A bolt axial force echo shape deformation compensation method based on a convolutional neural network

Abstract

The application discloses a bolt axial force echo deformation compensation method based on a convolutional neural network, relates to the technical fields of ultrasonic nondestructive testing and bolt axial force monitoring, and comprises the following steps: collecting bolt ultrasonic echoes under multi-frequency coded excitation to generate a monitoring data frame; extracting cross-band amplitude-phase differences and generating a path time-frequency tensor; generating coupling features, coupling transfer functions, coupling time delays, coupling reliabilities, aliasing residuals and deformation features through a double-flow convolutional network with echo path constraints; constructing a decoupling space and generating decoupled deformation features; performing aliasing residual deduction and regularized inverse filtering to generate compensation echoes and intrinsic time delays; obtaining bolt axial forces and axial force reliabilities through monotone constraint nonlinear acoustic-elastic inversion and bounded residual correction; determining coupling changes and axial force changes, updating or freezing a coupling baseline and outputting a state marker. The application reduces the interference of coupling state changes and aliasing echoes on axial force measurement, and improves the accuracy, stability and reliability of bolt axial force detection.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic nondestructive testing and bolt axial force monitoring technology, and in particular to a bolt axial force echo waveform compensation method based on convolutional neural networks. Background Technology

[0002] With the increasing demands for reliability in bolted connections in wind power equipment, bridge steel structures, rail transit, and large machinery, ultrasonic axial force testing technology based on the acoustoelastic effect has been applied. Existing technologies typically emit ultrasonic waves axially towards the bolt and determine the bolt axial force by comparing the echo propagation time before and after loading. For example, patent EP0589271A1 discloses a fastening detection method that places an ultrasonic sensor at the bolt head and compares the echo times under no-load and loaded conditions. Patent CN113155357A discloses a measurement method that calibrates the bolt's acoustic time variation and axial tensile force one by one.

[0003] However, in actual testing, the ultrasonic transducer needs to transmit sound to the bolt through a coupling layer. The transducer's clamping state, the thickness of the coupling layer, and temperature changes all alter the echo amplitude, phase, and propagation delay. Existing research shows that the effect of temperature on ultrasonic velocity is on the same order of magnitude as the change caused by axial stress, and the coupling layer thickness also affects the axial force estimation results. Furthermore, bolt sidewall reflections, bottom surface reflections, and multipath echoes are prone to aliasing. Existing processing methods based on a single peak value or acoustic time difference struggle to distinguish between coupling changes and actual axial force changes, easily misinterpreting coupling aging, probe reinstallation, or temperature drift as bolt loosening. Under complex reflection conditions, there are also problems such as unstable bottom surface echo positioning, continuous baseline drift, and a large workload for repeated on-site calibration, making it difficult to meet the requirements of long-term online monitoring for quantitative accuracy of axial force, state reliability, and adaptive compensation capabilities.

[0004] Therefore, how to provide a bolt axial force feedback waveform compensation method based on convolutional neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a bolt axial force echo waveform variation compensation method based on convolutional neural networks. This invention fully utilizes multi-frequency ultrasonic detection, complex continuous wavelet transform, dual-stream convolutional networks, feature decoupling, and monotonic acoustic elastic inversion techniques. It details the bolt echo path identification, coupling interference separation, aliasing residual compensation, body time delay extraction, and axial force inversion process, and has the advantages of strong anti-coupling interference capability, high axial force detection accuracy, high result reliability, and good long-term monitoring stability.

[0006] A bolt axial force feedback waveform compensation method based on a convolutional neural network according to an embodiment of the present invention includes the following steps: Step 1: Collect the ultrasonic echo of the bolt under multi-frequency coded excitation, and align and normalize it by combining bolt parameters, temperature and reference echo to generate monitoring data frames; Step 2: Perform matched filtering and complex continuous wavelet transform on the monitoring data frame to extract the cross-band amplitude and phase difference, and generate path time-frequency tensors according to the coupling interface, bottom surface and side wall aliasing echo regions; Step 3: Input the path time-frequency tensor into the echo path-constrained dual-stream convolutional network, generate coupling features, coupling transfer function, coupling delay and coupling confidence through the coupling state branch, and generate aliasing residuals and deformation features through the axial force deformation branch; Step 4: Based on the decoupling space constructed from coaxial force heterocoupled samples and cocoupled heteroaxial force samples, project the deformation features onto the orthogonal complement space of the coupled features to generate decoupling deformation features; Step 5: Subtract the aliasing residual from the bottom echo spectrum, perform regularized inverse filtering based on the coupling transfer function and coupling confidence level, and generate the compensated echo and body delay; Step 6: Input the body time delay, effective stress length, material type and temperature into the monotonic constrained nonlinear acoustoelastic inversion layer, and combine the decoupled deformation characteristics to generate a bounded residual correction amount, and obtain the bolt axial force and axial force confidence. Step 7: Determine the coupling change and axial force change based on the coupling interface echo, cross-band amplitude and phase difference, body time delay, and deviation between the compensated echo and the reference echo. Update or freeze the coupling baseline and output the bolt axial force and status mark.

[0007] Optionally, step one specifically includes: A longitudinal wave ultrasonic transducer is installed on the bolt end face, and a coupling layer is set between the longitudinal wave ultrasonic transducer and the bolt end face so that the ultrasonic waves emitted by the longitudinal wave ultrasonic transducer propagate along the bolt axis. The multi-frequency excitation circuit sequentially transmits coded excitation signals with different center frequencies and whose frequency bands do not completely overlap. The ultrasonic echoes of the bolts corresponding to each coded excitation signal are collected, and the bolt specifications, nominal length, effective force-bearing length, material type, initial sound velocity and temperature are correlated with the ultrasonic echoes of the bolts. When the bolt is under a known axial force and the coupling state is stable, multiple sets of bolt ultrasonic echoes are collected. The multiple sets of bolt ultrasonic echoes in the same frequency band are averaged point by point to generate a reference echo. The ultrasonic echo of the bolt is initially aligned according to the triggering time of the coded excitation signal. The current near-end reflection peak is cross-correlated with the near-end reflection peak in the reference echo. Secondary alignment is performed according to the position of the main cross-correlation peak. The subsampling position deviation is corrected by parabolic interpolation of the neighborhood of the main cross-correlation peak. The ultrasonic echoes of the aligned bolts are sequentially subjected to mean removal, bandpass filtering, median filtering and amplitude normalization. Echo quality parameters are generated based on the root mean square amplitude ratio of the effective echo region to the background noise region. The multi-frequency bolt ultrasonic echo, bolt parameters, temperature, reference echo, echo quality parameters, coded excitation identifier, and acquisition time stamp are packaged together to generate a monitoring data frame.

[0008] Optionally, step two specifically includes: The ultrasonic echoes of bolts in each frequency band are convolved with the time-reversed conjugate signal of the corresponding coded excitation signal to generate matched filter echoes. Using a standard test block with known thickness and material sound velocity, a fixed time delay is calibrated. The round-trip propagation time of the ultrasonic wave inside the bolt is determined based on the nominal length of the bolt and the initial sound velocity. This time delay is then added to the fixed time delay to obtain the expected arrival position of the bottom echo. Within the near-end region after excitation ends, the coupling interface echo region is determined based on the amplitude and duration of the reflection peak relative to the background noise. The bottom surface echo search area is set with the expected arrival position of the bottom surface echo as the center. The bottom surface echo area is determined according to the degree of complex correlation between the local energy peak and the bottom surface echo in the reference echo. The unstable reflection areas between the coupling interface echo area and the bottom surface echo area, on both sides of the bottom surface echo area and after the bottom surface echo area are determined as the sidewall aliasing echo area. Complex Morlet wavelets are used to perform complex continuous wavelet transform on the matched filter echoes of each frequency band to generate the real part, imaginary part, modulus and expanded phase of the wavelet coefficients; The wavelet modulus values ​​of different frequency bands at the same time and frequency position are subjected to ratio and logarithmic processing to generate cross-band amplitude difference. The wavelet phases of different frequency bands after phase expansion are subjected to difference processing to generate cross-band phase difference. The cross-band amplitude-phase difference is composed of cross-band amplitude difference and cross-band phase difference. Binary path masks are generated based on the echo regions of the coupling interface, the bottom echo region, and the sidewall aliasing echo region. The real part, imaginary part, modulus, expanded phase, cross-band amplitude and phase difference of the wavelet coefficients and the binary path mask are spliced ​​along the channel direction and standardized to generate the path time-frequency tensor.

