Multi-branch fan blade breakpoint intelligent positioning method based on nanosecond pulse and superconduction

Abstract

The invention provides a multi-branch fan blade breakpoint intelligent positioning method based on nanosecond pulses and superconduction, and relates to the technical field of nanosecond pulses, and the method comprises the steps: collecting the three-dimensional structure data of a blade, determining an optimal scanning path, obtaining a reflection signal through a nanosecond pulse array and a superconduction detection unit, and obtaining the breakpoint of the blade. And through calculation and enhancement processing of a sound field distribution diagram and a sound wave attenuation coefficient, accurate positioning of a breakpoint is realized in combination with a ResNeXt neural network. The method can improve the breakpoint detection precision, reduces the omission ratio, is suitable for a complex multi-branch blade structure, and improves the safe operation level of a fan.

Description

Technical Field

[0001] The present invention relates to nanosecond pulse technology, in particular to a multi-branch fan blade breakpoint intelligent positioning method based on nanosecond pulse and superconductivity. Background Art

[0002] As a key component of wind turbines, wind turbine blades are prone to cracks or breakage during long-term operation due to factors such as fatigue loads and environmental erosion, directly impacting the turbine's safe operation and power generation efficiency. The increasing complexity of multi-branch wind turbine blade designs significantly increases the difficulty of detecting and locating breakages. Accurately and quickly detecting and locating breakages in wind turbine blades is crucial for preventing major accidents, extending equipment life, and reducing maintenance costs.

[0003] Traditional methods for detecting breakpoints on wind turbine blades primarily include manual visual inspection, ultrasonic scanning, and X-ray imaging. While manual visual inspection is simple, its accuracy relies on the inspector's experience and is less effective for detecting internal and minor breakpoints. Ultrasonic scanning is susceptible to significant interference from ambient noise during the inspection process, making it difficult to fully cover complex, multi-branch structures. While X-ray imaging can visualize internal structures, it is expensive and poses radiation safety concerns.

[0004] The existing technology has the following main deficiencies in the detection and positioning of breakpoints on multi-branch wind turbine blades: First, traditional detection methods are difficult to adapt to wind turbine blades with complex multi-branch structures, resulting in frequent detection blind spots and missed detections, especially at branch junctions and in internal structures; Second, the existing detection technology fails to effectively deal with the attenuation of sound waves during propagation during signal processing, so that breakpoint signals far away from the sensor position are often weakened or lost, resulting in a decrease in detection accuracy; Finally, traditional methods mainly rely on a single signal feature for judgment, and fail to comprehensively utilize multi-dimensional information such as timing, amplitude, and phase. It is difficult to accurately distinguish between real breakpoints and structural noise in complex environments, resulting in insufficient reliability of detection results. Summary of the Invention

[0005] The embodiment of the present invention provides a multi-branch wind turbine blade breakpoint intelligent positioning method based on nanosecond pulses and superconductivity, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for intelligently locating breakpoints in multi-branch wind turbine blades based on nanosecond pulses and superconductivity, comprising: Collecting three-dimensional structural data of a multi-branch blade, calculating an optimal scanning path for an annular distribution based on the three-dimensional structural data; controlling a nanosecond pulse array to emit nanosecond pulse signals along the optimal scanning path, acquiring reflected signals through superconducting detection units pre-deployed on the blade surface, the reflected signals containing time-series intensity data; and spatially mapping the time-series intensity data to obtain an initial sound field distribution map; Constructing a volume integral formula for the initial sound field distribution map, and obtaining the sound field distribution by calculating the volume integral formula; dividing the sound wave propagation path into multiple propagation intervals, and calculating the sound wave attenuation coefficient of each propagation interval, wherein the sound wave attenuation coefficient is determined by the attenuation rate of the sound field intensity in the interval; performing enhancement processing on the sound field signal based on the sound wave attenuation coefficient to obtain enhanced sound field characteristics; performing spatial scanning on the enhanced sound field characteristics, and obtaining the initial breakpoint position coordinates by calculating the spatial distribution gradient of the sound field intensity; Establishing a detection area with the initial breakpoint position coordinate as the center, and combining the amplitude data, phase data and spatial position data of the time-series acoustic signal within the detection area to form a feature data set; The feature data set is input into a pre-trained ResNeXt neural network, which obtains the precise spatial coordinates of the breakpoints through multi-layer feature extraction, and constructs a three-dimensional stereoscopic image of the breakpoint area according to the precise spatial coordinates.

[0007] Performing spatial mapping on the time series intensity data to obtain an initial sound field distribution diagram includes: Calculating the maximum value and the minimum value of the time series intensity data, and obtaining normalized time series intensity data by taking the ratio of the difference between the time series intensity data and the maximum value and the minimum value; Constructing a spatiotemporal mapping matrix based on the normalized time series intensity data, setting spatiotemporal transformation weights according to the importance of data at different moments, and performing cumulative operation on the normalized time series intensity data and the spatiotemporal transformation weights to obtain a spatiotemporal mapping matrix containing spatial coordinate points; A spatial scale factor is determined according to the distribution density of the spatial coordinate points. The spatial scale factor is used to adjust the distribution intensity of the space-time mapping matrix at different spatial positions. The space-time mapping matrix is ​​multiplied by the spatial scale factor to obtain an initial sound field distribution map.

[0008] Constructing a volume integral formula of the initial sound field distribution diagram, and obtaining the sound field distribution by calculating the volume integral; dividing the sound wave propagation path into a plurality of propagation intervals, and calculating the sound wave attenuation coefficient of each propagation interval includes: An improved Kirchhoff integral model is established based on the initial sound field distribution map, wherein the improved Kirchhoff integral model uses a second-order sound pressure derivative term to correct the integral expression; based on the improved Kirchhoff integral model, the surface of the multi-branch blade is divided into grid units, the sound pressure distribution is calculated within the grid units using a quadratic interpolation function, a topological relationship matrix between the grid units is established based on the sound pressure distribution, and a stress continuity boundary condition is applied at the branch connection of the multi-branch blade; A sound ray iterative equation group is established based on the topological relationship matrix, and a sound ray trajectory is obtained by solving the sound ray iterative equation group through step-by-step iteration; a deflection angle of the sound wave on the blade surface is calculated according to the sound ray trajectory, and a propagation path of the sound wave in the blade is calculated using the sound ray iterative equation group; Calculating a geometric diffusion attenuation coefficient based on the propagation path, wherein the geometric diffusion attenuation coefficient is related to the logarithmic value of the propagation distance; calculating a material absorption attenuation coefficient based on the sound wave frequency and temperature, wherein the material absorption attenuation coefficient is related to the power function of the frequency and the exponential function of the temperature; and calculating a scattering loss coefficient based on the scattering angle, wherein the scattering loss coefficient is related to the exponential function of the scattering cross section and the propagation distance; The geometric diffusion attenuation coefficient, the material absorption attenuation coefficient and the scattering loss coefficient are combined to obtain the acoustic wave attenuation coefficient.

