Flange fatigue life prediction system and method using ultrasonic flaw detection

Abstract

The invention discloses a system and a method for predicting the fatigue life of a flange by utilizing ultrasonic flaw detection, and belongs to the field of artificial intelligence. Comprising an aggressive frequency domain spot feature extraction module, a microdefect enhancement recognition module, an acoustic echo database training module, a damage accumulation-stress wave mapping model construction module, an active echo disturbance injection module, a fatigue life progressive prediction module, a real-time feedback closed-loop control module and a multi-scale self-checking redundancy mechanism module. Spot feature extraction and defect identification share high-frequency phase data to generate a focusing map, and registration with historical defect features is carried out in a database; the mapping model constructs a stress-damage relation according to the stress concentration coefficient and the frequency domain characteristics; the disturbance injection module applies directional excitation to potential fatigue points, and the prediction module constructs a life evolution curve based on response; the feedback control module fits and corrects the database and the disturbance strategy; the self-checking mechanism improves the prediction robustness and self-adaptability through waveform deviation and mismatch correction. The beneficial effects are that the structure safety and the service life management intelligence are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and more particularly to a flange fatigue life prediction system and method using ultrasonic flaw detection. Background Art

[0002] Traditional flange fatigue life prediction methods rely heavily on stress-strain analysis, finite element modeling, and periodic manual inspection, making it difficult to accurately and accurately understand the microevolution of fatigue damage in real time. In recent years, ultrasonic nondestructive testing (NDT) has been widely used for fatigue crack detection due to its high sensitivity and penetration. The introduction of technologies such as TOFD and phased array has significantly enhanced the ability to identify early defects within flange structures. Furthermore, the relationship between acoustic signals and material damage behavior has gradually become clearer, providing a data foundation for fatigue life prediction.

[0003] However, existing ultrasonic testing systems still have several shortcomings: limited ability to extract tiny defect features, a lack of a systematic acoustic echo database, low accuracy in modeling the correlation between acoustic waves and fatigue damage, and insufficient real-time feedback and self-evolution capabilities for predictive models. Furthermore, most systems lack multi-scale redundant verification mechanisms, making them prone to misjudgment in complex service environments. Summary of the Invention

[0004] The present invention aims to provide a flange fatigue life prediction system and method using ultrasonic testing to address the issues raised in the aforementioned background technology: limited ability to extract minute defect features, lack of a systematic acoustic echo database, low accuracy in modeling the correlation between acoustic waves and fatigue damage, and insufficient real-time feedback and self-evolution capabilities of the prediction model. Furthermore, most systems lack multi-scale redundant verification mechanisms, making them prone to misjudgment in complex service environments.

[0005] Technical solution: A flange fatigue life prediction system using ultrasonic flaw detection includes: an aggressive frequency domain speckle feature extraction module, a flange structural micro-defect enhanced recognition module, an acoustic echo database training module, a damage accumulation-stress wave mapping model construction module, an active echo disturbance injection module, a fatigue life progressive prediction module, a real-time feedback closed-loop control module, and a multi-scale self-checking redundancy mechanism module; The aggressive frequency-domain speckle feature extraction module and the micro-defect enhanced identification module work together by sharing a high-frequency phase data channel, using a speckle phase interference matrix to form a primary defect focus map. This primary defect focus map is then cross-registered with the historical structural defect evolution data via the acoustic echo database training module. The mapping model construction module establishes a multivariate stress-damage evolution model based on the measured correlation between the stress concentration coefficient of the focus area and the frequency-domain scattering characteristics. The active disturbance injection module injects directional disturbances into the target area using potential fatigue points identified by the established multivariate stress-damage evolution model. The disturbance response is processed by the fatigue life progressive prediction module using a variable-parameter multidimensional residual network to construct a dynamic evolution curve of fatigue life. The feedback closed-loop control module performs difference fitting on the dynamic evolution curve with real-time monitoring data and reversely corrects the echo database and injection strategy. The multi-scale self-checking redundancy mechanism module establishes a waveform standard deviation boundary criterion and an automatic iterative correction mechanism for the mismatch interval for each prediction result, forming a predictive control path that integrates feedforward modeling, adaptive disturbance input, real-time correction, and multi-scale robustness assessment.