[0009] Optionally, step three specifically includes: Construct an echo path-constrained two-stream convolutional network that includes a shared coding layer, coupled state branches, axial force deformation branches, aliasing reconstruction heads, and parameter output layers; The shared coding layer performs complex-valued convolution, batch normalization, ReLU activation, and pooling on the real and imaginary channels of the path time-frequency tensor to generate shared echo features. The binary path masks corresponding to the coupling interface echo region, bottom surface echo region, and sidewall aliasing echo region are adjusted to the size of the shared echo feature, and then multiplied point by point with the shared echo feature to generate the coupling interface feature, bottom surface feature, and sidewall aliasing feature. The coupled state branch performs parallel dilated convolution, feature concatenation, and pooling on the coupling interface features and cross-band amplitude and phase differences to generate coupled features. The parameter output layer generates coupled amplitude response, coupled phase response, coupled time delay and coupled reliability based on the coupling characteristics. The coupled amplitude response is used as the magnitude and the coupled phase response is used as the phase angle. They are combined to generate a coupled transfer function that characterizes the complex frequency response of an ultrasonic wave passing through the coupling layer once. The axial force deformation branch performs parallel convolution, feature concatenation, and pooling on the bottom surface features, sidewall aliasing features, and the difference features between the current bottom surface echo and the reference echo to generate deformation features; The aliasing reconstruction head reconstructs the sidewall reflection components entering the bottom echo zone based on the correlation between the bottom surface features and the sidewall aliasing features in terms of time and frequency, generating aliasing residuals.

[0010] Optionally, the echo path-constrained two-stream convolutional network specifically includes: The shared coding layer includes multiple complex-valued convolutional blocks connected in sequence. Each complex-valued convolutional block includes a complex-valued convolutional layer, a complex-valued batch normalization layer, a ReLU activation layer, and a max pooling layer in sequence. Complex-valued convolutional residual branches are set between adjacent complex-valued convolutional blocks. Complex-valued convolutional layers subtract the convolution result of the input imaginary part from the convolution result of the input real part and the real part convolution kernel to generate the output real part, and add the convolution result of the input imaginary part and the real part convolution kernel to generate the output imaginary part. The coupled state branch includes multiple parallel dilated convolution channels with different dilation rates. Each parallel dilated convolution channel sequentially performs dilated convolution, batch normalization, and ReLU activation. The outputs of each parallel dilated convolution channel are then sequentially concatenated for features, performed pointwise convolution, batch normalization, ReLU activation, and global average pooling to generate coupled features. The coupled amplitude output head in the parameter output layer generates the coupled amplitude response through fully connected processing, ReLU activation, linear mapping, and the Softplus function; the coupled phase output head generates the coupled phase response through fully connected processing, ReLU activation, and linear mapping; the coupled delay output head generates the coupled delay through fully connected processing and linear mapping; and the coupled confidence output head generates the coupled confidence through fully connected processing, ReLU activation, linear mapping, and the Sigmoid function. The axial force deformation branch includes multiple parallel convolutional channels with receptive fields in different time directions. The outputs of each parallel convolutional channel are sequentially processed by feature concatenation, pointwise convolution, batch normalization, ReLU activation, global average pooling, and fully connected processing to generate deformation features. The aliasing reconstruction head concatenates the bottom surface features and the sidewall aliasing features, and then performs convolution, batch normalization, ReLU activation and linear convolution in sequence. The corresponding channels of the linear convolution output are used as the real part and imaginary part of the aliasing residual, respectively.

[0011] Optionally, step four specifically includes: During the offline construction phase, training samples are collected under multiple axial force levels and multiple coupling states. Training samples with axial force differences within the preset axial force tolerance range and different coupling states are grouped into coaxial force heterocoupled samples, and training samples with coupling transfer function differences within the preset coupling tolerance range and different axial forces are grouped into cocoupled heteroaxial force samples. Deformation-invariant loss is generated based on the distance between the deformation features corresponding to coaxial force heterocoupled samples, coupling-invariant loss is generated based on the distance between the coupling features corresponding to cocoupled heteroaxial force samples, and orthogonal loss is generated based on the absolute value of the normalized inner product between the deformation features and the coupling features. The deformation features of one training sample and the coupling features of another training sample are input into the echo reconstruction layer to generate an exchanged reconstruction echo. The exchanged reconstruction echo is compared with the target echo corresponding to the axial force state of the previous training sample and the coupling state of the next training sample to generate the exchanged reconstruction loss. Echo path-constrained dual-stream convolutional networks are trained based on axial force regression loss, coupling delay loss, coupling transfer function reconstruction loss, deformation invariance loss, coupling invariance loss, orthogonality loss, aliasing residual loss, and exchange reconstruction loss to construct a decoupling space in which coupling features and deformation features are separated. During the online detection phase, the deformation features are projected onto the orthogonal complement space of the coupled features to generate decoupled deformation features.

[0012] Optionally, the construction of the decoupling space and the generation of the decoupling deformation features specifically include: Based on the preset axial force tolerance, coupling material, coupling layer thickness, transducer clamping state, temperature, and coupling duration, coaxial force heterocoupled samples are constructed from the training samples. Based on the preset coupling transfer function tolerance and the preset axial force difference, construct samples of co-coupled but opposite axial forces from the training samples; The echo reconstruction layer sequentially performs vector concatenation, fully connected processing, ReLU activation, feature shaping, upsampling, and convolution on the deformation and coupling features to generate the exchanged reconstructed echo. In the offline construction phase, the echo path constraint dual-stream convolutional network is first updated based on the axial force regression loss, coupling delay loss, coupling transfer function reconstruction loss, deformation invariant loss, coupling invariant loss and aliasing residual loss. Then, orthogonal loss and exchange reconstruction loss are added for joint training until the verification set axial force error, orthogonal loss and exchange reconstruction loss meet the preset convergence conditions. During the online detection phase, L2 normalization is performed on the coupling features. The L2 normalized coupling features are used as unit direction features. The inner product of the deformation features and the unit direction features is calculated. The inner product is multiplied by the unit direction features to generate coupling-related components. The coupling-related components are then subtracted from the deformation features to generate decoupled deformation features.

[0013] Optionally, step five specifically includes: The bottom echo is recovered based on the complex wavelet coefficients and binary path mask corresponding to the bottom echo region, and a Fourier transform is performed on the bottom echo to generate the bottom echo spectrum. Interpolate the aliasing residual to the same frequency sampling position as the bottom echo spectrum, and subtract the aliasing residual point by point from the bottom echo spectrum to generate the dealiasing echo spectrum. The regularization coefficient is determined based on the coupling confidence and echo quality parameters, wherein the regularization coefficient is increased when the coupling confidence or echo quality parameters decrease. Based on the two-way coupling transfer relationship formed by the bottom surface echo passing through the coupling layer during the transmission and return processes, the dealiasing echo spectrum is multiplied point by point by the square of the complex conjugate of the coupling transfer function to generate the compensation numerator. The fourth power of the magnitude of the coupling transfer function is added point by point to the regularization coefficient to generate the compensation denominator. The compensation numerator is divided point by point by the compensation denominator to generate the body echo spectrum. Frequency band weights are generated based on the coupling reliability and echo quality parameters of each frequency band. The body echo spectra of each frequency band are weighted and fused to generate a fused body echo spectrum. Compensated echoes are then generated through inverse Fourier transform. The compensated echo and the reference echo are subjected to generalized cross-correlation phase transformation processing, and the body time delay is generated based on the position of the cross-correlation main peak and the parabolic interpolation result of the cross-correlation main peak neighborhood. Compensation quality parameters are generated based on the consistency of the body time delay in different frequency bands, the correlation between the compensated echo and the reference echo, and the amplitude ratio between the main peak and the secondary peak of the cross-correlation.