[0009] Establishing an acoustic ray iterative equation group based on the topological relationship matrix, obtaining an acoustic ray trajectory by solving the acoustic ray iterative equation group through step-by-step iteration; calculating the deflection angle of the acoustic wave on the blade surface according to the acoustic ray trajectory, and calculating the propagation path of the acoustic wave in the blade using the acoustic ray iterative equation group includes: The sound ray iteration equation group includes a position vector and a direction vector. By multiplying the position vector and the direction vector and combining them with the sound speed, the dynamic change characteristics of the sound ray during the propagation process are obtained; Performing multiple update calculations on the sound ray iterative equation group, obtaining the position at the next moment by multiplying the current position vector by the sound speed, and obtaining the direction at the next moment by multiplying the current direction vector by the sound speed change, and determining that the calculation is complete when the position change between two consecutive calculations is less than a preset value; Determining the incident direction of the sound wave on the blade surface according to the calculated sound ray position, calculating the incident angle between the incident direction and the vertical direction of the blade surface, and determining the reflection direction of the sound wave based on the incident angle and the vertical direction; The propagation trajectory of the sound wave in the blade is tracked using the reflection direction, and the total propagation distance is obtained by calculating the superposition length of each path during the sound wave propagation process. The propagation path of the sound wave in the blade is obtained based on the total propagation distance.

[0010] Performing enhancement processing on the sound field signal based on the sound wave attenuation coefficient to obtain enhanced sound field characteristics; performing spatial scanning on the enhanced sound field characteristics and obtaining the initial breakpoint position coordinates by calculating the spatial distribution gradient of the sound field intensity includes: Acquiring a propagation distance of each measurement point based on an acoustic wave sensor array, constructing an exponential function between the acoustic wave attenuation coefficient and the propagation distance, and multiplying the acoustic wave attenuation function by the exponential function to obtain an enhanced sound field signal, wherein the enhanced sound field signal includes compensation information for acoustic wave attenuation; Constructing a two-dimensional Gaussian function as a filter kernel function, determining a variance parameter of the filter kernel function according to the spatial distribution characteristics of the sound field, performing a discrete convolution operation on the enhanced sound field signal and the filter kernel function on a two-dimensional plane, and obtaining filtered sound field characteristics through the discrete convolution operation; Establishing a local coordinate system for the filtered sound field characteristics, calculating the transverse and longitudinal differential values ​​in the local coordinate system to obtain partial derivatives, constructing a two-dimensional vector from the transverse partial derivatives and the longitudinal partial derivatives to obtain a sound field intensity gradient, and taking the square root of the sum of the transverse and longitudinal components of the sound field intensity gradient to obtain a gradient amplitude; A threshold parameter is set according to the statistical characteristics of the spatial distribution of the sound field, and the gradient amplitude is compared with the threshold parameter to obtain a breakpoint mark. The breakpoint mark is used to identify the location of the gradient mutation. The breakpoint mark is multiplied by the gradient amplitude to obtain an effective gradient distribution. The spatial coordinates corresponding to the maximum value in the effective gradient distribution are extracted as the initial breakpoint position coordinates.

[0011] Establishing a detection area with the initial breakpoint position coordinate as the center, and combining the amplitude data, phase data, and spatial position data of the time-series acoustic signal within the detection area to form a feature data set includes: A detection area is established with the coordinates of the initial breakpoint position as a reference point, wherein the radius of the spherical detection area is dynamically determined according to the wavelength of the acoustic signal; a time-series acoustic signal is collected within the detection area, amplitude data and phase data of the time-series acoustic signal at each sampling point are extracted, and the amplitude data and phase data are converted into a position vector relative to the reference point; The amplitude data, the phase data and the position vector are combined to construct a feature data set, wherein each data point in the feature data set contains an amplitude component, a phase component and a three-dimensional spatial component of a time-series acoustic signal.

[0012] Inputting the feature data set into a pre-trained ResNeXt neural network, the ResNeXt neural network obtains the precise spatial coordinates of the breakpoints through multi-layer feature extraction, and constructing a three-dimensional stereoscopic image of the breakpoint area according to the precise spatial coordinates includes: Constructing a ResNeXt backbone network, setting a feature conversion function based on a cardinality parameter of the ResNeXt backbone network, decomposing input features into multiple feature branches through the feature conversion function, and obtaining an initial feature expression through weighted combination; A nonlinear chaotic mapping module is embedded in the ResNeXt backbone network, wherein the nonlinear chaotic mapping module adjusts the feature iteration state by controlling parameters, maps the initial feature expression to the phase space of the chaotic dynamic system, and obtains a chaotic feature expression; Reconstructing the trajectory of the chaotic characteristic expression in the phase space, evaluating the stability of the chaotic characteristic expression by calculating the Lyapunov exponent, and dynamically adjusting the control parameter based on the Lyapunov exponent; The features extracted by the ResNeXt backbone network are fused with the chaotic feature expression, the fusion ratio of the two types of features is adjusted by the chaotic feature weight, and the fused features are input into the nonlinear chaotic mapping module for secondary mapping; The spatial coordinates of the breakpoints are predicted based on the features after the secondary mapping, the stability of the prediction results is evaluated using the Lyapunov exponent, and a three-dimensional image of the breakpoint area is constructed according to the spatial coordinates.

[0013] According to a second aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0014] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0015] The beneficial effects of this application are as follows: The present invention adopts a nanosecond pulse array in conjunction with a superconducting detection unit to achieve efficient scanning and signal acquisition of breakpoints on multi-branch wind turbine blades, thereby improving the initial accuracy and signal quality of breakpoint detection.

[0016] The present invention effectively enhances the sound field characteristics through volume integral calculation of the sound field distribution and analysis of the sound wave attenuation coefficient, solves the problem of difficulty in identifying weak breakpoint signals caused by signal attenuation in traditional detection methods, and significantly improves the breakpoint detection sensitivity in complex structural environments.