[0006] Preferably, the radical frequency domain speckle feature extraction module includes an excitation signal discrete conversion unit, a speckle spectrum enhancement unit and a phase interference fusion array; The excitation signal discrete conversion unit uses an analog-to-digital converter with a sampling rate of not less than 1 GHz to perform frequency domain segment conversion on the ultrasonic echo signal to generate an energy transition curve between continuous frequency segments; the speckle spectrum enhancement unit uses the Hilbert transform and instantaneous frequency gradient convolution algorithm to extract the non-stationary components in the frequency domain signal and construct a complex wave field matrix. The phase interference fusion array is a set of variable phase delay matrices implemented based on FPGA. The control accuracy of each array element reaches within 5 nanoseconds. It is used to calculate the phase difference map between adjacent spots and reconstruct the coherent speckle peak heat map through wavelet domain phase matching. The coherent speckle peak heat map serves as the input basis for subsequent defect identification and regional focusing.

[0007] Preferably, each array element of the phase interference fusion array is constructed based on a phase-locked loop array controller, and its clock synchronization adopts a temperature-controlled crystal oscillator with a reference frequency stability better than ±0.0001%. Differential signals are used to combat EMI noise during data transmission, and a synchronous phase-locked mechanism is used to achieve phase amplitude consistency control error of each channel in the phase interference fusion array not exceeding 0.2dB, ultimately achieving detection and significant amplification of tiny phase offsets in echoes within the nanometer scale, and providing a multi-phase interference layer for defect identification.

[0008] Preferably, the flange structure micro-defect enhanced identification module includes an adaptive alignment control unit, a non-Gaussian wavefront reconstruction unit and a defect morphology reconstruction engine; The adaptive alignment control unit uses third-order B-spline interpolation based on the reference echo path comparison algorithm to perform band-by-band morphological matching on the measured scattering path and construct a deformation alignment map. The non-Gaussian wavefront reconstruction unit uses wavelet packet energy aggregation and bispectral entropy feature extraction technology to reconstruct the degree of wavefront distortion in the frequency-time joint space. Based on the above output, the defect morphology reconstruction engine uses three-dimensional voxel interpolation and Gaussian kernel density estimation to reconstruct a submillimeter scale defect space morphology map, and outputs a defect distribution model containing X, Y, and Z coordinates and corresponding scattering intensities, which is used to construct a damage propagation path map.

[0009] Preferably, the wavelet packet energy aggregation algorithm adopted by the non-Gaussian wavefront reconstruction unit selects the db20 wavelet basis, takes the maximum energy frequency band for focusing after the fifth layer of decomposition, adopts a 30ms frame length in the dual spectrum entropy calculation and eliminates the non-stationary interference component through the multi-window averaging method, realizes high-resolution decoupling and reconstruction between the abnormal wavefront and the nonlinear response of the material, and identifies weak scattering foreign objects with an amplitude lower than 10% of the baseline in the 5MHz signal.

[0010] Preferably, the active echo perturbation injection module includes a nanosecond photothermal perturbation unit and a stress wave response capture array; The photothermal disturbance unit uses a 980nm pulsed laser with a modulation frequency of 10kHz. The single-pulse energy density is controlled within 0.4J / cm², acting on the surface of the potential fatigue area. The stress wave response capture array is composed of 64-channel piezoelectric ceramic probes. Each channel records the stress waveform changes induced by the disturbance in real time. The relative time synchronization error between the probes is controlled within 3 nanoseconds. It is used to calculate the stress response differences in each area before and after the disturbance and output the spatial change map of the disturbance response.

[0011] Preferably, the stress wave response capture array adopts a self-clock locking mechanism based on an asynchronous sampling jitter elimination algorithm, and each sampling node has an embedded delay line modulation device. The external disturbance signal and the array sampling time are dynamically aligned through a bilateral locking reference frame mechanism, so that the delay jitter of the disturbance echo sampling in all channels does not exceed 1.2ns, thereby avoiding phase drift errors caused by channel asynchrony during the modeling stage.

