in, This is the frequency-selective fading depth parameter.

[0073] Since the noise floor power spectral density and the frequency selective fading depth parameter have different dimensions, the spectrum energy mapping component cannot be directly superimposed in the frequency domain. The spectrum energy mapping component first performs a unified dimension normalization process on the two parameters: calculate the maximum value of the noise floor power spectral density and the frequency selective fading depth parameter in their respective preset value ranges, and divide the two parameters by the corresponding maximum value to obtain normalized scalar parameters with values ​​ranging from 0 to 1.

[0074] The normalized noise floor power spectral density is used as the flat distribution basis across the entire bandwidth. The normalized frequency-selective fading depth parameter is used as the peak value at the center frequency to generate a Gaussian fading distribution. The flat distribution and the Gaussian distribution are superimposed to generate the total interference energy spectrum of the wireless channel in the frequency domain. The spectrum energy mapping component has a preset interference energy threshold. Continuous frequency bands in the total interference energy spectrum whose energy values ​​exceed the interference energy threshold are marked as interference energy distribution intervals. Each interference energy distribution interval includes parameters such as the interval's start frequency, end frequency, center frequency, and bandwidth. Based on the frequency domain distribution of the total interference energy spectrum, the spectrum energy mapping component can generate multiple discontinuous interference energy distribution intervals, each interval corresponding to a high-interference sub-band in the frequency domain.

[0075] The notch filter execution component connects to the spectrum energy mapping component and the wireless receiver to acquire the interference energy distribution range output by the spectrum energy mapping component and the image data stream output by the wireless receiver. The notch filter execution component first performs a two-dimensional Fourier transform on each frame of the two-dimensional image data stream to obtain the image's two-dimensional spectrum data. The horizontal axis of the two-dimensional spectrum data represents the horizontal spatial frequency, and the vertical axis represents the vertical spatial frequency. Each coordinate point corresponds to a complex value of a spatial frequency component, containing amplitude and phase information. The notch filter execution component converts the frequency parameters of the interference energy distribution range into spatial frequency coordinates in the two-dimensional spectrum, locating the set of frequency coordinates corresponding to the interference energy distribution range. This set of frequency coordinates contains all spatial frequency coordinate points in the two-dimensional spectrum that fall within the interference energy distribution range.

[0076] The notch filter execution component constructs a corresponding second-order infinite impulse response (IRR) notch filter based on the center frequency and bandwidth of each interference sub-band corresponding to the frequency coordinate set. Multiple IRR notch filters are combined to form a notch filter bank. The notch filter execution component uses the notch filter bank to filter the two-dimensional spectral data of the image, setting the amplitude of the spectral components within the frequency coordinate set to zero while preserving the phase information. After filtering, the notch filter execution component performs an inverse two-dimensional Fourier transform on the processed two-dimensional spectral data, converting the spectral data from the frequency domain back to the spatial domain, obtaining two-dimensional filtered image data. The notch filter execution component then outputs the filtered image data to the physical feature-guided recognition processor.

[0077] The notch filter execution component includes a pole arbiter and a quality factor calculator. The pole arbiter is used to configure the poles of a second-order infinite impulse response (IRR) notch filter. The system function of the IRR is designed based on the pole-zero configuration of the complex frequency plane. The zeros of the notch filter are set on the unit circle of the complex frequency plane, corresponding to the notch center frequency, to achieve complete notching of the corresponding frequency components. The pole arbiter sets conjugate pole pairs corresponding to the zeros inside the unit circle of the complex frequency plane according to the center frequency of the interference sub-band. The argument of the conjugate pole pairs is consistent with the argument of the zeros, ensuring that the phase characteristics of the notch filter remain linear within the passband.

[0078] The quality factor calculator connects to the spectrum energy mapping component to obtain the center frequency and bandwidth of each interference energy distribution interval. The ratio of the center frequency to the bandwidth is used as the quality factor of the corresponding second-order infinite impulse response notch filter. The quality factor characterizes the frequency selectivity of the notch filter; the higher the quality factor, the narrower the notch bandwidth and the stronger the frequency selectivity. The corresponding calculation expression is:

[0079]

[0080] in, The quality factor of a second-order infinite impulse response notch filter. The center frequency of the notch filter. This represents the bandwidth of the notch filter.