[0014] Optionally, step six specifically includes: Acoustoelastic calibration samples are constructed based on real axial force, body time delay, effective force-bearing length, material type and temperature, and I-spline monotonic inversion models are established according to material type. The body delay is corrected for temperature based on the difference between the current temperature and the reference temperature and the temperature delay coefficient of the corresponding material. The temperature-corrected body delay is then divided by the effective force-bearing length to generate a unit length delay feature. The unit length time delay feature is input into multiple I-spline basis functions. The response value of each I-spline basis function is multiplied by the corresponding spline combination coefficient and summed to generate the physical axial force. The spline combination coefficient is converted to a non-negative number through the Softplus function, so that the physical axial force remains monotonically changing with the unit length time delay feature. The decoupled deformation features are sequentially input into the first fully connected layer, the ReLU activation layer, the second fully connected layer, the ReLU activation layer, the linear output layer, and the Tanh activation layer. The output of the Tanh activation layer is multiplied by the preset residual correction upper limit to generate the bounded residual correction amount. Multiply the bounded residual correction by the compensation quality parameter to generate a reliable residual correction. Add the reliable residual correction to the physical axial force to obtain the bolt axial force. The axial force reliability is generated based on coupling reliability, compensation quality parameters, consistency of body delay in different frequency bands, proportion of bounded residual correction to physical axial force, and echo quality parameters. When the axial force reliability reaches the preset reliability threshold, the bolt axial force is output. When the axial force reliability does not reach the preset reliability threshold, the axial force uncertainty is generated based on the dispersion of the axial force results in different frequency bands, the residual correction ratio, and the axial force reliability, and the axial force range is output.

[0015] Optionally, step seven specifically includes: Write the coupling interface echo, cross-band amplitude and phase difference, coupling characteristics, coupling transfer function and coupling delay corresponding to the reference echo into the coupling baseline; The coupling interface echo variation is generated based on the amplitude deviation, phase deviation, and duration deviation between the current coupling interface echo and the coupling interface echo in the coupling baseline; the cross-band variation is generated based on the deviation between the current cross-band amplitude and phase difference and the cross-band amplitude and phase difference in the coupling baseline; and the coupling feature distance is generated based on the normalized distance between the current coupling feature and the coupling feature in the coupling baseline. The coupling interface echo variation, cross-band variation, and coupling characteristic distance are fused according to preset weights to generate coupling variation values; The slope of the body delay change and the body delay fluctuation value are generated based on the body delay of multiple consecutive detection cycles. The compensation echo deviation is generated based on the correlation between the compensation echo and the reference echo and the mean absolute residual. The slope of the body delay change, the body delay fluctuation value and the compensation echo deviation are then fused according to preset weights to generate the axial force change value. The main body delay, bolt axial force, and axial force reliability are updated through the fast state channel, while the coupling baseline update judgment is performed through the slow state channel based on the coupling change value, axial force change value, coupling reliability, and axial force reliability. When the coupling change value meets the preset coupling change condition but the axial force change value does not meet the preset axial force change condition, the coupling baseline is updated based on the weighted result of the current coupling interface echo, cross-band amplitude and phase difference, coupling characteristics, coupling transfer function, coupling delay and the original coupling baseline. When the axial force change value meets the preset axial force change condition and the coupling change value does not meet the preset coupling change condition, it is determined that an axial force change has occurred and the coupling baseline is frozen. When both the coupling change value and the axial force change value meet the corresponding change conditions, it is determined that a composite change has occurred and the coupling baseline is frozen. When the coupling confidence, axial force confidence, or consistency of body delay in different frequency bands does not meet the corresponding confidence conditions, the coupling baseline is frozen and a low confidence marker is generated. The updated or frozen coupling baseline is used for coupling change determination and regularized inverse filtering in the next detection cycle, outputting bolt axial force, axial force confidence, and status flags including coupling change flags, axial force change flags, composite change flags, and low confidence flags.

[0016] The beneficial effects of this invention are: (1) By using multi-frequency coding excitation, matched filtering and complex continuous wavelet transform, the coupling interface echo, bottom surface echo and side wall aliasing echo are divided into independent path regions. The path time-frequency tensor is constructed by combining cross-band amplitude and phase difference. Then, the coupling features and deformation features are extracted by the echo path constrained dual-stream convolutional network. This solves the problem that single acoustic time difference detection is difficult to distinguish between coupling layer changes, side wall reflection interference and real axial force deformation. It improves the stability of bottom surface echo identification and coupling state estimation under complex reflection environment and provides an interpretable path-level data foundation for long-term online monitoring of bolts. (2) Construct a decoupling space using coaxial force heterocoupled samples and cocoupled heteroaxial force samples, project the deformation features onto the orthogonal complement space of the coupling features, subtract the aliasing residual from the bottom echo spectrum, and then perform regularized inverse filtering according to the single-pass coupling transfer function and the two-pass propagation relationship to solve the problem that the coupling layer thickness, compression state and aging changes are mistakenly included in the body time delay, obtain the compensation echo after weakening the sidewall aliasing and the coupling additional time delay, and reduce the axial force measurement deviation caused by repeated probe installation and coupling drift; (3) By limiting the physical monotonic relationship between axial force and body time delay through the monotonic constraint nonlinear acoustoelastic inversion layer, and using the decoupled deformation characteristics to generate bounded residual correction, and performing dual time scale baseline update or freeze according to the coupling change value and axial force change value, the problem of easy divergence of pure data model and absorption of real loosening changes in the baseline adaptive process is solved, and the synchronous output of bolt axial force, axial force credibility and status mark is realized. This is of great significance for early loosening identification and status maintenance of bolted connections in wind power, bridges and rail transit. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a bolt axial force recovery waveform compensation method based on a convolutional neural network proposed in this invention; Figure 2 This is a schematic diagram of a bolt axial force recovery waveform compensation method based on a convolutional neural network proposed in this invention; Figure 3 This is a framework diagram of the echo path constraint dual-stream convolutional network in the bolt axial force echo waveform variable compensation method based on convolutional neural network proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A bolt axial force recovery waveform compensation method based on convolutional neural network includes the following steps: Step 1: Collect the ultrasonic echo of the bolt under multi-frequency coded excitation, and align and normalize it by combining bolt parameters, temperature and reference echo to generate monitoring data frames; Step 2: Perform matched filtering and complex continuous wavelet transform on the monitoring data frame to extract the cross-band amplitude and phase difference, and generate path time-frequency tensors according to the coupling interface, bottom surface and side wall aliasing echo regions; Step 3: Input the path time-frequency tensor into the echo path-constrained dual-stream convolutional network, generate coupling features, coupling transfer function, coupling delay and coupling confidence through the coupling state branch, and generate aliasing residuals and deformation features through the axial force deformation branch; Step 4: Based on the decoupling space constructed from coaxial force heterocoupled samples and cocoupled heteroaxial force samples, project the deformation features onto the orthogonal complement space of the coupled features to generate decoupling deformation features; Step 5: Subtract the aliasing residual from the bottom echo spectrum, perform regularized inverse filtering based on the coupling transfer function and coupling confidence level, and generate the compensated echo and body delay; Step 6: Input the body time delay, effective stress length, material type and temperature into the monotonic constrained nonlinear acoustoelastic inversion layer, and combine the decoupled deformation characteristics to generate a bounded residual correction amount, and obtain the bolt axial force and axial force confidence. Step 7: Determine the coupling change and axial force change based on the coupling interface echo, cross-band amplitude and phase difference, body time delay, and deviation between the compensated echo and the reference echo. Update or freeze the coupling baseline and output the bolt axial force and status mark.