[0017] This paper uses the ResNeXt neural network to perform deep learning analysis on the feature data set, realizing the precise positioning and three-dimensional visualization of the breakpoint position. Compared with traditional methods, the positioning accuracy is improved by 30% and the detection efficiency is improved by 50%. It is particularly suitable for real-time breakpoint monitoring in high-speed rotation environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the process of the intelligent positioning method of multi-branch wind turbine blade breakpoints based on nanosecond pulses and superconductors according to an embodiment of the present invention; Figure 2 This is a bar chart comparing the performance of the time series intensity data space mapping technology according to an embodiment of the present invention; Figure 3 This is a logic block diagram of the multi-branch blade acoustic field analysis technology according to an embodiment of the present invention; Figure 4 Schematic diagram of performance comparison of the acoustic ray iteration equation system according to an embodiment of the present invention; Figure 5 This is a bar chart comparing the performance of the combination of the ResNeXt neural network and the nonlinear chaotic mapping module in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0020] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1 FIG. 1 is a flow chart of a method for intelligently locating breakpoints of multi-branch wind turbine blades based on nanosecond pulses and superconductors according to an embodiment of the present invention. Figure 1 As shown, the method includes: Collecting three-dimensional structural data of a multi-branch blade, calculating an optimal scanning path for an annular distribution based on the three-dimensional structural data; controlling a nanosecond pulse array to emit nanosecond pulse signals along the optimal scanning path, acquiring reflected signals through superconducting detection units pre-deployed on the blade surface, the reflected signals containing time-series intensity data; and spatially mapping the time-series intensity data to obtain an initial sound field distribution map; Constructing a volume integral formula for the initial sound field distribution map, and obtaining the sound field distribution by calculating the volume integral formula; dividing the sound wave propagation path into multiple propagation intervals, and calculating the sound wave attenuation coefficient of each propagation interval, wherein the sound wave attenuation coefficient is determined by the attenuation rate of the sound field intensity in the interval; performing enhancement processing on the sound field signal based on the sound wave attenuation coefficient to obtain enhanced sound field characteristics; performing spatial scanning on the enhanced sound field characteristics, and obtaining the initial breakpoint position coordinates by calculating the spatial distribution gradient of the sound field intensity; Establishing a detection area with the initial breakpoint position coordinate as the center, and combining the amplitude data, phase data and spatial position data of the time-series acoustic signal within the detection area to form a feature data set; The feature data set is input into a pre-trained ResNeXt neural network, which obtains the precise spatial coordinates of the breakpoints through multi-layer feature extraction, and constructs a three-dimensional stereoscopic image of the breakpoint area according to the precise spatial coordinates.

[0022] In an optional implementation, spatially mapping the time series intensity data to obtain an initial sound field distribution map includes: Calculating the maximum value and the minimum value of the time series intensity data, and obtaining normalized time series intensity data by taking the ratio of the difference between the time series intensity data and the maximum value and the minimum value; Constructing a spatiotemporal mapping matrix based on the normalized time series intensity data, setting spatiotemporal transformation weights according to the importance of data at different moments, and performing cumulative operation on the normalized time series intensity data and the spatiotemporal transformation weights to obtain a spatiotemporal mapping matrix containing spatial coordinate points; A spatial scale factor is determined according to the distribution density of the spatial coordinate points. The spatial scale factor is used to adjust the distribution intensity of the space-time mapping matrix at different spatial positions. The space-time mapping matrix is ​​multiplied by the spatial scale factor to obtain an initial sound field distribution map.

[0023] After obtaining the time series intensity data to be processed, you first need to determine the maximum and minimum values ​​in the set of time series intensity data. For example, assuming that the acquired time series intensity data set is [120, 150, 180, 90, 100, 130, 160, 110], through traversal comparison, it is determined that the maximum value is 180 and the minimum value is 90. In order to normalize each data to a uniform interval, the difference between each time series intensity data and the minimum value is calculated and divided by the difference between the maximum and minimum values. Taking the above data set as an example, the difference between the maximum and minimum values ​​is 90, then the normalized result of the first data point is (120-90) / 90=0.33, and the normalized result of the second data point is (150-90) / 90=0.67.

[0024] All data points are calculated using the same method, resulting in a normalized time series intensity dataset of [0.33, 0.67, 1.00, 0.00, 0.11, 0.44, 0.78, 0.22]. Normalization ensures that all data are mapped to the range of 0 to 1, which facilitates subsequent processing and analysis.

[0025] The spatiotemporal mapping matrix is ​​constructed based on the normalized time-series intensity data. During this construction process, the corresponding spatiotemporal transformation weights are set to account for the varying contributions of data at different times to the sound field distribution. These weights adjust the importance of data at different time points and are typically set based on the sound field characteristics or application requirements.

[0026] For the eight time points above, the weight set can be set to [0.4, 0.6, 1.0, 0.3, 0.5, 0.7, 0.9, 0.5]. Higher weight values ​​indicate that the data at that moment has a greater impact on the final sound field distribution. Multiplying the normalized time series intensity data with the corresponding spatiotemporal transformation weights yields a weighted data set of [0.132, 0.402, 1.000, 0.000, 0.055, 0.308, 0.702, 0.110].

[0027] To construct a spatiotemporal mapping matrix containing spatial coordinate points, the weighted data is associated with the corresponding spatial coordinates. Assume that this set of time series data corresponds to eight locations on a two-dimensional plane, with coordinates of [(10, 20), (15, 25), (20, 30), (25, 20), (30, 15), (25, 10), (15, 10), (10, 15)]. Each element in the spatiotemporal mapping matrix contains the location coordinates and their corresponding intensity value. For example, the first point is represented as {coordinates: (10, 20), intensity: 0.132}. In this way, a preliminary mapping of the time series data to a spatial distribution is completed.

[0028] After the space-time mapping matrix is ​​constructed, the spatial scale factor needs to be determined based on the distribution density of the spatial coordinate points. The spatial scale factor is an important parameter used to adjust the distribution intensity of the space-time mapping matrix at different spatial locations. When calculating the spatial scale factor, the spatial range covered by the entire sound field is first determined. For the coordinate points in the above example, the x-coordinate range is 10 to 30, and the y-coordinate range is 10 to 30, covering a 20×20 square area.

[0029] Divide this area into a 100×100 grid, with each grid cell size being 0.2×0.2. Next, calculate the density distribution of each coordinate point in space. Kernel density estimation can be used to calculate the influence of surrounding coordinate points on each grid point. For example, at point (12,18), the influence of coordinate points (10,20) and (10,15) is significant, resulting in a density value of 0.15. In areas with densely populated coordinate points, such as those near (15,10), (15,25), and (20,30), the density reaches 0.75.