[0012] Preferably, the fatigue life progressive prediction module constructs a damage evolution prediction structure based on a tensor time series residual network, wherein the input layer includes a stress perturbation spectrum, a microcrack extension vector sequence and a corresponding echo morphology matrix. The tensor time series residual network adopts a three-layer LSTM nested structure and a bidirectional residual connection mechanism to extract the evolution trend, and updates the fatigue life prediction value at each progressive time step. A confidence boundary calculation submodule based on Bayesian probability regression is introduced during the network training process to control the fluctuation range of the prediction value, and finally outputs the fatigue life prediction curve and its confidence band that evolve over time.

[0013] Preferably, the input echo morphology matrix is ​​cross-convolved with the stress perturbation spectrum through a three-dimensional channel fusion structure, the convolution kernel is initialized using the Xavier method and the DropBlock regularization technology is used in training to control the overfitting risk, the LSTM structure uses a structural damage delay factor embedded in the gate unit to deal with the non-uniform propagation of microcracks, the damage extension vector maintains a feature coverage of more than 95% through principal component dimensionality reduction, and after each network update, the residual error is fed back to the acoustic echo database to optimize the modeling weights in real time, thereby realizing a high-frequency, high-precision self-closed-loop iterative optimization of the fatigue life prediction process.

[0014] Preferably, the flange fatigue life prediction method using ultrasonic flaw detection includes a method for using a flange fatigue life prediction system using ultrasonic flaw detection.

[0015] Compared with the prior art, the advantages of the present invention are: (1) The characteristics of small defects are extracted by speckle interferometry and high-frequency phase interference matrix, which significantly improves the sensitivity and resolution of early microcrack recognition.

[0016] (2) By integrating multiple batches of fatigue test data with labeled defect evolution samples, an acoustic echo database that can be used for real-time comparison is constructed to achieve high-throughput automation of defect identification.

[0017] Construct a multivariate damage accumulation-stress wave mapping model: establish a coupled regression relationship between the stress concentration factor and the acoustic scattering characteristics to enhance the prediction model's ability to describe the fatigue damage evolution path under actual service conditions.

[0018] (3) Different from passive detection, the system can actively inject directional disturbances into key fatigue points to improve the accuracy and specificity of fatigue response monitoring.

[0019] (4) Realize dynamic fitting and model self-updating between fatigue life prediction and actual monitoring data, significantly improving the timeliness and adaptability of the model.

[0020] (5) Through waveform standard deviation boundary, frequency domain mismatch analysis and model automatic iteration, the robustness of the prediction path is enhanced and the error self-correction is achieved.

[0021] (6) Use AI models such as deep embedding clustering and residual networks to reduce the dimension of acoustic data, match it, and predict the life curve to improve prediction accuracy and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a schematic diagram of the overall system of a flange fatigue life prediction system using ultrasonic flaw detection according to the present invention; DETAILED DESCRIPTION

[0023] For examples, see Figure 1 , a flange fatigue life prediction system using ultrasonic flaw detection includes: a radical frequency domain speckle feature extraction module, a flange structure micro-defect enhanced recognition module, an acoustic echo database training module, a damage accumulation-stress wave mapping model construction module, an active echo disturbance injection module, a fatigue life progressive prediction module, a real-time feedback closed-loop control module and a multi-scale self-checking redundancy mechanism module; The radical frequency-domain speckle feature extraction module and the micro-defect enhanced identification module work together by sharing a high-frequency phase data channel, using the speckle phase interference matrix to form a primary defect focus map. The primary defect focus map is then cross-registered with historical structural defect evolution data via the acoustic echo database training module. The mapping model construction module establishes a multivariate stress-damage evolution model based on the measured correlation between the stress concentration factor and the frequency-domain scattering characteristics in the focus area. The active disturbance injection module injects directional disturbances into the target area using potential fatigue points identified by the established multivariate stress-damage evolution model. The disturbance response is processed by the fatigue life progressive prediction module using a variable-parameter multidimensional residual network to construct a dynamic evolution curve of fatigue life. The feedback closed-loop control module performs difference fitting between the dynamic evolution curve and real-time monitoring data and reversely corrects the echo database and injection strategy. The multi-scale self-checking redundancy mechanism module establishes waveform standard deviation boundary criteria and an automatic iterative correction mechanism for the mismatch interval for each prediction result, forming a predictive control path that integrates feedforward modeling, adaptive disturbance input, real-time correction, and multi-scale robustness assessment.