[0081] The pole configurator adjusts the polar radius distance of the conjugate pole pair from the unit circle in the complex frequency plane based on the quality factor. The polar radius distance ranges from 0 to 1. The closer the polar radius distance is to 1, the higher the quality factor of the notch filter, the narrower the notch bandwidth, and the greater the notch depth. The expression for calculating the polar radius distance is:

[0082]

[0083] in, The distance between the polar radii of the conjugate pole pair is denoted as . The sampling frequency of the image data is denoted as . The pole locator configures conjugate pole pairs based on the pole radius distance and the center frequency. Combined with the zero configuration, it generates the system function of a second-order infinite impulse response notch filter, with the corresponding expression being:

[0084]

[0085] in, This is the system function of a second-order infinite impulse response notch filter. Let z be the z-transform variable of the complex frequency plane. This is the digital angular frequency corresponding to the center frequency of the notch filter. is the polar radius distance between the conjugate pole pairs.

[0086] The notch filter execution component combines second-order infinite impulse response notch filters corresponding to each interference energy distribution interval in a cascade manner to form a complete notch filter bank. The system function of the cascaded notch filter bank is the product of the system functions of all individual second-order infinite impulse response notch filters, and can simultaneously notch filter the spectral components in multiple interference sub-bands.

[0087] Table 3. Parameter Configuration Table for Notch Filter Bank

[0088] Number of points in a two-dimensional Fourier transform 3072×2048 The number of horizontal and vertical points in the two-dimensional Fourier transform of image data is consistent with the image resolution. Maximum number of second-order IIR notch filters cascaded 16 Maximum number of second-order IIR notch filters cascaded by the notch filter bank Notch filter center frequency resolution 1Hz Minimum step value for setting the center frequency of the notch filter Quality factor range 10-1000 The settable range of the quality factor for a second-order IIR notch filter. Polar radius distance range 0.9-0.999 The configurable range of the distance between conjugate poles and pole radius. Interference energy threshold setting coefficient 2 The interference energy threshold is twice the average interference energy.

[0089] Specifically, the parameters defined in Table 3 are used to standardize the design and implementation process of notch filter banks, ensuring the accuracy and efficiency of frequency domain adaptive filtering. The number of two-dimensional Fourier transform points is kept consistent with the image resolution, avoiding the loss of spectral resolution; the maximum number of second-order infinite impulse response notch filters cascades matches the maximum number of interference sub-bands in complex electromagnetic environments; the range of quality factor and polar radius distance covers the filtering requirements of all scenarios from broadband interference to narrowband interference; the setting of the interference energy threshold coefficient balances the completeness of interference filtering and the degree of preservation of effective image information.

[0090] In this embodiment, the signal-to-noise ratio and multipath delay spread parameters are converted into noise floor power spectral density and frequency-selective fading depth parameters through a spectral energy mapping component, thereby generating an interference energy distribution range. A notch filter execution component locates the set of frequency coordinates corresponding to the interference energy distribution range in the two-dimensional spectrum. A second-order infinite impulse response (IRR) notch filter is designed using a pole locator and a quality factor calculator. Multiple IRR notch filters are cascaded to form a notch filter bank. After filtering the image spectrum, the filtered image data is obtained through an inverse two-dimensional Fourier transform. This embodiment achieves frequency-domain adaptive filtering based on channel state, which can selectively filter out channel interference overlapping with the image's inherent frequencies while preserving the image's effective texture structure.

[0091] refer to Figure 4 In another preferred embodiment, the physical feature-guided recognition processor includes a material attenuation modeling component and a weighted graph generation component. The material attenuation modeling component pre-stores standard structural parameters of the tension clamp, including the nominal thickness of the aluminum tube, the nominal diameter of the internal steel core, and the coaxiality tolerance between the aluminum tube and the steel core. It also pre-stores a table of linear attenuation coefficients of X-rays in aluminum and steel materials under different tube voltages. The material attenuation modeling component obtains the tube voltage and tube current parameters of the X-ray source corresponding to this flaw detection operation. Based on the tube voltage parameters, it retrieves the corresponding linear attenuation coefficients for aluminum and steel materials from the table. Simultaneously, it obtains the initial X-ray intensity parameters of the X-ray source, which are calculated based on the tube current and exposure time of the X-ray source.