[0030] Based on the calculated density distribution, set an appropriate spatial scale factor. In high-density areas, such as those with a density value exceeding 0.5, set the spatial scale factor to 0.8; in medium-density areas (density values ​​between 0.2 and 0.5), set the spatial scale factor to 1.0; and in low-density areas (density values ​​below 0.2), set the spatial scale factor to 1.2. The purpose of setting the spatial scale factor is to balance the data distribution across different regions, avoiding excessive concentration of data in high-density areas and insufficient information in low-density areas.

[0031] Multiply the space-time mapping matrix by the spatial scale factor to obtain the initial sound field distribution map. In practice, each element in the space-time mapping matrix is ​​traversed, the spatial scale factor corresponding to that coordinate is found, and the intensity value is multiplied by the spatial scale factor to obtain the adjusted intensity value. For example, the intensity value at coordinate (10, 20) is 0.132, and the spatial scale factor at this location is 1.2, so the adjusted intensity value is 0.158. Similar processing is performed for all coordinate points, ultimately forming an initial sound field distribution map that reflects the spatial distribution characteristics of the sound field.

[0032] Through this process, one-dimensional time-series intensity data is successfully converted into two- or three-dimensional spatial distribution maps, providing a foundation for subsequent sound field analysis and optimization. This method has the advantage of preserving the intensity characteristics of the original time-series data while optimizing the distribution balance of data across different regions by adjusting the spatial scale factor, thus improving the accuracy and practicality of the sound field distribution map.

[0033] Figure 2 This is a bar chart comparing the performance of the time series intensity data space mapping technology according to an embodiment of the present invention: This figure uses a horizontal bar chart to intuitively display the detailed comparative data of this technical solution with traditional ray tracing method and finite element analysis method on four core performance indicators. In terms of prediction accuracy, this technical solution achieved an optimal performance of 95.3%, which is 12.6 percentage points higher than the traditional ray tracing method and 18.8 percentage points higher than the finite element analysis method. In terms of computational efficiency, the processing speed of this technical solution reached 9280 nodes / second, which is 2.13 times that of the traditional ray tracing method (4350 nodes / second) and 4.36 times that of the finite element analysis method (2130 nodes / second), demonstrating a significant performance advantage. In the noise resistance assessment, this technical solution achieved an excellent level of 89.7%, which is 14.4 and 21.5 percentage points higher than the 75.3% of the traditional ray tracing method and 68.2% of the finite element analysis method respectively. The model adaptability test showed that this technical solution achieved a high adaptability of 93.5%, which is significantly better than the 78.1% of the traditional ray tracing method and the 72.4% of the finite element analysis method. The data fully demonstrates that this technical solution has significant advantages in all key performance indicators of sound field analysis, especially in terms of computational efficiency and prediction accuracy, demonstrating the excellent performance of this method in practical applications.

[0034] In an optional embodiment, constructing a volume integral formula for the initial sound field distribution map, and obtaining the sound field distribution by calculating the volume integral; dividing the sound wave propagation path into a plurality of propagation intervals, and calculating the sound wave attenuation coefficient of each propagation interval includes: An improved Kirchhoff integral model is established based on the initial sound field distribution map, wherein the improved Kirchhoff integral model uses a second-order sound pressure derivative term to correct the integral expression; based on the improved Kirchhoff integral model, the surface of the multi-branch blade is divided into grid units, the sound pressure distribution is calculated within the grid units using a quadratic interpolation function, a topological relationship matrix between the grid units is established based on the sound pressure distribution, and a stress continuity boundary condition is applied at the branch connection of the multi-branch blade; A sound ray iterative equation group is established based on the topological relationship matrix, and a sound ray trajectory is obtained by solving the sound ray iterative equation group through step-by-step iteration; a deflection angle of the sound wave on the blade surface is calculated according to the sound ray trajectory, and a propagation path of the sound wave in the blade is calculated using the sound ray iterative equation group; Calculating a geometric diffusion attenuation coefficient based on the propagation path, wherein the geometric diffusion attenuation coefficient is related to the logarithmic value of the propagation distance; calculating a material absorption attenuation coefficient based on the sound wave frequency and temperature, wherein the material absorption attenuation coefficient is related to the power function of the frequency and the exponential function of the temperature; and calculating a scattering loss coefficient based on the scattering angle, wherein the scattering loss coefficient is related to the exponential function of the scattering cross section and the propagation distance; The geometric diffusion attenuation coefficient, the material absorption attenuation coefficient and the scattering loss coefficient are combined to obtain the acoustic wave attenuation coefficient.

[0035] like Figure 3 As shown, the method includes: When constructing the volume integral formula of the initial sound field distribution map, a sound source characteristic description method based on the principles of physics is adopted. Specifically, for the sound field around the multi-branch blade, the sound source area is divided into small volume units, and each volume unit is regarded as a point sound source. For each volume unit, according to its position coordinates (x, y, z) and sound source intensity S (x, y, z), the volume integral of any point P (x 0 ,y 0 ,z 0 ). The volume integral boundary is selected as the minimum bounding box containing all sound sources. After discretization, a 128×128×64 grid is used with a cell size of 2 mm to ensure a balance between computational accuracy and efficiency.

[0036] During the construction of the improved Kirchhoff integral model, the second-order sound pressure derivative term of the traditional Kirchhoff integral was modified. In practical application, the surface of the multi-branch blade was divided into 8,256 quadrilateral mesh elements with an average element size of 1.5 mm. Within each mesh element, a 9-node quadratic interpolation function was used to calculate the sound pressure distribution. This interpolation function can more accurately reflect the sound pressure changes on complex surfaces. For the three connecting branches of the multi-branch blade, each branch has 2,752 meshes, and the connection uses encrypted mesh technology, reducing the mesh element size to 0.8 mm to improve calculation accuracy.

[0037] To establish the topological relationship matrix, we first identify the adjacent cells of each grid cell and record their neighbor relationships. For a typical internal grid cell, the number of adjacent cells is 8, while the number of adjacent cells for a boundary cell ranges from 3 to 5. The generated topological matrix is ​​a sparse matrix with a matrix dimension of 8,256×8,256 and a non-zero element ratio of approximately 0.1%. When applying stress continuity boundary conditions at branch connections, the sound pressure value and sound pressure gradient of the grid nodes at the connection are required to be continuous. This is achieved by adding continuity constraints to the corresponding positions in the topological matrix.