[0024] Specifically, the system is integrated into the structural health monitoring module at the flange connection. The system transmits and receives frequency-domain ultrasonic signals by deploying several sets of high-frequency ultrasonic transducers at the inspection site. All submodules utilize heterogeneous collaborative processing based on an FPGA and embedded ARM architecture. A master control script written in Python runs on a Linux system, communicating via Ethernet with a backend server. The backend server runs a neural network built with TensorFlow for fatigue life prediction curve analysis. All signal processing and stress mapping model construction steps are implemented in the on-site module and synchronized to the cloud with a 5-second refresh rate.

[0025] Specifically, the acoustic echo database training module includes an ultrasonic echo feature vector acquisition unit, a structural defect history label integration unit, and a high-dimensional feature normalization encoder. This module extracts echo waveforms by framing multiple batches of flange fatigue test data and, based on feature selection algorithms such as ReliefF and Lasso regression, extracts representative acoustic features (such as waveform kurtosis, bandwidth broadening coefficient, and spectral slope). Subsequently, a training set is constructed using structural defect labels (including defect type, spatial morphology, and evolution history). A deep embedding clustering network is used to reduce the dimensionality of high-dimensional features, generating an acoustic echo feature database suitable for online comparison. This supports subsequent echo signal similarity matching and defect evolution tracing.

[0026] Specifically, the damage accumulation-stress wave mapping model construction module is based on a composite multivariate regression framework and includes a stress concentration factor solver, an acoustic scattering pattern analyzer, and an evolution path modeler. First, the local stress field distribution in the focal area is extracted using the finite element method, and the stress enhancement factor caused by the presence of microcracks in the structure is calculated. Second, multi-channel scattering waveforms in the area are collected, and the scattering phase spectrum is reconstructed using the wavelet packet domain. Machine learning methods (such as support vector machines and XGBoost) are combined to analyze the correlation between stress field changes and scattering characteristics. Finally, a composite evolution model is established based on the fusion of damage accumulation theory (Miner's law correction model) and a dynamic fatigue growth model. This model can predict the state of each fatigue hotspot in the structure as it evolves with load and time.

[0027] Specifically, the real-time feedback closed-loop control module includes a difference fitter, a reverse correction strategy unit, and a strategy adaptation scheduling engine. Based on the difference between the currently predicted fatigue life dynamic evolution curve and the real-time collected stress wave response, the system uses spline interpolation and local weighted regression (LOWESS) to perform fitting residual analysis on the two and identify potential offset intervals. When the deviation exceeds the set tolerance boundary, the module automatically triggers the feature correction process of the echo database (including abnormal waveform deletion and echo label update), and calls the new perturbation scheme in the perturbation injection strategy library for regional replacement. The strategy adaptation scheduling engine dynamically evaluates the strategy execution effect based on the reinforcement learning framework, and optimizes the injection frequency and perturbation energy distribution to maintain the timeliness and accuracy of the system prediction model.

[0028] Specifically, the multi-scale self-verification redundancy mechanism module consists of a residual standard deviation estimator, a multi-scale matching engine, and an automatic correction feedback system. The module first extracts the residual volatility of the corresponding fatigue life curve in each prediction, calculates the standard deviation range, and defines the effective prediction bandwidth. Next, it performs wavelet domain overlap consistency analysis and frequency domain mismatch measurement on the prediction results of adjacent time periods to identify high-frequency mismatch regions. Finally, it calls upon the correction feedback system to dynamically adjust key parameters in the prediction model (such as LSTM forget gate weights and DropBlock dropout rate) to enhance model robustness. If prediction inconsistency is detected three or more times in a row, the system automatically triggers the model fine-tuning process, using recently collected real-world data for rapid iterative optimization to achieve multi-scale closed-loop convergence control of the fatigue life prediction path.