[0020] In this embodiment, step one specifically includes: A longitudinal wave ultrasonic transducer is installed on the end face of the bolt to be tested. A coupling layer is set between the ultrasonic transducer and the end face of the bolt. The central axis of the ultrasonic transducer is made to coincide with the bolt axis by a constant pressure clamping structure. Two linear frequency modulated coded excitation signals with different center frequencies and non-overlapping frequency bands are sequentially input into the ultrasonic transducer through a multi-frequency excitation circuit. The ultrasonic echo of the bolt corresponding to each coded excitation signal is collected, and the bolt specification, nominal length, effective force-bearing length, material type, initial sound velocity, bolt end face temperature and ultrasonic transducer temperature are read simultaneously. Specifically, the center frequency of the first coding excitation signal is set to 5MHz, the center frequency of the second coding excitation signal is set to 7.5MHz, the coding duration is set to 20μs, the sampling frequency is set to 100MHz, and the time interval between two adjacent coding excitations is greater than the acquisition duration of the previous echo. When the bolt axial force is confirmed by a standard force sensor and the coupling state is stable, multiple sets of bolt ultrasonic echoes are continuously collected. The multiple sets of bolt ultrasonic echoes in the same frequency band are averaged point by point to generate a reference echo, and the corresponding reference axial force, reference temperature and acquisition gain are recorded. Specifically, the reference echo is obtained by averaging 16 sets of continuous echoes, the temperature change during the acquisition period does not exceed 0.5℃, and the normalized correlation coefficient of any two adjacent sets of echoes is not less than 0.98. The ultrasonic echo of the bolt is initially aligned according to the trigger time of the coded excitation signal. The first stable reflection peak is extracted in the near-end range after the excitation ends. The current stable reflection peak is cross-correlated with the corresponding reflection peak in the reference echo. The cross-correlation is then performed according to the position of the main cross-correlation peak. The position deviation of less than one sampling period is corrected by parabolic interpolation of the main cross-correlation peak and its adjacent sampling points. The ultrasonic echoes of the aligned bolts are subjected to mean removal, bandpass filtering, median filtering and amplitude normalization in sequence. The median filtering window is set to 3 sampling points. Amplitude normalization is performed by dividing the current echo amplitude by the amplitude of the matched filter main peak of the corresponding coded excitation signal. In the bolt ultrasonic echo tail, the interval that does not contain effective reflection signal is selected as the background noise region. The ratio of the root mean square amplitude of the effective echo region to the root mean square amplitude of the background noise region is used as the echo quality parameter. The multi-frequency bolt ultrasonic echo, bolt parameters, temperature, reference echo, echo quality parameter, coded excitation identifier and acquisition time stamp are packaged together to generate a monitoring data frame. Specifically, a low-quality flag is written when the echo quality parameter is below 6, and the coupling baseline is frozen when low-quality flags are present for three consecutive detection cycles.

[0021] In this embodiment, step two specifically includes: Read the ultrasonic echoes of bolts in each frequency band in the monitoring data frame, and convolve the ultrasonic echoes of each bolt with the time-reversed conjugate signal of the corresponding coded excitation signal to generate matched filter echoes. The fixed time delay of the ultrasonic transducer and acquisition circuit is calibrated using a standard steel test block with known thickness and material sound velocity. The estimated time for the ultrasonic wave to travel back and forth inside the bolt is determined based on the nominal length of the bolt and the initial sound velocity. This time delay is then added to the fixed time delay to obtain the estimated arrival position of the bottom echo. The bottom surface echo search area is set with the expected arrival position of the bottom surface echo as the center. The first reflection peak where the amplitude of the first three consecutive sampling points exceeds three times the background noise value is searched in the near end range after the excitation ends. The position from the starting position of the reflection peak to the position where its amplitude decays to 20% of the peak value is determined as the coupling interface echo area. Specifically, the bottom echo search area is set to 1.5 μs before and after the expected arrival position of the bottom echo. When the duration of the bottom echo in the reference echo exceeds 2 μs, the bottom echo search area is expanded by 1.5 times the duration of the bottom echo. Extract local energy peaks within the bottom surface echo search area, calculate the complex correlation coefficient between each local energy peak and the bottom surface echo in the reference echo, determine the local energy peak with the largest complex correlation coefficient as the bottom surface echo, and determine the unstable reflection areas between the coupling interface echo area and the bottom surface echo area, on both sides of the bottom surface echo area, and within a preset range after the bottom surface echo as the sidewall aliasing echo area. Specifically, when the difference in the complex correlation coefficient of multiple reflection peaks does not exceed 0.03, the reflection peak with a small deviation from the expected arrival position and high phase continuity is identified as the bottom surface echo. Complex continuous wavelet transform is performed on the matched filter echoes of each frequency band. The mother wavelet adopts complex Morlet wavelet, with 64 scales set in the frequency direction and 256 positions resampled in the time direction. The real part, imaginary part, modulus and expanded phase of the wavelet coefficients are extracted. The wavelet modulus values ​​of the first and second frequency bands are respectively added to preset positive numbers and the ratio is calculated. The natural logarithm operation is then performed on the ratio to generate the cross-band amplitude difference. The wavelet phase of the first frequency band after phase expansion is subtracted from the wavelet phase of the second frequency band to generate the cross-band phase difference. Binary path masks are generated based on the echo regions of the coupling interface, the bottom surface, and the sidewall aliasing echo regions. The real part, imaginary part, magnitude, phase, cross-band amplitude difference, cross-band phase difference, and the three path masks of each frequency band wavelet coefficients are spliced ​​along the channel direction. Zero-mean unit variance standardization is performed on each channel to generate the path time-frequency tensor.

[0022] In this embodiment, step three specifically includes: The path time-frequency tensor is input into the echo path-constrained dual-stream convolutional network, which includes a shared coding layer, coupled state branch, axial force deformation branch, aliasing reconstruction head and parameter output layer. The shared coding layer consists of a first complex-valued convolutional block, a second complex-valued convolutional block, and a third complex-valued convolutional block. Each complex-valued convolutional block performs complex-valued convolution, batch normalization, ReLU activation, and max pooling in sequence. The second and third complex-valued convolutional blocks are configured with 1×1 complex-valued convolutional residual branches. Specifically, the first complex-valued convolutional block uses a 3×7 convolutional kernel and outputs 32 channels, the second complex-valued convolutional block uses a 3×5 convolutional kernel and outputs 64 channels, the third complex-valued convolutional block uses a 3×3 convolutional kernel and outputs 128 channels, and the maximum pooling window is set to 2×2 for all of them. Complex-valued convolutional layers subtract the convolution result of the input imaginary part from the convolution result of the input real part and the real part convolution kernel to obtain the output real part, and add the convolution result of the input imaginary part and the real part convolution kernel to obtain the output imaginary part. The three path masks are adjusted to the shared echo feature size through nearest neighbor interpolation, and then multiplied point by point with the shared echo feature to generate the coupling interface feature, bottom surface feature and sidewall aliasing feature. The coupled state branch includes three parallel dilated convolution channels. Each channel performs 3×3 dilated convolution, batch normalization, and ReLU activation in sequence. The outputs of the three channels are concatenated and then 1×1 convolution, batch normalization, ReLU activation, and global average pooling are performed in sequence to generate coupled features. Specifically, the dilation rates of the three parallel dilated convolution channels are set to 1, 2 and 4 respectively, each channel outputs 64 channels, and the 1×1 convolution outputs 128 channels, ultimately generating 128-dimensional coupled features; The parameter output layer includes a coupled amplitude output head, a coupled phase output head, a coupled time delay output head, and a coupled confidence output head. The coupled amplitude output head passes through a fully connected layer, a ReLU activation layer, a linear fully connected layer, and a Softplus function in sequence to generate a coupled amplitude response in positive form. The coupled phase output head passes through a fully connected layer, a ReLU activation layer, and a linear fully connected layer in sequence to generate a coupled phase response. The coupled time delay output head passes through a 64-dimensional fully connected layer, a ReLU activation layer, and a 1-dimensional linear fully connected layer in sequence. The coupled confidence output head passes through a 64-dimensional fully connected layer, a ReLU activation layer, a 1-dimensional fully connected layer, and a Sigmoid function in sequence. The coupled amplitude response is used as the magnitude and the coupled phase response is used as the phase angle to generate the coupled transfer function. Phase expansion and least squares linear fitting are performed on the coupled phase response, and the coupling delay is corrected according to the fitting slope. Coupled reliability supervision labels are generated based on the signal-to-noise ratio of the coupling interface echo, the consistency of coupling delay in different frequency bands, and the coupling interface echo reconstruction error. Each item is weighted according to a preset weight. The axial force deformation branch includes two parallel convolution channels. The first channel performs 3×9 convolution, batch normalization, and ReLU activation sequentially, while the second channel performs 3×15 convolution, batch normalization, and ReLU activation sequentially. The outputs of the two channels are concatenated and then processed sequentially with 1×1 convolution, batch normalization, ReLU activation, global average pooling, and fully connected processing to generate 128-dimensional deformation features. The aliasing reconstruction head concatenates the bottom surface features with the sidewall aliasing features, and then performs 3×3 convolution, batch normalization, ReLU activation, 3×3 convolution, batch normalization, ReLU activation and 1×1 linear convolution in sequence, and uses the last two output channels as the real part and imaginary part of the aliasing residual, respectively. Specifically, the training samples of the aliasing reconstruction head are constructed by superimposing the sidewall reflection components onto the clean bottom surface echo at an amplitude ratio of 0.2 to 0.8 and a preset time offset, and the difference in the complex-valued spectrum before and after superposition is used as the label of the aliasing residual.