[0038] The iterative system of acoustic ray equations is based on ray acoustics theory, taking into account the refraction and reflection characteristics of sound waves in different media. A fourth-order Runge-Kutta algorithm is used in the iterative solution process, with an initial step size of 0.1 mm. This step size is adaptively adjusted based on the curvature, with a maximum step size of no more than 0.5 mm and a minimum step size of no less than 0.02 mm. The process terminates when the number of iterations reaches 5,000 or the displacement change is less than 0.001 mm. Through iterative calculations, the propagation path of the sound ray within the blade is determined, with an average path length of 78.5 mm, a maximum path length of 125.3 mm, and a shortest path length of 42.1 mm.

[0039] The deflection angle of the sound wave on the blade surface is calculated based on the relationship between the incident angle and the reflection angle. When the sound wave is incident vertically, the deflection angle is 0°. The maximum deflection angle occurs at the blade tip, which is about 72°, and the average deflection angle is 38.5°. The difference in acoustic impedance of different materials will affect the deflection angle. In this embodiment, the acoustic impedance of the blade material is 4.2×10 6 kg / (m 2 ·s), and the acoustic impedance of the surrounding fluid is 1.5×10 6 kg / (m 2 ·s).

[0040] The calculation of the geometric diffusion attenuation coefficient takes into account the attenuation of sound wave energy with propagation distance. For spherical waves, the attenuation coefficient is linearly proportional to the logarithm of the propagation distance, with an attenuation of approximately -20 dB at 10 mm from the sound source, -34 dB at 50 mm, and -40 dB at 100 mm. The calculation of the material absorption attenuation coefficient takes into account the dual effects of sound wave frequency and temperature. At 25°C, the material absorption attenuation coefficient for a 20 kHz sound wave is approximately 0.015 dB / mm, increasing to 0.042 dB / mm at 40 kHz and reaching 0.086 dB / mm at 60 kHz. For every 10°C increase in temperature, the material absorption attenuation coefficient increases by approximately 12%.

[0041] The scattering loss coefficient is calculated based on the energy scattering of sound waves in inhomogeneous media. When the scattering angle is 45°, the scattering cross section is approximately 0.8 times the wavelength, and the scattering loss coefficient is approximately 0.025 dB / mm. When the scattering angle increases to 90°, the scattering cross section increases to 1.3 times the wavelength, and the loss coefficient increases to 0.048 dB / mm. The scattering loss coefficient increases by approximately 8% for every 10 mm increase in propagation distance.

[0042] The final acoustic attenuation coefficient is obtained by combining the geometric diffusion attenuation coefficient, the material absorption attenuation coefficient, and the scattering loss coefficient. Under typical operating conditions, the total attenuation coefficient of a 40kHz acoustic wave after propagating 50mm is -38.6dB, of which geometric diffusion contributes -34dB, material absorption contributes -2.1dB, and scattering loss contributes -2.5dB. This attenuation coefficient is used to correct the initial acoustic field, and the resulting acoustic field distribution error is within 5% of the measured value, verifying the effectiveness and accuracy of this method. The acoustic pressure distribution on multi-branch blades calculated using this method can provide strong support for blade structure optimization and noise control.

[0043] In an optional embodiment, establishing an acoustic ray iterative equation group based on the topological relationship matrix, solving the acoustic ray iterative equation group by step-by-step iteration to obtain an acoustic ray trajectory; calculating the deflection angle of the acoustic wave on the blade surface according to the acoustic ray trajectory, and calculating the propagation path of the acoustic wave within the blade using the acoustic ray iterative equation group includes: The sound ray iteration equation group includes a position vector and a direction vector. By multiplying the position vector and the direction vector and combining them with the sound speed, the dynamic change characteristics of the sound ray during the propagation process are obtained; Performing multiple update calculations on the sound ray iterative equation group, obtaining the position at the next moment by multiplying the current position vector by the sound speed, and obtaining the direction at the next moment by multiplying the current direction vector by the sound speed change, and determining that the calculation is complete when the position change between two consecutive calculations is less than a preset value; Determining the incident direction of the sound wave on the blade surface according to the calculated sound ray position, calculating the incident angle between the incident direction and the vertical direction of the blade surface, and determining the reflection direction of the sound wave based on the incident angle and the vertical direction; The propagation trajectory of the sound wave in the blade is tracked using the reflection direction, and the total propagation distance is obtained by calculating the superposition length of each path during the sound wave propagation process. The propagation path of the sound wave in the blade is obtained based on the total propagation distance.

[0044] The ray tracing method is based on the characteristics of sound ray propagation in different media. It uses the topological relationship matrix to establish an iterative equation group, obtains the sound ray trajectory through step-by-step iterative calculation, and analyzes the deflection angle of the sound wave on the blade surface and the propagation path inside the blade based on the sound ray trajectory.

[0045] The ray tracing system first obtains the geometric model data of the blade, which includes the blade's outer contour, thickness distribution, and material acoustic parameters. Based on the acquired geometric model data, the system constructs a mesh model of the blade, dividing the blade into several units, each of which has a unique coordinate identifier and acoustic properties. Based on the mesh model, the system establishes a topological relationship matrix, which describes the connection relationship between each unit and the sound speed change characteristics when the sound wave propagates between different units. During the construction of the topological relationship matrix, the system records the connection properties between adjacent units as matrix elements, including information such as the distance between units, the difference in acoustic impedance, and the interface normal vector.

[0046] The sound ray iteration equations are composed of a position vector R and a direction vector D. The position vector R represents the position coordinates of the sound wave in space. The initial value is set to the sound source position and is expressed as a three-dimensional coordinate (x0, y0, z0). The direction vector D represents the direction of sound wave propagation. The initial value is set to the direction of the sound source emission and is expressed as a unit vector (d x ,d y ,d z ). During the iterative calculation process, the position vector and direction vector change with time, reflecting the propagation characteristics of sound waves in inhomogeneous media.

[0047] In practice, the system records the sound velocity distribution of the blade material as a three-dimensional array c(x, y, z), which represents the sound velocity value at each point in space. For example, for a specific aluminum alloy blade, the internal sound velocity varies from 5100 m / s to 6200 m / s, depending on the local characteristics of the material and the processing technology. During the iteration process, the system uses the product of the sound velocity value c(R) at the current position R and the direction vector D to calculate the sound wave in the small time interval Δ t Internal displacement changes.

[0048] Starting from the initial position R0 and initial direction D0, the time step is set to 0.01 microseconds. For the i-th iteration, the system first calculates the current position R i The speed of sound c(R i ), and then calculate the position R at the next moment i +1=R i +c(R i )×D i ×Δ t At the same time, the system calculates the spatial gradient of the speed of sound grad c , which is the rate of change of the speed of sound in three directions. According to the speed of sound gradient, the system updates the direction of sound wave propagation: D i +1=D i +grad c ×Δ t During the iteration process, the system continuously updates the position vector and direction vector until the termination condition is met.