[0029] The radical frequency domain speckle feature extraction module includes an excitation signal discrete conversion unit, a speckle spectrum enhancement unit, and a phase interference fusion array; The excitation signal discrete conversion unit uses an analog-to-digital converter with a sampling rate of not less than 1GHz to perform frequency domain segment transformation on the ultrasonic echo signal and generate an energy transition curve between continuous frequency segments; the speckle spectrum enhancement unit uses the Hilbert transform and instantaneous frequency gradient convolution algorithm to extract the non-stationary components in the frequency domain signal and construct a complex wave field matrix. The phase interference fusion array is a set of variable phase delay matrices implemented based on FPGA. The control accuracy of each array element is within 5 nanoseconds. It is used to calculate the phase difference map between adjacent spots and reconstruct the coherent speckle peak heat map through wavelet domain phase matching. The coherent speckle peak heat map serves as the input basis for subsequent defect identification and regional focusing.

[0030] Specifically, the analog-to-digital converter model is TIADS54J60, with a sampling depth of 14 bits; the Hilbert transform is implemented through Python in the embedded OpenCV library, and the instantaneous frequency gradient algorithm processes every 40ms data segment in a sliding window manner; the phase interferometer fusion array is implemented using a Virtex-7 FPGA, and the delay offset step size of each phase-controlled element is set to 1ns. The acquisition results are transmitted back to the main control module via PCIe to generate a three-dimensional thermal map image, whose peak area reflects the initial spot information of the defect.

[0031] Each array element of the phase interference fusion array is built based on a phase-locked loop array controller. Its clock synchronization uses a temperature-controlled crystal oscillator with a reference frequency stability better than ±0.0001%. Differential signals are used to combat EMI noise during data transmission. A synchronous phase-locked mechanism is used to achieve phase amplitude consistency control error of no more than 0.2dB for each channel in the phase interference fusion array. Ultimately, tiny phase offsets in the echo can be detected and significantly amplified within the nanometer scale, and a multi-phase interference layer is provided for defect identification.

[0032] Specifically, the phase-locked loop control chip uses the Analog Devices AD9576, and the temperature-controlled crystal oscillator is a 10MHz OCXO with a frequency stability better than ±0.00005%. Differential signal lines connect to the array elements using shielded twisted-pair cables. All control signals are transmitted via LVDS format. Coherent speckle multi-layer overlay processing is performed using a Matlab script, supporting up to 16 layers of interference patterns, which can be refreshed and displayed in real time on the operation and maintenance interface.

[0033] The flange structure micro-defect enhanced recognition module includes an adaptive alignment control unit, a non-Gaussian wavefront reconstruction unit, and a defect morphology reconstruction engine; The adaptive alignment control unit uses third-order B-spline interpolation based on the reference echo path comparison algorithm to perform band-by-band morphological matching on the measured scattering path and construct a deformation alignment map. The non-Gaussian wavefront reconstruction unit uses wavelet packet energy aggregation and bispectral entropy feature extraction technology to reconstruct the degree of wavefront distortion in the frequency-time joint space. Based on the above output, the defect morphology reconstruction engine uses three-dimensional voxel interpolation and Gaussian kernel density estimation to reconstruct the sub-millimeter scale defect space morphology map, and outputs a defect distribution model containing X, Y, Z coordinates and corresponding scattering intensities, which is used to construct a damage propagation path map.

[0034] Specifically, the reference path in the adaptive alignment control unit is determined by defect-free data from the initial flaw detection. The third-order B-spline interpolation algorithm used is implemented in SciPy, and 2000 scattered point data are used for fitting during each alignment. Non-Gaussian wavefront reconstruction is performed using the wavelet processing module on the LabVIEW platform. The output defect voxel map is rendered in 3D using ParaView, with a display accuracy of up to 0.1 mm³.

[0035] The wavelet packet energy aggregation algorithm used by the non-Gaussian wavefront reconstruction unit adopts the db20 wavelet basis. After the fifth-level decomposition, the maximum energy frequency band is taken for focusing. A 30ms frame length is used in the dual-spectrum entropy calculation, and the multi-window averaging method is used to eliminate non-stationary interference components, achieving high-resolution decoupling and reconstruction between the abnormal wavefront and the nonlinear response of the material, and identifying weak scattering foreign objects with an amplitude lower than 10% of the baseline in the 5MHz signal.