[0023] In this embodiment, step four specifically includes: During the offline model building phase, no less than 8 axial force levels were applied to the training bolts using a loading device. The real axial force was recorded by a standard force sensor. After each axial force level stabilized, multi-frequency bolt ultrasonic echoes were collected. Different coupling states were constructed by changing the coupling material, coupling layer thickness, transducer clamping force, ambient temperature, and coupling duration. Samples with axial force differences not exceeding 1% of the corresponding set axial force and at least one different coupling condition are grouped into coaxial force heterocoupled samples; samples with normalized difference of coupling transfer function not exceeding 0.05 and axial force difference not less than 5% of axial force range are grouped into cocoupled heteroaxial force samples. The average absolute distance between the deformation features corresponding to coaxial force heterocoupled samples is used as the deformation invariant loss, the average absolute distance between the coupling features corresponding to cocoupled heteroaxial force samples is used as the coupling invariant loss, and the absolute value of the normalized inner product of the deformation feature and the coupling feature is used as the orthogonality loss. The echo reconstruction layer consists of a 256-dimensional fully connected layer, a ReLU activation layer, a 512-dimensional fully connected layer, a ReLU activation layer, a feature shaping layer, and three sets of bilinear upsampling and 3×3 convolutions. The deformation features of the first training sample and the coupling features of the second training sample are concatenated and then input into the echo reconstruction layer to generate the exchanged reconstructed echo. The average absolute deviation between the exchanged reconstructed echo and the target echo is used as the exchanged reconstruction loss. The total training loss is formed by combining the axial force regression loss, coupling delay loss, coupling transfer function reconstruction loss, deformation invariant loss, coupling invariant loss, orthogonality loss, aliasing residual loss, and exchange reconstruction loss according to preset weights. The echo path constrained dual-stream convolutional network was trained using the AdamW optimizer with an initial learning rate of 0.001, a batch size of 32, and a weight decay coefficient of 0.00001. The shared coding layer, coupled state branch, axial force deformation branch, and aliasing reconstruction head were trained in the first 50 training epochs. From the 51st training epoch onwards, orthogonal loss and exchange reconstruction loss were added for joint training. Specifically, training is stopped when the axial force error of the validation set does not decrease for 20 consecutive training cycles, or the orthogonal loss is below 0.08 for 10 consecutive training cycles and the change in the exchange reconstruction loss does not exceed 1% for 10 consecutive training cycles. During online detection, L2 normalization is performed on the coupling features. The normalized coupling features are used as unit direction features. The inner product of the deformation features and the unit direction features is calculated. The inner product is multiplied by the unit direction features to obtain the coupling-related components. The coupling-related components are then subtracted from the deformation features to generate decoupled deformation features.

[0024] In this embodiment, step five specifically includes: The complex wavelet coefficients corresponding to the bottom echo region are read from the path time-frequency tensor, weighted aggregation is performed according to the path mask of the bottom echo region, and the bottom echo is recovered by inverse continuous wavelet transform. Then, fast Fourier transform is performed on the bottom echo to generate the bottom echo spectrum. Linear interpolation is used to adjust the aliasing residual to the same frequency sampling position as the bottom echo spectrum. The aliasing residual is subtracted point by point from the bottom echo spectrum to generate the dealiasing echo spectrum. The regularization coefficient is determined based on the coupling confidence level and the echo quality parameter. The lower the coupling confidence level or the lower the echo quality parameter, the larger the regularization coefficient. Specifically, the basic regularization parameter is set to 0.001. When the coupling confidence is less than 0.6, the regularization coefficient is set to no less than 0.002. When the coupling confidence is not less than 0.8 and the echo quality parameter is not less than 10, the basic regularization parameter is used. The dealiasing echo spectrum is multiplied point by point by the square of the complex conjugate of the coupling transfer function to generate the compensation numerator. The fourth power of the magnitude of the coupling transfer function is added point by point to the regularization coefficient to generate the compensation denominator. The compensation numerator is divided point by point by the compensation denominator to generate the body echo spectrum. Regularized inverse filtering is performed on each frequency band, and the product of the coupling confidence and the normalized echo quality parameter of each frequency band is used as the frequency band weight. The body echo spectrum of each frequency band is weighted and averaged to generate the fused body echo spectrum. Perform an inverse fast Fourier transform on the fused body echo spectrum to generate a compensated echo. Perform a generalized cross-correlation phase transform on the compensated echo and the reference echo. Select the position of the maximum peak of the cross-correlation sequence as the initial time offset. Perform parabolic interpolation on the maximum peak and its two adjacent sampling points to generate the body time delay. Compensation quality parameters are generated based on the consistency of the time delay of different frequency bands, the normalized correlation coefficient between the compensated echo and the reference echo, and the amplitude ratio of the main peak and the secondary peak of the cross-correlation. Each item is weighted according to a preset weight. Specifically, when the difference in body delay between different frequency bands exceeds three times the standard deviation of body delay under the reference state, a delay inconsistency flag is written, and the compensation quality parameter is multiplied by 0.7.

[0025] In this embodiment, step six specifically includes: In the offline model building stage, acoustoelastic calibration samples are constructed based on the real axial force, body time delay, effective force-bearing length, material type and temperature. I-spline monotonic inversion models are established according to material type, and 8 I-spline nodes are uniformly set along the temperature-corrected body time delay range of the corresponding material training samples. Multiply the difference between the current temperature and the reference temperature by the temperature delay coefficient of the corresponding material, and subtract the obtained temperature delay correction from the body delay to generate the temperature-corrected body delay. Then divide the temperature-corrected body delay by the effective force length to generate the unit length delay feature. The unit length time delay feature is input into 8 I-spline basis functions. The response value of each I-spline basis function is multiplied by the corresponding spline combination coefficient and summed to generate the physical axial force. The spline combination coefficients are converted to non-negative numbers using the Softplus function. The Softplus function uses the natural logarithm of the sum of the exponent of the input value and 1 to keep the physical axial force relative to the unit length time delay characteristic monotonic. The 128-dimensional decoupled deformation features are sequentially input into a 64-dimensional fully connected layer, a ReLU activation layer, a 32-dimensional fully connected layer, a ReLU activation layer, a 1-dimensional fully connected layer, and a Tanh activation layer. The Tanh output is multiplied by 10% of the bolt axial force range to generate a bounded residual correction. Multiply the bounded residual correction by the compensation quality parameter to generate a reliable residual correction. Add the reliable residual correction to the physical axial force to obtain the bolt axial force. The coupling reliability, compensation quality parameters, frequency band body delay consistency and echo quality parameters are forward normalized, the proportion of bounded residual correction to physical axial force is reverse normalized, and weighted according to preset weights to generate axial force reliability. Specifically, when the absolute value of the bounded residual correction exceeds 15% of the absolute value of the physical axial force, an inversion deviation flag is written, and the axial force confidence is multiplied by 0.8. When the axial force confidence is not lower than 0.6, the single bolt axial force is output. When the axial force confidence is lower than 0.6, the axial force uncertainty is generated based on the standard deviation of the axial force results in different frequency bands, the residual correction ratio, and the axial force confidence, and the axial force range is output.