[0092] The material attenuation modeling component establishes a two-dimensional cross-sectional geometric model of the tension clamp based on its structural parameters. The geometric model defines the coordinate ranges of the aluminum tube region, the steel core region, and the aluminum-steel interface region. For any pixel coordinate in the image... The material attenuation modeling component calculates the thickness of the aluminum tube penetrated by X-rays at that coordinate based on the geometric model. With the thickness of the penetrating steel core By combining the initial radiation intensity and the linear attenuation coefficients of aluminum and steel materials, a linear attenuation coefficient equation is constructed. The expression of the linear attenuation coefficient equation is consistent with the expression in the previous embodiment, and the meaning of the parameters remains the same.

[0093] The weighted graph generation component is connected to the material attenuation modeling component and the frequency domain adaptive filter to obtain the linear attenuation coefficient equation constructed by the material attenuation modeling component and the filtered image data output by the frequency domain adaptive filter. The weighted graph generation component first processes the coordinates of each pixel in the filtered image data. Corresponding actual grayscale value Converted to actual radiation intensity The transformation relationship is converted into a linear mapping relationship, and the expression is:

[0094]

[0095] in, and The linear mapping coefficients are preset based on the detector's calibration parameters. The weight map generation component assigns each pixel coordinate... Substituting into the linear attenuation coefficient equation, the theoretical ray intensity corresponding to this coordinate is obtained. Then convert the theoretical ray intensity into theoretical grayscale value. The weighted image generation component calculates the absolute value of the difference between the theoretical grayscale value and the actual grayscale value, using it as the initial weight corresponding to that pixel coordinate. The corresponding calculation expression is:

[0096]

[0097] in, pixel coordinates The corresponding initial weights.

[0098] The weight map generation component normalizes the initial weights corresponding to all pixels in the entire image, mapping the values ​​of the initial weights to between 0 and 1. The normalization formula is as follows:

[0099]

[0100] in, pixel coordinates The corresponding normalized weight values, This represents the minimum initial weight for the entire image. This represents the maximum initial weight for the entire image. The weight map generation component combines the normalized weight values ​​corresponding to all pixels into a two-dimensional matrix, generating a ray attenuation physical feature-guided weight map with the same size as the filtered image data.

[0101] The weighted map generation component includes a boundary fitter and a smooth transition processor. The boundary fitter, based on the aluminum tube thickness and internal steel core diameter parameters of the tension clamp, combined with the grayscale distribution characteristics of the filtered image data, delineates the aluminum tube region, steel core region, and aluminum-steel interface region in the image. The boundary fitter first performs edge detection processing on the filtered image data, extracting the inner and outer edge contours of the aluminum tube and the outer edge contour of the steel core. Then, it fits and corrects the edge contours using nominal structural parameters to obtain accurate region boundaries. The aluminum-steel interface region is a ring-shaped area between the inner edge of the aluminum tube and the outer edge of the steel core, with a width preset to 5 pixels according to the nominal structural parameters.

[0102] The boundary fitter applies a geometric transition compensation coefficient to the pixel coordinates within the aluminum-steel interface region. This coefficient compensates for the X-ray intensity attenuation deviation caused by X-ray scattering in the interface. The value of the geometric transition compensation coefficient is related to the distance from the pixel coordinate to the aluminum-steel boundary; the closer to the boundary, the larger the value. The boundary fitter then substitutes the geometric transition compensation coefficient into the linear attenuation coefficient equation to correct the calculated theoretical X-ray intensity. The updated linear attenuation coefficient equation is as follows:

[0103]

[0104] in, pixel coordinates The corresponding geometric transition compensation coefficient ranges from 0.8 to 1.2.

[0105] The smooth transition processor is connected to the boundary fitter to obtain the edge pixel coordinates of the aluminum-steel interface region. The initial weights at these edge pixel coordinates are then spatially convolved using a Gaussian kernel function to eliminate abrupt weight changes at the aluminum-steel interface region, achieving a smooth weight transition. The expression for the Gaussian kernel function is:

[0106]

[0107] in, The Gaussian kernel function in coordinates The value at that location, The standard deviation of the Gaussian kernel function is set to 1.5. The smooth transition processor uses a 3×3 Gaussian kernel to perform convolution operations on the initial weights of the aluminum-steel interface region. The initial weights after convolution are then input into the subsequent normalization process to generate a ray attenuation physical feature-guided weight map with smooth edges.