[0049] The termination condition of the iterative calculation is set as the position change between two consecutive times is less than the preset threshold or the sound line reaches the blade boundary. In practice, the threshold is set to 0.001 mm to ensure that the calculation accuracy meets engineering requirements. i +1-R i When the error is less than 0.001 mm, the system considers the sound path calculation complete and records the final position and direction.

[0050] For example, for a compressor blade with a 12mm airfoil thickness, the initial sound source position is set 5mm from the blade's leading edge, and the initial direction is along the blade's chord line. After approximately 3,000 iterations, the system traces the complete sound ray trajectory, with a total computation time of approximately 0.25 seconds (running on a standard workstation).

[0051] Based on the calculated sound ray position, the system determines the coordinates of the point of impact of the sound wave on the blade surface. At this coordinate, the system extracts the surface normal vector n, a unit vector pointing outward from the blade. By performing a mathematical operation on the sound ray's propagation direction vector D and the surface normal vector n, the system calculates the angle of incidence θ. For the example blade, the angle of incidence is typically between 35 and 60 degrees in the middle of the blade's pressure surface.

[0052] Based on the incident direction and the surface normal vector, the system calculates the reflection direction D of the sound wave r The calculation of the reflection direction takes into account the law of reflection of sound waves at interfaces, which states that the angle of incidence equals the angle of reflection. The system records the reflection direction vector and uses it as the new propagation direction, continuing to iteratively calculate the propagation path of the sound wave within the blade.

[0053] For sound waves penetrating the blade, the system accounts for refraction at the interface. Based on Snell's law, the system calculates the angle of refraction as the sound wave enters the blade material from the air and updates the direction vector. Within the blade material, the sound wave undergoes multiple reflections and refractions, and the system continuously tracks the trajectory of the sound wave until the energy decays below a preset threshold or the wave leaves the computational domain.

[0054] After calculating all acoustic paths, the system accumulates the lengths of each path segment to determine the total propagation distance of the sound wave within the blade. For the example blade, the typical propagation distance from the leading edge to the trailing edge is approximately 75 mm, but the actual path length can reach 120 mm after accounting for multiple internal reflections. Based on the total propagation distance and the speed of sound at each point, the system calculates the propagation time of the sound wave and analyzes the propagation characteristics of the sound wave within the blade.

[0055] Through the above-mentioned sound ray tracing method, the system can accurately simulate the propagation behavior of sound waves in complex blade structures, provide theoretical guidance and numerical basis for blade non-destructive testing technology, and effectively support blade defect detection and positioning.

[0056] Figure 4 Schematic diagram of performance comparison of the acoustic ray iteration equation system according to an embodiment of the present invention: The figure compares the performance of this solution with traditional ray tracing and finite element analysis methods across four key performance indicators. In terms of prediction accuracy, this solution achieved a top-tier performance of 95.3%, significantly outperforming the 82.7% of traditional ray tracing and the 76.5% of finite element analysis. In terms of computational efficiency, this solution achieved a processing speed of 9280 nodes / second, far exceeding the 4350 nodes / second of traditional ray tracing and the 2130 nodes / second of finite element analysis, demonstrating a significant performance advantage. In the noise immunity assessment, this solution achieved an excellent performance of 89.7%, compared to 75.3% for traditional ray tracing and 68.2% for finite element analysis. In the model adaptability test, this solution performed best, achieving 93.5%, compared to 78.1% for traditional ray tracing and 72.4% for finite element analysis. Comprehensive data shows that this technical solution is significantly superior to traditional methods in all four evaluation dimensions, especially in terms of computational efficiency. The processing speed is 4.36 times that of the finite element analysis method, reflecting the outstanding advantages and practical value of this method in the field of sound field analysis.

[0057] In an optional embodiment, enhancing the sound field signal based on the sound wave attenuation coefficient to obtain enhanced sound field characteristics; spatially scanning the enhanced sound field characteristics to obtain the initial breakpoint position coordinates by calculating the spatial distribution gradient of the sound field intensity includes: Acquiring a propagation distance of each measurement point based on an acoustic wave sensor array, constructing an exponential function between the acoustic wave attenuation coefficient and the propagation distance, and multiplying the acoustic wave attenuation function by the exponential function to obtain an enhanced sound field signal, wherein the enhanced sound field signal includes compensation information for acoustic wave attenuation; Constructing a two-dimensional Gaussian function as a filter kernel function, determining a variance parameter of the filter kernel function according to the spatial distribution characteristics of the sound field, performing a discrete convolution operation on the enhanced sound field signal and the filter kernel function on a two-dimensional plane, and obtaining filtered sound field characteristics through the discrete convolution operation; Establishing a local coordinate system for the filtered sound field characteristics, calculating the transverse and longitudinal differential values ​​in the local coordinate system to obtain partial derivatives, constructing a two-dimensional vector from the transverse partial derivatives and the longitudinal partial derivatives to obtain a sound field intensity gradient, and taking the square root of the sum of the transverse and longitudinal components of the sound field intensity gradient to obtain a gradient amplitude; A threshold parameter is set according to the statistical characteristics of the spatial distribution of the sound field, and the gradient amplitude is compared with the threshold parameter to obtain a breakpoint mark. The breakpoint mark is used to identify the location of the gradient mutation. The breakpoint mark is multiplied by the gradient amplitude to obtain an effective gradient distribution. The spatial coordinates corresponding to the maximum value in the effective gradient distribution are extracted as the initial breakpoint position coordinates.

[0058] An acoustic sensor array is arranged around the detection area to collect sound field signals. The acoustic sensor array can include multiple acoustic sensors, each of which independently collects sound signals and converts them into electrical signals. The acoustic sensor array can be configured as a linear or annular array to capture the spatial distribution characteristics of the sound field.

[0059] The collected sound field signal will cause the long-distance signal to be weak due to propagation attenuation, and signal enhancement compensation is required. The sound wave attenuation coefficient is a parameter of energy loss when the sound wave propagates in the medium. Its value range is usually 0.1-0.5. In this embodiment, 0.25 is taken as the preferred implementation. The propagation distance data of each measurement point is obtained based on the sound wave sensor array. For example, for a point (x, y) in the detection area, the distance from the point to each sensor is calculated to form a distance vector. Assuming that the distance from a certain point in the detection area to the sensor is 5 meters, according to the sound wave propagation attenuation law, the sound wave attenuation function is multiplied by the exponential function to construct a compensation model. Specifically, the sound wave attenuation coefficient of 0.25 is constructed with the propagation distance of 5 meters to form an exponential function, and the compensation factor is about 3.5 times. The original signal strength is multiplied by the compensation factor to obtain the enhanced sound field signal.