[0036] Specifically, the Daubechies-20 (db20) wavelet basis function was used, and five-layer wavelet packet decomposition was implemented using the Python PyWavelets library. The frequency bands with the highest energy density were extracted for atlas construction. Bispectral entropy was calculated using a 60% overlapping sliding window averaging method to reduce noise interference. The results were used to generate heat maps to locate weak scattering points. The minimum detected signal amplitude was 8% of the original amplitude.

[0037] The active echo perturbation injection module includes a nanosecond photothermal perturbation unit and a stress wave response capture array; The photothermal disturbance unit uses a 980nm pulsed laser with a modulation frequency of 10kHz. The energy density of a single pulse is controlled within 0.4J / cm², acting on the surface of the potential fatigue area. The stress wave response capture array consists of 64-channel piezoelectric ceramic probes. Each channel records the stress waveform changes induced by the disturbance in real time. The relative time synchronization error between the probes is controlled within 3 nanoseconds. It is used to calculate the stress response differences in each area before and after the disturbance and output the spatial change map of the disturbance response.

[0038] Specifically, the 980nm laser is a CoherentOBIS series model with a maximum power of 50mW. The controller is set to output a modulated pulse every 0.1ms, which is focused on a 10mm² area through an optical fiber. The piezoelectric probe is made of PZT-5H material. The signal is amplified by an LNA and then sampled. The sampling system is connected to the controller via an SPI interface. The response data is output to the server in a two-dimensional matrix format for stress response difference calculation.

[0039] The stress wave response capture array adopts a self-clocked locking mechanism based on an asynchronous sampling jitter elimination algorithm. Each sampling node has an embedded delay line modulation device. The external disturbance signal is dynamically aligned with the array sampling time through a bilaterally locked reference frame mechanism, ensuring that the delay jitter of the disturbance echo sampling in all channels does not exceed 1.2ns, thereby avoiding phase drift errors caused by channel asynchrony during the modeling stage.

[0040] Specifically, each sampling node has a built-in adjustable delay line chip (such as the IDT5V9885). The sampling clock is synchronized via a distributed reference clock network. Sampling error statistics are automatically detected and adjusted by a built-in module in the FPGA, keeping the delay error within ±0.8ns. Data is uploaded in real time via USB 3.0, and dynamic alignment uses a cross-correlation function to calculate the maximum cross-correlation point for time alignment.

[0041] The fatigue life progressive prediction module constructs a damage evolution prediction structure based on the tensor time series residual network. The input layer includes the stress perturbation spectrum, the microcrack extension vector sequence and the corresponding echo morphology matrix. The tensor time series residual network adopts a three-layer LSTM nested structure and a bidirectional residual connection mechanism to extract the evolution trend and update the fatigue life prediction value at each progressive time step. During the network training process, a confidence boundary calculation submodule based on Bayesian probability regression is introduced to control the fluctuation range of the prediction value. Finally, the fatigue life prediction curve and its confidence band that evolve over time are output.

[0042] Specifically, the tensor temporal residual network was implemented using TensorFlow 2.0. The input tensor size was [100, 64, 64]. The three-layer LSTM had 128, 64, and 32 hidden units, respectively. The training set consisted of 1200 sets of historical fatigue data. Training was performed using the Adam optimizer with a learning rate of 0.001 and a loss function combining weighted mean squared error (MSE) and KL divergence. Prediction results were updated every 30 minutes, and confidence intervals were calculated at a 95% confidence level.

[0043] The input echo morphology matrix is ​​cross-convolved with the stress perturbation spectrum through a three-dimensional channel fusion structure. The Xavier method is used for convolution kernel initialization, and DropBlock regularization technology is used during training to control the risk of overfitting. The LSTM structure uses a structural damage delay factor embedded in the gated unit to address the non-uniform propagation of microcracks. The damage propagation vector maintains a feature coverage rate of over 95% through principal component dimensionality reduction. After each network update, the residual error is fed back to the acoustic echo database to optimize the modeling weights in real time, achieving self-closed-loop iterative optimization of the high-frequency and high-precision fatigue life prediction process.