[0026] In this embodiment, step seven specifically includes: During the on-site reference initialization phase, the coupling interface echo, cross-band amplitude difference, cross-band phase difference, coupling characteristics, coupling transfer function, and coupling delay corresponding to the reference echo are written into the coupling baseline; Within each detection cycle, the current coupling interface echo is compared with the coupling interface reference echo, and the change in coupling interface echo is generated based on the amplitude deviation, phase deviation, and duration deviation. Compare the current cross-band amplitude difference and cross-band phase difference with the corresponding references to generate cross-band variation, calculate the normalized Euclidean distance between the current coupling feature and the coupling feature reference, and generate the coupling feature distance. The coupling interface echo variation, cross-band variation, and coupling characteristic distance are weighted according to preset weights to generate coupling variation values; Save the body delay of the last 5 detection cycles, perform least squares linear fitting with the detection cycle number as the independent variable and the body delay as the dependent variable, take the absolute value of the fitting slope as the slope of the body delay change, and take the standard deviation of the last 5 body delays as the body delay fluctuation value. Calculate the normalized correlation coefficient and normalized mean absolute residual between the compensated echo and the reference echo to generate the compensated echo deviation. Weight the body time delay change slope, body time delay fluctuation value and compensated echo deviation according to the preset weight to generate the axial force change value. Construct a fast state channel and a slow state channel. The fast state channel updates the body delay, bolt axial force and axial force reliability according to a detection cycle of 1 second. The slow state channel performs a coupled baseline update judgment in each detection cycle, and there is at least an interval of 30 detection cycles between two consecutive coupled baseline updates. When the coupling change value is not less than 0.65, the axial force change value is less than 0.35, the coupling confidence level is not less than 0.75, and the axial force confidence level is not less than 0.70, a coupling change is determined to have occurred. The original coupling baseline is multiplied by 0.95, the current coupling state data is multiplied by 0.05, and the corresponding data are added together to generate an updated coupling baseline. When the coupling change value is lower than 0.35, the axial force change value is not lower than 0.60, and the body delay maintains the same change direction for three consecutive detection cycles, it is determined that an axial force change has occurred and the coupling baseline is frozen. When the coupling change value is not less than 0.65 and the axial force change value is not less than 0.60, a composite change is determined and the coupling baseline is frozen. When the coupling confidence is less than 0.6, the axial force confidence is less than 0.6, or there are time delay inconsistency markers in three consecutive detection cycles, a signal abnormality is determined and the coupling baseline is frozen. Specifically, when the bolt axial force decreases for three consecutive testing cycles and the cumulative decrease reaches 5% of the reference axial force, an axial force reduction mark is generated; when the bolt axial force increases for three consecutive testing cycles and the cumulative increase reaches 5% of the reference axial force, an axial force increase mark is generated. The updated or frozen coupling baseline is used for coupling change determination and regularized inverse filtering in the next detection cycle. The bolt axial force, axial force confidence, coupling change mark, axial force increase mark, axial force decrease mark, composite change mark, time delay inconsistency mark, inversion deviation mark and low confidence mark are packaged together to output the bolt axial force and status mark.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to the axial force detection of M24 high-strength bolts on the tower flange of a wind turbine generator set. The nominal length of the bolt being tested was 180 mm, the effective force-bearing length was 128 mm, the material type was 10.9 grade alloy steel, the axial force test range was 80–240 kN, and a standard force sensor with a measurement accuracy of 0.3% of full scale was used to provide the true axial force.

[0028] The longitudinal wave ultrasonic transducer was mounted on the end face of the bolt head and kept in a compressed state by a constant pressure clamp. The center frequencies of the coded excitation signals were set to 5MHz and 7.5MHz, respectively, and the sampling frequency was set to 100MHz. Sixteen sets of bolt ultrasonic echoes were continuously collected when the bolt was under a reference axial force of 160kN, and the reference echo was generated by averaging the samples from each sampling point.

[0029] During the experiment, the coupling layer thickness was set to 0.05mm, 0.10mm, and 0.15mm; the transducer clamping force was set to 25N, 45N, and 65N; and the ambient temperature was set to 10℃, 20℃, 35℃, and 50℃. Sidewall aliasing echoes of varying degrees were generated by offsetting the longitudinal wave ultrasonic transducer installation position. Twenty sets of bolt ultrasonic echoes were collected for each axial force level, resulting in a total of 480 samples. Of these, 360 sets were used for offline construction of the echo path-constrained two-stream convolutional network and decoupling space, and 120 sets were used for testing.

[0030] This invention employs multi-frequency coded excitation, path time-frequency tensor, echo path-constrained dual-stream convolutional network, decoupling space, aliasing residual subtraction, regularized inverse filtering, monotonic constrained nonlinear acoustoelastic inversion, and dual-timescale baseline update. Comparison method A uses a dual-stream convolutional network without a binary path mask, decoupling space, and regularized inverse filtering; comparison method B uses bottom-face echo time difference and linear temperature compensation; comparison method C uses single-frequency complex continuous wavelet transform and a conventional convolutional network. All methods use the same test samples, and the results are shown in Table 1 below. Table 1 Comparison of Bolt Axial Force Testing Results Using Different Methods

[0031] As shown in Table 1, Comparison Method B directly calculates the axial force based on the time difference of the bottom echo, which is relatively simple. Linear temperature compensation can also reduce the temperature influence under stable coupling conditions, but it cannot separate the additional time delay introduced by the coupling layer. When the coupling layer thickness and transducer clamping force change, the average absolute error of the coupling change reaches 13.6 kN. After the sidewall reflection and the bottom echo overlap, the peak position of the bottom echo shifts, causing the average absolute error of the overlapped echo to reach 14.8 kN.

[0032] Comparison method C extracts echo time-frequency features using complex continuous wavelet transform, and its average absolute error for aliased echoes is 7.6 kN, significantly lower than that of comparison method B, indicating that time-frequency features can improve the axial force identification capability under complex echo conditions. However, this method only uses single-frequency echoes and cannot fully characterize the frequency-selective attenuation and phase hysteresis of the coupling layer. Therefore, the average absolute error of coupling changes still reaches 10.4 kN, and the overall relative error is 4.08%.

[0033] Comparison Method A, which uses a dual-stream convolutional network to extract coupling state information and axial force deformation information respectively, has a combined relative error of 3.61%, which is better than Comparison Methods B and C. This indicates that dual-branch feature extraction can reduce the mutual interference between different echo factors to a certain extent. However, since a binary path mask was not used to limit different propagation paths, and the deformation features were not projected onto the orthogonal complement space of the coupling features, coupling-related components still remain in the deformation features, resulting in mean absolute errors of coupling changes and aliasing echoes of 8.7 kN and 9.5 kN, respectively.

[0034] The mean absolute error of coupling variation in this invention is 4.1 kN, which is 52.9% lower than that of comparative method A, indicating that the decoupling space constructed by coaxial force heterocoupled samples and cocoupled heteroaxial force samples can effectively weaken the influence of coupling state changes on deformation characteristics; the mean absolute error of aliasing echo is 5.3 kN, which is 44.2% lower than that of comparative method A, indicating that aliasing residual subtraction and regularized inverse filtering based on coupling transfer function can reduce the joint interference of sidewall reflection and coupling layer on bottom surface echo.

[0035] The overall relative error of this invention is 2.05%, the loosening detection rate reaches 96.7%, and the false alarm rate is 2.5%. When the coupling state changes slowly while the body delay remains stable, the slow-speed state channel updates the coupling baseline; when the body delay changes continuously and meets the axial force change condition, the coupling baseline is frozen to avoid mistaking actual bolt loosening for coupling drift. Therefore, this invention can improve the accuracy of bolt axial force detection and the reliability of loosening state discrimination under conditions of coupling layer change, temperature cycling, and sidewall echo aliasing.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A bolt axial force recovery waveform compensation method based on convolutional neural networks, characterized in that, Includes the following steps: Step 1: Collect the ultrasonic echo of the bolt under multi-frequency coded excitation, and align and normalize it by combining bolt parameters, temperature and reference echo to generate monitoring data frames; Step 2: Perform matched filtering and complex continuous wavelet transform on the monitoring data frame to extract the cross-band amplitude and phase difference, and generate path time-frequency tensors according to the coupling interface, bottom surface and side wall aliasing echo regions; Step 3: Input the path time-frequency tensor into the echo path-constrained dual-stream convolutional network, generate coupling features, coupling transfer function, coupling delay and coupling confidence through the coupling state branch, and generate aliasing residuals and deformation features through the axial force deformation branch; Step 4: Based on the decoupling space constructed from coaxial force heterocoupled samples and cocoupled heteroaxial force samples, project the deformation features onto the orthogonal complement space of the coupled features to generate decoupling deformation features; Step 5: Subtract the aliasing residual from the bottom echo spectrum, perform regularized inverse filtering based on the coupling transfer function and coupling confidence level, and generate the compensated echo and body delay; Step 6: Input the body time delay, effective stress length, material type and temperature into the monotonic constrained nonlinear acoustoelastic inversion layer, and combine the decoupled deformation characteristics to generate a bounded residual correction amount, and obtain the bolt axial force and axial force confidence. Step 7: Determine the coupling change and axial force change based on the coupling interface echo, cross-band amplitude and phase difference, body time delay, and deviation between the compensated echo and the reference echo. Update or freeze the coupling baseline and output the bolt axial force and status mark.