[0108] refer to Figure 5 The physical feature-guided recognition processor includes a feature extraction backbone network and a weight fusion component. The feature extraction backbone network adopts an encoder structure, containing multiple convolutional and pooling layers connected in series. Specifically, the backbone network contains 5 convolutional blocks, each containing two 3×3 convolutional layers, one batch normalization layer, and one ReLU activation function layer. Each convolutional block ends with a 2×2 max-pooling layer for downsampling the feature map. The feature extraction backbone network performs layer-by-layer convolution and pooling processing on the input filtered image data. Each convolutional block outputs a convolutional feature map at a corresponding scale, ultimately outputting 5 multi-scale convolutional feature maps at different scales. The sizes of the multi-scale convolutional feature maps are 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image, with 64, 128, 256, 512, and 1024 channels, respectively.

[0109] The weight fusion component is connected to the weight map generation component and the feature extraction backbone network to obtain the ray attenuation physical feature-guided weight map and the multi-scale convolutional feature map. The weight fusion component performs bilinear interpolation on the ray attenuation physical feature-guided weight map, scaling the weight map to the same size as the convolutional feature map at each scale, resulting in a spatial attention matrix corresponding to the multi-scale convolutional feature map. Each element of the spatial attention matrix has a value between 0 and 1, corresponding to each spatial position of the convolutional feature map at the corresponding scale.

[0110] The weighted fusion component performs element-wise multiplication of the spatial attention matrix at each scale with the corresponding convolutional feature map. This weights the feature channels at each spatial location of the convolutional feature map, amplifying the feature responses of regions with higher weights and suppressing those of regions with lower weights, outputting a weighted feature map for each scale. The expression for the element-wise multiplication operation is:

[0111]

[0112] in, For the first Weighted feature maps at various scales in coordinates eigenvalues ​​at that location For the first Convolutional feature maps at various scales in coordinates eigenvalues ​​at that location For the first Spatial attention matrix at each scale in coordinate The value at that location.

[0113] refer to Figure 6 The physical feature-guided recognition processor includes a classification head component connected to a weighted fusion component. This component acquires weighted feature maps at all scales, performs global average pooling on each scale's weighted feature map to obtain a one-dimensional feature vector for each scale, and concatenates the feature vectors from all scales along their channel dimensions to obtain a one-dimensional feature vector that fuses multi-scale information. The classification head component contains two fully connected layers. The first fully connected layer maps the concatenated feature vector to a 256-dimensional feature space, and the second fully connected layer maps the 256-dimensional feature vector to a 2-dimensional output space, corresponding to the classification probabilities of porosity defects and inclusion defects, respectively. The final output is the recognition results for porosity and inclusion defects, including the defect category, confidence level, location coordinates, and contour range.

[0114] The classification head component includes a receptive field pyramid pooler and a focus loss calculator. The classification head component operates in both offline training and online flaw detection modes.

[0115] In online flaw detection mode, the receptive field pyramid pooler performs adaptive average pooling at different scales on the highest-scale weighted feature map output by the weight fusion component. The pooling scales are 1×1, 2×2, 3×3, and 6×6, respectively, resulting in pooled feature maps of the corresponding scales. Each pooled feature map is flattened to obtain a one-dimensional feature vector of the corresponding scale. The feature vectors of all scales are concatenated to obtain a multi-scale pooled feature vector. The receptive field pyramid pooler concatenates the multi-scale pooled feature vector with the global average pooled feature vector of the aforementioned multi-scale convolutional feature map and inputs it into the subsequent fully connected mapping layer, thereby improving the feature's ability to represent defects of different sizes.

[0116] In offline training mode, the receptive field pyramid pooler concatenates the feature vectors obtained from pooling at different scales along the channel dimension. The concatenated feature vector has a dimension of 2048. The concatenated feature vector is then input into a fully connected mapping layer. The first fully connected layer maps the 2048-dimensional feature vector to 256 dimensions, and the second fully connected layer maps the 256-dimensional feature vector to a 2-dimensional classification output space, outputting the defect classification probability for each pixel.