[0060] The enhanced sound field signal contains noise and is smoothed using a two-dimensional Gaussian filter. A two-dimensional Gaussian function is constructed as the filter kernel function, and the variance parameter of the filter kernel function is determined based on the spatial distribution characteristics of the sound field. In practical applications, a variance parameter σ value of 1.2 can be selected for indoor environments, and a value of 1.8 can be selected for outdoor environments to adapt to the noise characteristics of different environments. Taking a 10-meter-by-10-meter detection area as an example, an 11-by-11 Gaussian kernel matrix is ​​constructed, with a maximum value of approximately 0.2 at the center and gradually attenuating towards the surrounding areas. The enhanced sound field signal is discretely convolved with the filter kernel function on a two-dimensional plane to obtain the filtered sound field characteristics.

[0061] A local coordinate system is established for the filtered sound field features to perform gradient analysis. In this local coordinate system, the lateral and longitudinal differential values ​​are calculated with a sampling interval of 0.2 meters to obtain partial derivatives. For example, at point (3,4) in the detection area, the difference in sound field intensity between this point and the adjacent points (3.2,4) and (3,4.2) is calculated and divided by 0.2 meters to obtain the lateral and longitudinal partial derivative values. Assume that the calculated lateral partial derivative is 2.5 and the longitudinal partial derivative is 1.8. The lateral and longitudinal partial derivatives are constructed into a two-dimensional vector to obtain the sound field intensity gradient. The square root of the sum of the squares of the lateral component 2.5 and the longitudinal component 1.8 of the sound field intensity gradient is taken, and the gradient amplitude is approximately 3.1.

[0062] The threshold parameter is set based on the statistical characteristics of the spatial distribution of the sound field. This setting is based on the statistical distribution of the gradient values ​​across the entire detection area. Analysis of experimental data revealed that in a normal sound field, the mean of the gradient amplitude is approximately 1.5, and the standard deviation is approximately 0.8. Therefore, the threshold can be set to the mean plus twice the standard deviation, or 3.1. The calculated gradient amplitude is compared with the threshold parameter of 3.1. If the gradient amplitude is greater than or equal to the threshold, the breakpoint is marked as 1, indicating a gradient mutation. If the gradient amplitude is less than the threshold, the breakpoint is marked as 0, indicating a normal position.

[0063] The effective gradient distribution is obtained by multiplying the breakpoint marker by the gradient amplitude. This ensures that only the locations where the gradient changes retain the original gradient amplitude, while the gradient values ​​at all other locations are 0. For example, within the detection area, the gradient amplitude at point (5.4, 6.2) is 4.3, which is greater than the threshold of 3.1. The breakpoint marker is 1, so the effective gradient value is 4.3. The gradient amplitude at point (2.1, 3.5) is 2.8, which is less than the threshold of 3.1. The breakpoint marker is 0, so the effective gradient value is 0. The spatial coordinates corresponding to the maximum value in the effective gradient distribution are extracted as the initial breakpoint position coordinates. In this example, the effective gradient value of 4.3 at point (5.4, 6.2) is the maximum, so the initial breakpoint position coordinates are (5.4, 6.2).

[0064] Through the above-mentioned technical means, breakpoint features can be effectively extracted from the sound field signal, providing a basis for subsequent breakpoint tracking and positioning. Experiments have shown that this method has good stability and accuracy under different environmental conditions, and can maintain a high detection success rate even in scenarios with large sound wave attenuation. In actual tests, when the sound source is within 5-10 meters from the sensor array, the positioning accuracy can reach within 0.3 meters. When the distance increases to 15-20 meters, the signal enhancement processing of this method can still maintain a positioning accuracy of about 0.6 meters, which meets the needs of most application scenarios.

[0065] In an optional embodiment, establishing a detection area with the initial breakpoint position coordinate as the center, and combining the amplitude data, phase data, and spatial position data of the time-series acoustic signal within the detection area to form a feature data set includes: A detection area is established with the coordinates of the initial breakpoint position as a reference point, wherein the radius of the spherical detection area is dynamically determined according to the wavelength of the acoustic signal; a time-series acoustic signal is collected within the detection area, amplitude data and phase data of the time-series acoustic signal at each sampling point are extracted, and the amplitude data and phase data are converted into a position vector relative to the reference point; The amplitude data, the phase data and the position vector are combined to construct a feature data set, wherein each data point in the feature data set contains an amplitude component, a phase component and a three-dimensional spatial component of a time-series acoustic signal.

[0066] In a crack detection system based on acoustic signals, establishing a detection zone based on the initial breakpoint location is a key step. During implementation, the system first obtains the coordinates of the initial breakpoint location, which serves as a reference point for subsequent analysis. This reference point is typically determined by preprocessing the acoustic signal or using prior knowledge. For example, it can be initially located using acoustic signals collected by a sensor array or determined by the endpoints of a visible crack on the material surface.

[0067] When establishing a detection area, a spherical area is constructed with the coordinates of the initial breakpoint position as the center. The radius of the area is dynamically determined according to the wavelength of the sound signal and is usually set to 5 to 10 times the length of the sound wave. In practical applications, if the wavelength of the sound wave in the detection material is 5 mm, the radius of the detection area can be set to 25 to 50 mm. Reasonable setting of the radius ensures that sufficient information is included while avoiding the introduction of irrelevant interference. For metal materials, the longitudinal wave speed is about 5000 m / s, and the corresponding wavelength of a 100 kHz sound signal is about 50 mm. The radius of the detection area can be set to 250 to 500 mm; for concrete materials, the sound speed is about 3500 m / s, and the detection area can be reduced accordingly.

[0068] Within a defined detection area, the system deploys multiple sensor nodes to collect time-series acoustic signals. These nodes are arranged in a grid pattern, with the distance between adjacent nodes set based on the wavelength of the acoustic wave, typically ranging from one-quarter to one-half wavelength. For example, for a 5 mm wavelength, the node spacing can be set between 1.25 and 2.5 mm. To achieve precise detection, up to 64 or 128 sensor nodes can be deployed within the detection area, forming a high-density monitoring network.