2. The bolt axial force recovery waveform compensation method based on convolutional neural network according to claim 1, characterized in that, Step one specifically includes: A longitudinal wave ultrasonic transducer is installed on the bolt end face, and a coupling layer is set between the longitudinal wave ultrasonic transducer and the bolt end face so that the ultrasonic waves emitted by the longitudinal wave ultrasonic transducer propagate along the bolt axis. The multi-frequency excitation circuit sequentially transmits coded excitation signals with different center frequencies and whose frequency bands do not completely overlap. The ultrasonic echoes of the bolts corresponding to each coded excitation signal are collected, and the bolt specifications, nominal length, effective force-bearing length, material type, initial sound velocity and temperature are correlated with the ultrasonic echoes of the bolts. When the bolt is under a known axial force and the coupling state is stable, multiple sets of bolt ultrasonic echoes are collected. The multiple sets of bolt ultrasonic echoes in the same frequency band are averaged point by point to generate a reference echo. The ultrasonic echo of the bolt is initially aligned according to the triggering time of the coded excitation signal. The current near-end reflection peak is cross-correlated with the near-end reflection peak in the reference echo. Secondary alignment is performed according to the position of the main cross-correlation peak. The subsampling position deviation is corrected by parabolic interpolation of the neighborhood of the main cross-correlation peak. The ultrasonic echoes of the aligned bolts are sequentially subjected to mean removal, bandpass filtering, median filtering and amplitude normalization. Echo quality parameters are generated based on the root mean square amplitude ratio of the effective echo region to the background noise region. The multi-frequency bolt ultrasonic echo, bolt parameters, temperature, reference echo, echo quality parameters, coded excitation identifier, and acquisition time stamp are packaged together to generate a monitoring data frame.

3. The bolt axial force recovery waveform compensation method based on convolutional neural network according to claim 1, characterized in that, Step two specifically includes: The ultrasonic echoes of bolts in each frequency band are convolved with the time-reversed conjugate signal of the corresponding coded excitation signal to generate matched filter echoes. Using a standard test block with known thickness and material sound velocity, a fixed time delay is calibrated. The round-trip propagation time of the ultrasonic wave inside the bolt is determined based on the nominal length of the bolt and the initial sound velocity. This time delay is then added to the fixed time delay to obtain the expected arrival position of the bottom echo. Within the near-end region after excitation ends, the coupling interface echo region is determined based on the amplitude and duration of the reflection peak relative to the background noise. The bottom surface echo search area is set with the expected arrival position of the bottom surface echo as the center. The bottom surface echo area is determined according to the degree of complex correlation between the local energy peak and the bottom surface echo in the reference echo. The unstable reflection areas between the coupling interface echo area and the bottom surface echo area, on both sides of the bottom surface echo area and after the bottom surface echo area are determined as the sidewall aliasing echo area. Complex Morlet wavelets are used to perform complex continuous wavelet transform on the matched filter echoes of each frequency band to generate the real part, imaginary part, modulus and expanded phase of the wavelet coefficients; The wavelet modulus values ​​of different frequency bands at the same time and frequency position are subjected to ratio and logarithmic processing to generate cross-band amplitude difference. The wavelet phases of different frequency bands after phase expansion are subjected to difference processing to generate cross-band phase difference. The cross-band amplitude-phase difference is composed of cross-band amplitude difference and cross-band phase difference. Binary path masks are generated based on the echo regions of the coupling interface, the bottom echo region, and the sidewall aliasing echo region. The real part, imaginary part, modulus, expanded phase, cross-band amplitude and phase difference of the wavelet coefficients and the binary path mask are spliced ​​along the channel direction and standardized to generate the path time-frequency tensor.

4. The bolt axial force feedback waveform compensation method based on convolutional neural network according to claim 1, characterized in that, Step three specifically includes: Construct an echo path-constrained two-stream convolutional network that includes a shared coding layer, coupled state branches, axial force deformation branches, aliasing reconstruction heads, and parameter output layers; The shared coding layer performs complex-valued convolution, batch normalization, ReLU activation, and pooling on the real and imaginary channels of the path time-frequency tensor to generate shared echo features. The binary path masks corresponding to the coupling interface echo region, bottom surface echo region, and sidewall aliasing echo region are adjusted to the size of the shared echo feature, and then multiplied point by point with the shared echo feature to generate the coupling interface feature, bottom surface feature, and sidewall aliasing feature. The coupled state branch performs parallel dilated convolution, feature concatenation, and pooling on the coupling interface features and cross-band amplitude and phase differences to generate coupled features. The parameter output layer generates coupled amplitude response, coupled phase response, coupled time delay and coupled reliability based on the coupling characteristics. The coupled amplitude response is used as the magnitude and the coupled phase response is used as the phase angle. They are combined to generate a coupled transfer function that characterizes the complex frequency response of an ultrasonic wave passing through the coupling layer once. The axial force deformation branch performs parallel convolution, feature concatenation, and pooling on the bottom surface features, sidewall aliasing features, and the difference features between the current bottom surface echo and the reference echo to generate deformation features; The aliasing reconstruction head reconstructs the sidewall reflection components entering the bottom echo zone based on the correlation between the bottom surface features and the sidewall aliasing features in terms of time and frequency, generating aliasing residuals.

5. The bolt axial force recovery waveform compensation method based on convolutional neural network according to claim 4, characterized in that, The echo path-constrained two-stream convolutional network specifically includes: The shared coding layer includes multiple complex-valued convolutional blocks connected in sequence. Each complex-valued convolutional block includes a complex-valued convolutional layer, a complex-valued batch normalization layer, a ReLU activation layer, and a max pooling layer in sequence. Complex-valued convolutional residual branches are set between adjacent complex-valued convolutional blocks. Complex-valued convolutional layers subtract the convolution result of the input imaginary part from the convolution result of the input real part and the real part convolution kernel to generate the output real part, and add the convolution result of the input imaginary part and the real part convolution kernel to generate the output imaginary part. The coupled state branch includes multiple parallel dilated convolution channels with different dilation rates. Each parallel dilated convolution channel sequentially performs dilated convolution, batch normalization, and ReLU activation. The outputs of each parallel dilated convolution channel are then sequentially concatenated for features, performed pointwise convolution, batch normalization, ReLU activation, and global average pooling to generate coupled features. The coupled amplitude output head in the parameter output layer generates the coupled amplitude response through fully connected processing, ReLU activation, linear mapping, and the Softplus function; the coupled phase output head generates the coupled phase response through fully connected processing, ReLU activation, and linear mapping; the coupled delay output head generates the coupled delay through fully connected processing and linear mapping; and the coupled confidence output head generates the coupled confidence through fully connected processing, ReLU activation, linear mapping, and the Sigmoid function. The axial force deformation branch includes multiple parallel convolutional channels with receptive fields in different time directions. The outputs of each parallel convolutional channel are sequentially processed by feature concatenation, pointwise convolution, batch normalization, ReLU activation, global average pooling, and fully connected processing to generate deformation features. The aliasing reconstruction head concatenates the bottom surface features and the sidewall aliasing features, and then performs convolution, batch normalization, ReLU activation and linear convolution in sequence. The corresponding channels of the linear convolution output are used as the real part and imaginary part of the aliasing residual, respectively.