[0117] The focus loss calculator is used to calculate the loss value during the network training phase and update the network parameters. During the online flaw detection phase, the focus loss calculator is not loaded; only the forward inference process is executed. The focus loss calculator obtains the label matrix of the porosity and inclusion defect identification results corresponding to the training samples. The label matrix is ​​a binary matrix; positive samples correspond to defect areas and have a value of 1, while negative samples correspond to background areas and have a value of 0. The focus loss calculator first calculates the binary cross-entropy loss based on the label matrix and the output vector of the fully connected mapping layer. Then, it modulates the cross-entropy loss using an exponential modulation factor, reducing the loss weight of easily classified negative samples and increasing the loss weight of difficult-to-classify positive samples, thus solving the problem of imbalanced positive and negative samples in defect detection scenarios. The expression for the focus loss is:

[0118]

[0119] in, For the focal loss value, This is the balancing weight factor for positive and negative samples, set to a value of 0.25. This represents the model's predicted probability for the sample. The focus parameter is set to 2, which adjusts the weight increase for difficult-to-classify samples. The focus loss calculator transmits the calculated loss value to each convolutional layer of the feature extraction backbone network through the backpropagation algorithm, and updates the convolutional kernel parameters of the backbone network according to the gradient descent algorithm, thus completing the network training optimization.

[0120] Table 4 Parameter Configuration Table for Linear Attenuation Coefficient Equation

[0121] X-ray source tube voltage range 160kV-225kV The tube voltage of the X-ray source used in flaw detection operations can be set within a certain range. Linear attenuation coefficient of aluminum material (160kV) 0.75cm⁻¹ Linear attenuation coefficient of X-rays in aluminum at 160kV tube voltage Linear attenuation coefficient of steel material (160kV) 2.3cm⁻¹ Linear attenuation coefficient of X-rays in steel at 160kV tube voltage Linear attenuation coefficient of aluminum material (225kV) 0.55cm⁻¹ Linear attenuation coefficient of X-rays in aluminum material at 225kV tube voltage Linear attenuation coefficient of steel material (225kV) 1.8cm⁻¹ Linear attenuation coefficient of X-rays in steel under 225kV tube voltage Aluminum tube nominal thickness range 8mm-15mm The nominal thickness of tension clamp aluminum tubes can be set within a certain range. Nominal diameter range of steel core 10mm-30mm The nominal diameter of the internal steel core of the tension clamp can be set within a certain range. Range of values ​​for geometric transition compensation coefficient 0.8-1.2 The settable range of the geometric transition compensation coefficient in the aluminum-steel interface area Gaussian kernel standard deviation 1.5 The standard deviation of the Gaussian kernel function used in weighted smoothing

[0122] Specifically, the parameters defined in Table 4 are used to standardize the construction of the linear attenuation coefficient equation and the generation process of the weight map, ensuring the accuracy of the weight map guided by the physical characteristics of X-ray attenuation. The values ​​of the linear attenuation coefficient under different tube voltages conform to the physical laws of X-ray-matter interaction. The structural parameters of the aluminum tube and steel core cover the various types of tension clamps commonly used in transmission lines. The setting of the geometric transition compensation coefficient and Gaussian kernel function effectively compensates for the scattering effect at the aluminum-steel interface, eliminates abrupt weight changes, and improves the smoothness and accuracy of the weight map.

[0123] In this embodiment, a linear attenuation coefficient equation is constructed by combining the material attenuation modeling component with the tension clamp structural parameters and the X-ray linear attenuation coefficient. A weight map generation component calculates the initial weight of each pixel and performs normalization to generate a weight map guided by the physical features of X-ray attenuation. A boundary fitter and a smooth transition processor complete the weight correction and smoothing of the aluminum-steel interface region. A feature extraction backbone network extracts multi-scale convolutional feature maps from the filtered image data. A weight fusion component fuses the spatial attention matrix with the convolutional feature maps to obtain a weighted feature map. A classification head component combined with a receptive field pyramid pooler completes the fusion and classification of multi-scale features. A focus loss calculator completes the network training and optimization. This embodiment utilizes the physical attenuation law of X-rays to guide the defect identification process, amplifying the feature response of abnormally low grayscale areas conforming to the attenuation law, suppressing grayscale fluctuations caused by edge scattering, and improving the completeness and accuracy of feature representation of micropores and inclusion defects.

Claims

1. A wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps, characterized in that, The system includes a wireless receiver, a channel state monitor, a frequency-domain adaptive filter, and a physical feature-guided identification processor. A flat panel detector acquires X-ray images of tension cable clamps and outputs an image data stream. The wireless receiver is communicatively connected to the flat panel detector to receive the image data stream. The channel state monitor is connected to the wireless receiver to extract the signal-to-noise ratio (SNR) and multipath delay spread parameters of the wireless transmission link. The frequency-domain adaptive filter is connected to both the channel state monitor and the wireless receiver. Based on the SNR and multipath delay spread parameters, the frequency-domain adaptive filter calculates and dynamically generates the interference energy distribution range of the wireless channel in the frequency domain. The system comprises a notch filter bank, a frequency-domain adaptive filter, and a physical feature-guided recognition processor connected to the frequency-domain adaptive filter. The physical feature-guided recognition processor extracts the grayscale distribution features of the filtered image data and calculates a radiation attenuation physical feature-guided weight map by combining the linear attenuation coefficient equation of X-rays in aluminum and steel materials. The physical feature-guided recognition processor inputs the filtered image data and the radiation attenuation physical feature-guided weight map into a recognition network to output the identification results of pores and inclusion defects.

2. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 1, characterized in that, The channel state monitor includes an orthogonal frequency division multiplexing (OFDM) signal analysis component and a delay spread calculation component. The OFDM signal analysis component performs a fast Fourier transform on the physical layer preamble of the image data stream received by the wireless receiver to obtain multiple subcarrier channel frequency domain response values. The delay spread calculation component performs an inverse fast Fourier transform on the multiple subcarrier channel frequency domain response values ​​to obtain a time-domain power delay distribution spectrum. The delay spread calculation component extracts multipath components exceeding a preset noise threshold from the time-domain power delay distribution spectrum and calculates the multipath delay spread parameter based on the arrival time and power value of the multipath components. The OFDM signal analysis component calculates the signal-to-noise ratio (SNR) based on the sum of the squares of the real and imaginary parts of the multiple subcarrier channel frequency domain response values.

3. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 1, characterized in that, The frequency domain adaptive filter includes a spectrum energy mapping component and a notch filtering component. The spectrum energy mapping component converts the signal-to-noise ratio into a noise floor power spectral density and converts the multipath delay spread parameter into a frequency selective fading depth parameter. The spectrum energy mapping component calculates the maximum values ​​of the noise floor power spectral density and the frequency selective fading depth parameter within their respective ranges. The two parameters are divided by their corresponding maximum values ​​to complete the unified dimension normalization process. The normalized noise floor power spectral density is used as the flat distribution basis of the entire bandwidth. The normalized frequency selective fading depth parameter is used as the peak value at the center frequency to generate a Gaussian fading distribution. The flat distribution and the Gaussian distribution are superimposed to generate the total interference energy spectrum. The spectrum energy mapping component adds the noise floor power spectral density to the frequency selective fading depth parameter to generate the interference energy distribution interval. The notch filter execution component locates the set of frequency coordinates corresponding to the interference energy distribution interval in the two-dimensional Fourier transform spectrum of the image data stream. The conversion scaling factor is the ratio of the maximum value of the horizontal spatial frequency of the image to the bandwidth of the wireless channel. The start frequency and end frequency of the interference energy distribution interval are multiplied by this ratio to obtain the corresponding horizontal spatial frequency range. The vertical spatial frequency range takes the same value interval as the horizontal spatial frequency range. All coordinate points in the two-dimensional spectrum that fall within the rectangular spatial frequency region enclosed on the two-dimensional Fourier transform spectrum with the calculated horizontal spatial frequency range as the horizontal boundary and the vertical spatial frequency range with the same value as the horizontal range as the vertical boundary are marked as the set of frequency coordinates.

4. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 1, characterized in that, The physical feature-guided recognition processor includes a material attenuation modeling component and a weight map generation component. The material attenuation modeling component acquires the aluminum tube thickness and internal steel core diameter parameters of the tension clamp and sets the initial radiation intensity of the X-ray source. The material attenuation modeling component completes imaging geometry calibration through a standard calibration plate, establishes a one-to-one correspondence between pixel coordinates and physical space coordinates, and determines the radiation projection path as a parallel ray perpendicular to the detector plane based on the relative position of the X-ray source and the detector. The material attenuation modeling component constructs the linear attenuation coefficient equation based on the initial radiation intensity, the linear attenuation coefficient of aluminum material corresponding to the aluminum tube thickness, and the linear attenuation coefficient of steel material corresponding to the internal steel core diameter. The weight map generation component calculates the thickness of the radiation penetrating the aluminum tube and steel core based on the physical space coordinates corresponding to the pixel coordinates, substitutes the calculated thickness value into the linear attenuation coefficient equation to solve for the theoretical gray value, and uses the absolute value of the difference between the theoretical gray value and the actual gray value as the initial weight. The weight map generation component normalizes the initial weight to generate the radiation attenuation physical feature-guided weight map.

5. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 1, characterized in that, The physical feature-guided recognition processor includes a feature extraction backbone network and a weight fusion component. The feature extraction backbone network contains multiple convolutional layers and pooling layers connected in series to downsample the filtered image data and output a multi-scale convolutional feature map. The weight fusion component performs bilinear interpolation on the ray attenuation physical feature-guided weight map to obtain a spatial attention matrix with the same size as the multi-scale convolutional feature map. The weight fusion component multiplies the spatial attention matrix element-wise with the multi-scale convolutional feature map to output a weighted feature map. The physical feature-guided recognition processor includes a classification head component. The classification head component performs global average pooling and fully connected mapping on the weighted feature map to output the identification results of pores and inclusion defects.

6. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 1, characterized in that, The wireless receiver includes a frame header parsing component and a data reassembly component. The frame header parsing component extracts the data frame header containing row number identifiers and column number identifiers from the image data stream. The data reassembly component reassembles the out-of-order received data frames into the image data stream in the form of a two-dimensional matrix according to the row number identifiers and column number identifiers. The channel state monitor synchronously reads the channel estimation pilot symbols in the physical layer link control frame when the frame header parsing component extracts the data frame header. The frequency domain adaptive filter directly performs frequency domain adaptive filtering operation on the data in the row buffer queue of the image data stream in the form of the two-dimensional matrix output by the data reassembly component.

7. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 2, characterized in that, The time delay spread calculation component includes a Hanning window function applicator and a moment calculator. The Hanning window function applicator multiplies a fixed-length Hanning window function with the time-domain power delay distribution spectrum in the time domain to truncate the multipath components within a preset time window. The moment calculator performs first-order moment calculation on the truncated time-domain power delay distribution spectrum to obtain the average multipath arrival time. The moment calculator performs second-order central moment calculation on the truncated time-domain power delay distribution spectrum to obtain the multipath delay variance. The time delay spread calculation component outputs the square root of the multipath delay variance as the multipath delay spread parameter to the frequency domain adaptive filter.

8. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 3, characterized in that, The notch filter execution component includes a pole arbiter and a quality factor calculator. The pole arbiter sets conjugate pole pairs inside the unit circle of the complex frequency plane according to the center frequency in the frequency coordinate set. The quality factor calculator calculates the bandwidth of the interference energy distribution range. The quality factor calculator uses the ratio of the center frequency to the bandwidth as the quality factor of the second-order IIR notch filter. The pole arbiter adjusts the polar radius distance of the conjugate pole pairs from the unit circle of the complex frequency plane according to the quality factor to generate the second-order IIR notch filter with a specified notch depth. The notch filter execution component combines the multiple second-order IIR notch filters in a cascaded manner to form the notch filter bank.

9. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 4, characterized in that, The weight map generation component includes a boundary fitter and a smooth transition processor. The boundary fitter delineates the aluminum tube region, the steel core region, and the aluminum-steel interface region in the filtered image data based on the aluminum tube thickness and the internal steel core diameter parameters. The boundary fitter adds a geometric transition compensation coefficient to the pixel coordinates within the aluminum-steel interface region to update the linear attenuation coefficient equation. The smooth transition processor performs spatial convolution operations on the initial weights at the edge pixel coordinates of the aluminum-steel interface region using a Gaussian kernel function. The smooth transition processor inputs the initial weights after the convolution operation into the normalization process to generate a ray attenuation physical feature guided weight map with smooth edges.

10. The wireless collaborative and intelligent identification system for X-ray flaw detection based on tension clamps according to claim 5, characterized in that, The classification head component includes a receptive field pyramid pooler and a focus loss calculator. The classification head component operates in offline training mode and online flaw detection mode. In the online flaw detection mode, the receptive field pyramid pooler performs adaptive average pooling at different scales on the weighted feature map to obtain multi-scale pooling feature vectors. The receptive field pyramid pooler concatenates the multi-scale pooling feature vectors by channel dimension and inputs them into a fully connected mapping layer to output the porosity and inclusion defect identification results. In offline training mode, the focus loss calculator obtains the label matrix of the porosity and inclusion defect identification results. The focus loss calculator calculates the cross-entropy loss based on the label matrix and the output vector of the fully connected mapping layer. The focus loss calculator generates a loss value based on the modulation factor of the exponential term of the cross-entropy loss and backpropagates it to the feature extraction backbone network to update the convolution kernel parameters of the feature extraction backbone network.

Images