[0044] Specifically, the three-dimensional channel fusion structure uses two 3×3×3 convolution kernels with ReLU activation, followed by BatchNormalization. DropBlock regularization maintains the block size at 5×5 during training and randomly drops 20% of the feature blocks, keeping GPU memory usage below 6GB during training. The tensor generated after convolution is connected to a fully connected network via a flatten layer for fatigue life curve prediction.

[0045] A flange fatigue life prediction method using ultrasonic flaw detection includes any one of claims 1-9 of the flange fatigue life prediction system using ultrasonic flaw detection.

[0046] The above shows and describes the basic principles, main features and advantages of the present invention; those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected; the scope of protection claimed in the present invention is defined by the attached claims and their equivalents.

Claims

1. A flange fatigue life prediction system using ultrasonic flaw detection, characterized in that: The flange fatigue life prediction system using ultrasonic flaw detection includes an aggressive frequency domain speckle feature extraction module, a flange structure micro-defect enhanced recognition module, an acoustic echo database training module, a damage accumulation-stress wave mapping model construction module, an active echo disturbance injection module, a fatigue life progressive prediction module, a real-time feedback closed-loop control module, and a multi-scale self-checking redundancy mechanism module; The aggressive frequency-domain speckle feature extraction module and the micro-defect enhanced identification module work together by sharing a high-frequency phase data channel, using a speckle phase interference matrix to form a primary defect focus map. This primary defect focus map is then cross-registered with historical structural defect evolution data via the acoustic echo database training module. The mapping model construction module establishes a multivariate stress-damage evolution model based on the measured correlation between the stress concentration coefficient and the frequency-domain scattering characteristics of the focus area. The active disturbance injection module injects directional disturbances into the target area using potential fatigue points identified by the established multivariate stress-damage evolution model. The disturbance response is processed by the fatigue life progressive prediction module using a variable-parameter multidimensional residual network to construct a dynamic evolution curve of fatigue life. The feedback closed-loop control module performs difference fitting on the dynamic evolution curve with real-time monitoring data and reversely corrects the echo database and injection strategy. The multi-scale self-checking redundancy mechanism module establishes a waveform standard deviation boundary criterion and an automatic iterative correction mechanism for the mismatch interval for each prediction result, forming a predictive control path that integrates feedforward modeling, adaptive disturbance input, real-time correction, and multi-scale robustness assessment.

2. The flange fatigue life prediction system using ultrasonic testing according to claim 1 is characterized in that: The radical frequency domain speckle feature extraction module includes an excitation signal discrete conversion unit, a speckle spectrum enhancement unit and a phase interference fusion array; The excitation signal discrete conversion unit uses an analog-to-digital converter with a sampling rate of not less than 1 GHz to perform frequency domain segment conversion on the ultrasonic echo signal to generate an energy transition curve between continuous frequency segments; the speckle spectrum enhancement unit uses the Hilbert transform and instantaneous frequency gradient convolution algorithm to extract the non-stationary components in the frequency domain signal and construct a complex wave field matrix. The phase interference fusion array is a set of variable phase delay matrices implemented based on FPGA. The control accuracy of each array element reaches within 5 nanoseconds. It is used to calculate the phase difference map between adjacent spots and reconstruct the coherent speckle peak heat map through wavelet domain phase matching. The coherent speckle peak heat map serves as the input basis for subsequent defect identification and regional focusing.

3. The flange fatigue life prediction system using ultrasonic testing according to claim 2 is characterized in that: Each array element of the phase interference fusion array is constructed based on a phase-locked loop array controller. Its clock synchronization uses a temperature-controlled crystal oscillator with a reference frequency stability better than ±0.0001%. Differential signals are used to combat EMI noise during data transmission. A synchronous phase-locked mechanism is used to ensure that the phase amplitude consistency of each channel in the phase interference fusion array does not exceed 0.2dB. Ultimately, it achieves the detection and significant amplification of tiny phase offsets in the echo within the nanometer scale, and provides a multi-phase interference layer for defect identification.