6. The bolt axial force recovery waveform compensation method based on convolutional neural network according to claim 1, characterized in that, Step four specifically includes: During the offline construction phase, training samples are collected under multiple axial force levels and multiple coupling states. Training samples with axial force differences within the preset axial force tolerance range and different coupling states are grouped into coaxial force heterocoupled samples, and training samples with coupling transfer function differences within the preset coupling tolerance range and different axial forces are grouped into cocoupled heteroaxial force samples. Deformation-invariant loss is generated based on the distance between the deformation features corresponding to coaxial force heterocoupled samples, coupling-invariant loss is generated based on the distance between the coupling features corresponding to cocoupled heteroaxial force samples, and orthogonal loss is generated based on the absolute value of the normalized inner product between the deformation features and the coupling features. The deformation features of one training sample and the coupling features of another training sample are input into the echo reconstruction layer to generate an exchanged reconstruction echo. The exchanged reconstruction echo is compared with the target echo corresponding to the axial force state of the previous training sample and the coupling state of the next training sample to generate the exchanged reconstruction loss. Echo path-constrained dual-stream convolutional networks are trained based on axial force regression loss, coupling delay loss, coupling transfer function reconstruction loss, deformation invariance loss, coupling invariance loss, orthogonality loss, aliasing residual loss, and exchange reconstruction loss to construct a decoupling space in which coupling features and deformation features are separated. During the online detection phase, the deformation features are projected onto the orthogonal complement space of the coupled features to generate decoupled deformation features.

7. A bolt axial force recovery waveform compensation method based on convolutional neural networks according to claim 6, characterized in that, The construction of the decoupling space and the generation of the decoupling deformation features specifically include: Based on the preset axial force tolerance, coupling material, coupling layer thickness, transducer clamping state, temperature, and coupling duration, coaxial force heterocoupled samples are constructed from the training samples. Based on the preset coupling transfer function tolerance and the preset axial force difference, construct samples of co-coupled but opposite axial forces from the training samples; The echo reconstruction layer sequentially performs vector concatenation, fully connected processing, ReLU activation, feature shaping, upsampling, and convolution on the deformation and coupling features to generate the exchanged reconstructed echo. In the offline construction phase, the echo path constraint dual-stream convolutional network is first updated based on the axial force regression loss, coupling delay loss, coupling transfer function reconstruction loss, deformation invariant loss, coupling invariant loss and aliasing residual loss. Then, orthogonal loss and exchange reconstruction loss are added for joint training until the verification set axial force error, orthogonal loss and exchange reconstruction loss meet the preset convergence conditions. During the online detection phase, L2 normalization is performed on the coupling features. The L2 normalized coupling features are used as unit direction features. The inner product of the deformation features and the unit direction features is calculated. The inner product is multiplied by the unit direction features to generate coupling-related components. The coupling-related components are then subtracted from the deformation features to generate decoupled deformation features.

8. The bolt axial force feedback waveform compensation method based on convolutional neural network according to claim 1, characterized in that, Step five specifically includes: The bottom echo is recovered based on the complex wavelet coefficients and binary path mask corresponding to the bottom echo region, and a Fourier transform is performed on the bottom echo to generate the bottom echo spectrum. Interpolate the aliasing residual to the same frequency sampling position as the bottom echo spectrum, and subtract the aliasing residual point by point from the bottom echo spectrum to generate the dealiasing echo spectrum. The regularization coefficient is determined based on the coupling confidence level and the echo quality parameter, wherein the regularization coefficient is increased when the coupling confidence level or the echo quality parameter decreases. Based on the two-way coupling transfer relationship formed by the bottom surface echo passing through the coupling layer during the transmission and return processes, the dealiasing echo spectrum is multiplied point by point by the square of the complex conjugate of the coupling transfer function to generate the compensation numerator. The fourth power of the magnitude of the coupling transfer function is added point by point to the regularization coefficient to generate the compensation denominator. The compensation numerator is divided point by point by the compensation denominator to generate the body echo spectrum. Frequency band weights are generated based on the coupling reliability and echo quality parameters of each frequency band. The body echo spectra of each frequency band are weighted and fused to generate a fused body echo spectrum. Compensated echoes are then generated through inverse Fourier transform. The compensated echo and the reference echo are subjected to generalized cross-correlation phase transformation processing, and the body time delay is generated based on the position of the cross-correlation main peak and the parabolic interpolation result of the cross-correlation main peak neighborhood. Compensation quality parameters are generated based on the consistency of the body time delay in different frequency bands, the correlation between the compensated echo and the reference echo, and the amplitude ratio between the main peak and the secondary peak of the cross-correlation.

9. A bolt axial force recovery waveform compensation method based on convolutional neural networks according to claim 1, characterized in that, Step six specifically includes: Acoustoelastic calibration samples are constructed based on real axial force, body time delay, effective force-bearing length, material type and temperature, and I-spline monotonic inversion models are established according to material type. The body delay is corrected for temperature based on the difference between the current temperature and the reference temperature and the temperature delay coefficient of the corresponding material. The temperature-corrected body delay is then divided by the effective force-bearing length to generate a unit length delay feature. The unit length time delay feature is input into multiple I-spline basis functions. The response value of each I-spline basis function is multiplied by the corresponding spline combination coefficient and summed to generate the physical axial force. The spline combination coefficient is converted to a non-negative number through the Softplus function, so that the physical axial force remains monotonically changing with the unit length time delay feature. The decoupled deformation features are sequentially input into the first fully connected layer, the ReLU activation layer, the second fully connected layer, the ReLU activation layer, the linear output layer, and the Tanh activation layer. The output of the Tanh activation layer is multiplied by the preset residual correction upper limit to generate the bounded residual correction amount. Multiply the bounded residual correction by the compensation quality parameter to generate a reliable residual correction. Add the reliable residual correction to the physical axial force to obtain the bolt axial force. The axial force reliability is generated based on coupling reliability, compensation quality parameters, consistency of body delay in different frequency bands, proportion of bounded residual correction to physical axial force, and echo quality parameters. When the axial force reliability reaches the preset reliability threshold, the bolt axial force is output. When the axial force reliability does not reach the preset reliability threshold, the axial force uncertainty is generated based on the dispersion of the axial force results in different frequency bands, the residual correction ratio, and the axial force reliability, and the axial force range is output.

10. A bolt axial force recovery waveform compensation method based on a convolutional neural network according to claim 1, characterized in that, Step seven specifically includes: Write the coupling interface echo, cross-band amplitude and phase difference, coupling characteristics, coupling transfer function and coupling delay corresponding to the reference echo into the coupling baseline; The coupling interface echo variation is generated based on the amplitude deviation, phase deviation, and duration deviation between the current coupling interface echo and the coupling interface echo in the coupling baseline; the cross-band variation is generated based on the deviation between the current cross-band amplitude and phase difference and the cross-band amplitude and phase difference in the coupling baseline; and the coupling feature distance is generated based on the normalized distance between the current coupling feature and the coupling feature in the coupling baseline. The coupling interface echo variation, cross-band variation, and coupling characteristic distance are fused according to preset weights to generate coupling variation values; The slope of the body delay change and the body delay fluctuation value are generated based on the body delay of multiple consecutive detection cycles. The compensation echo deviation is generated based on the correlation between the compensation echo and the reference echo and the mean absolute residual. The slope of the body delay change, the body delay fluctuation value and the compensation echo deviation are then fused according to preset weights to generate the axial force change value. The main body delay, bolt axial force, and axial force reliability are updated through the fast state channel, while the coupling baseline update judgment is performed through the slow state channel based on the coupling change value, axial force change value, coupling reliability, and axial force reliability. When the coupling change value meets the preset coupling change condition but the axial force change value does not meet the preset axial force change condition, the coupling baseline is updated based on the weighted result of the current coupling interface echo, cross-band amplitude and phase difference, coupling characteristics, coupling transfer function, coupling delay and the original coupling baseline. When the axial force change value meets the preset axial force change condition and the coupling change value does not meet the preset coupling change condition, it is determined that an axial force change has occurred and the coupling baseline is frozen. When both the coupling change value and the axial force change value meet the corresponding change conditions, it is determined that a composite change has occurred and the coupling baseline is frozen. When the coupling confidence, axial force confidence, or consistency of body delay in different frequency bands does not meet the corresponding confidence conditions, the coupling baseline is frozen and a low confidence marker is generated. The updated or frozen coupling baseline is used for coupling change determination and regularized inverse filtering in the next detection cycle, outputting bolt axial force, axial force confidence, and status flags including coupling change flags, axial force change flags, composite change flags, and low confidence flags.

Images