4. The flange fatigue life prediction system using ultrasonic flaw detection according to claim 1 is characterized in that: The flange structure micro-defect enhanced recognition module includes an adaptive alignment control unit, a non-Gaussian wavefront reconstruction unit and a defect morphology reconstruction engine; The adaptive alignment control unit uses third-order B-spline interpolation to perform band-by-band morphological matching on the measured scattering path based on the reference echo path comparison algorithm, and constructs a deformation alignment map. The non-Gaussian wavefront reconstruction unit uses wavelet packet energy aggregation and bispectral entropy feature extraction technology to reconstruct the degree of wavefront distortion in the frequency-time joint space. Based on the non-Gaussian wavefront reconstruction unit, the defect morphology reconstruction engine uses three-dimensional voxel interpolation and Gaussian kernel density estimation to reconstruct a submillimeter-scale defect space morphology map, outputs a defect distribution model containing X, Y, and Z coordinates and corresponding scattering intensities, and is used to construct a damage propagation path map.

5. The flange fatigue life prediction system using ultrasonic testing according to claim 4 is characterized in that: The wavelet packet energy aggregation algorithm adopted by the non-Gaussian wavefront reconstruction unit uses the db20 wavelet basis, takes the maximum energy frequency band for focusing after the fifth layer of decomposition, adopts a 30ms frame length in the dual spectrum entropy calculation and eliminates the non-stationary interference component through the multi-window averaging method, realizes high-resolution decoupling and reconstruction between the abnormal wavefront and the nonlinear response of the material, and identifies weak scattering foreign objects with an amplitude lower than 10% of the baseline in the 5MHz signal.

6. The flange fatigue life prediction system using ultrasonic testing according to claim 1 is characterized in that: The active echo perturbation injection module includes a nanosecond photothermal perturbation unit and a stress wave response capture array; The photothermal disturbance unit uses a 980nm pulsed laser with a modulation frequency of 10kHz. The single-pulse energy density is controlled within 0.4J / cm², acting on the surface of the potential fatigue area. The stress wave response capture array is composed of 64-channel piezoelectric ceramic probes. Each channel records the stress waveform changes induced by the disturbance in real time. The relative time synchronization error between the probes is controlled within 3 nanoseconds. It is used to calculate the stress response differences in each area before and after the disturbance and output the spatial change map of the disturbance response.

7. The flange fatigue life prediction system using ultrasonic testing according to claim 6, characterized in that: The stress wave response capture array adopts a self-clock locking mechanism based on an asynchronous sampling jitter elimination algorithm. Each sampling node has an embedded delay line modulation device. The external disturbance signal and the array sampling time are dynamically aligned through a bilaterally locked reference frame mechanism, ensuring that the delay jitter of the disturbance echo sampling in all channels does not exceed 1.2ns, thereby avoiding phase drift errors caused by channel asynchrony during the modeling stage.

8. The flange fatigue life prediction system using ultrasonic testing according to claim 1 is characterized in that: The fatigue life progressive prediction module constructs a damage evolution prediction structure based on a tensor time series residual network, in which the input layer includes a stress perturbation spectrum, a microcrack extension vector sequence, and a corresponding echo morphology matrix. The tensor time series residual network uses a three-layer LSTM nested structure and a bidirectional residual connection mechanism to extract the evolution trend, and updates the fatigue life prediction value at each progressive time step. During the network training process, a confidence boundary calculation submodule based on Bayesian probability regression is introduced to control the fluctuation range of the prediction value, and finally outputs the fatigue life prediction curve and its confidence band that evolve over time.

9. The flange fatigue life prediction system using ultrasonic testing according to claim 8, characterized in that: The input echo morphology matrix is ​​cross-convolved with the stress perturbation spectrum through a three-dimensional channel fusion structure. The convolution kernel is initialized using the Xavier method and the DropBlock regularization technology is used during training to control the risk of overfitting. The LSTM structure uses a structural damage delay factor embedded in the gate unit to deal with the non-uniform propagation of microcracks. The damage propagation vector maintains a feature coverage rate of more than 95% through principal component dimensionality reduction. After each network update, the residual error is fed back to the acoustic echo database to optimize the modeling weights in real time, achieving self-closed-loop iterative optimization of the high-frequency and high-precision fatigue life prediction process.

10. A flange fatigue life prediction method using ultrasonic testing, characterized in that: The flange fatigue life prediction method using ultrasonic flaw detection includes the method for using any one of the flange fatigue life prediction systems using ultrasonic flaw detection described in claims 1-9.

Images