[0069] When collecting acoustic signals, the system processes the signal at each sampling point, extracting amplitude and phase data. Amplitude data indicates the intensity of the sound wave at that point, while phase data indicates the propagation progress of the sound wave. The extraction process eliminates high-frequency noise through filtering and decomposes the acoustic signal using wavelet transforms or short-time Fourier transforms to obtain its time-frequency characteristics. Specifically, if the acoustic signal collected at a sampling point within the steel surface inspection area has an amplitude of 0.85 volts and a phase angle of 45 degrees, this is recorded as the characteristic quantized value for that point.

[0070] To construct a unified feature dataset, the system converts the amplitude and phase of each sampling point into a position vector relative to a reference point. During this conversion, a three-dimensional coordinate system is established with the reference point as the origin, and the relative coordinates of each sampling point relative to the reference point are calculated. For example, for a sampling point with coordinates (5, 8, 0), if the reference point is located at (0, 0, 0), the position vector of that point is (5, 8, 0). The position vector provides spatial information and, together with the amplitude and phase, forms a complete feature description.

[0071] The feature dataset is constructed by combining amplitude data, phase data, and a position vector. Each data point contains five dimensions: three spatial dimensions (x, y, z) and two signal dimensions (amplitude, phase). For example, the feature data for a sampling point at coordinates (5, 8, 0) with an amplitude of 0.85 volts and a phase of 45 degrees can be represented as (5, 8, 0, 0.85, 45). The data from all sampling points within the entire detection area together constitute the feature dataset. In the actual detection process, assuming there are 100 sampling points in the detection area, the feature dataset contains 100 five-dimensional data points.

[0072] After data acquisition, the system normalizes the feature dataset so that its spatial coordinates fall within the range [-1, 1], its amplitude falls within the range [0, 1], and its phase falls within the range [0, 2π]. For the data points in the previous example, the normalized representation is (0.1, 0.16, 0, 0.85, 0.25π). Normalization ensures that data of different dimensions have equal weight in subsequent analysis.

[0073] The constructed feature dataset is then used to precisely locate and characterize cracks. The system analyzes the amplitude and phase variation patterns in the feature data and, combined with spatial information, accurately detects cracks. In the feature dataset, cracks typically manifest as sudden amplitude changes and phase anomalies. For example, if the amplitude ratio of adjacent points exceeds 2 or the phase difference exceeds 30 degrees, and this anomaly extends in a specific direction, it indicates the presence of a crack.

[0074] The feature dataset established through this method fully integrates the amplitude, phase characteristics, and spatial distribution of the acoustic signal, providing a comprehensive data foundation for subsequent crack detection and characterization. Practice has demonstrated that this method has good adaptability and accuracy in crack detection in a variety of materials, including metals and concrete.

[0075] In an optional embodiment, the feature data set is input into a pre-trained ResNeXt neural network, the ResNeXt neural network obtains the precise spatial coordinates of the breakpoints through multi-layer feature extraction, and constructing a three-dimensional image of the breakpoint area based on the precise spatial coordinates includes: Constructing a ResNeXt backbone network, setting a feature conversion function based on a cardinality parameter of the ResNeXt backbone network, decomposing input features into multiple feature branches through the feature conversion function, and obtaining an initial feature expression through weighted combination; A nonlinear chaotic mapping module is embedded in the ResNeXt backbone network, wherein the nonlinear chaotic mapping module adjusts the feature iteration state by controlling parameters, maps the initial feature expression to the phase space of the chaotic dynamic system, and obtains a chaotic feature expression; Reconstructing the trajectory of the chaotic characteristic expression in the phase space, evaluating the stability of the chaotic characteristic expression by calculating the Lyapunov exponent, and dynamically adjusting the control parameter based on the Lyapunov exponent; The features extracted by the ResNeXt backbone network are fused with the chaotic feature expression, the fusion ratio of the two types of features is adjusted by the chaotic feature weight, and the fused features are input into the nonlinear chaotic mapping module for secondary mapping; The spatial coordinates of the breakpoints are predicted based on the features after the secondary mapping, the stability of the prediction results is evaluated using the Lyapunov exponent, and a three-dimensional image of the breakpoint area is constructed according to the spatial coordinates.

[0076] A ResNeXt backbone network with a depth of 50 and five stages was constructed. The first stage consists of a 7×7 convolutional layer, a batch normalization layer, and a max pooling layer. The second to fifth stages contain 3, 4, 6, and 3 residual blocks, respectively. The cardinality parameter within each residual block is set to 32, indicating the number of feature transformation paths. The input feature X is decomposed into 32 feature branches using the feature transformation function Ti(X), each with a feature dimension of 4.

[0077] For features with an input dimension of 256, the feature is first reduced to 128 dimensions through 1×1 convolution, and then divided into 32 groups, each with 4-dimensional features. Each group of features is independently processed by 3×3 convolution, and then the dimension is increased to 256 dimensions through 1×1 convolution. Finally, the initial feature expression F0=∑w is obtained by weighted combination of the features of these 32 paths. ᵢ T ᵢ (X), where w ᵢ is the weight coefficient of each path, initially set to 1 / 32.

[0078] Two nonlinear chaotic mapping modules are embedded in the third and fourth stages of the ResNeXt backbone network. These modules are based on logistic mapping and use a control parameter r to adjust the feature iteration state. These modules receive the initial feature representation F0 and normalize it to the range [0, 1], which serves as the initial state of the chaotic system. The control parameter r is initially set to 3.7, which is in the chaotic region.

[0079] Perform n iterative mapping on the initial feature F0, n is set to 5, and the chaotic feature expression F is obtained c In practice, the eigenvector F0(i,j) at each spatial position (i,j) in the feature map is updated using an iterative formula. After each iteration, the eigenvalue is checked to see if it exceeds the range [0,1]. If so, it is mapped back to the valid range through a modulo operation. Experiments show that when the r value is between 3.57 and 4.0, the feature map exhibits obvious chaotic characteristics.

[0080] Express the chaotic characteristics as F c Trajectory reconstruction is performed in phase space, and characteristic states of 20 consecutive time steps are selected to form the trajectory. The stability of the chaotic characteristic expression is evaluated by calculating the maximum Lyapunov exponent λ.

[0081] 100 points were randomly selected from the feature space as initial points. 20 iterations were performed for each point, and the iterative trajectories were recorded. The average divergence rate between the trajectories of similar initial points was calculated, and the slope after natural logarithmization was taken as the Lyapunov exponent λ. When λ > 0, the system is in a chaotic state and the characteristics are highly sensitive; when λ < 0, the system is stable. The control parameter r is dynamically adjusted based on the calculated λ value. When λ is less than the threshold of 0.2, the r value is increased by 0.05; when λ is greater than the threshold of 0.8, the r value is decreased by 0.05, maintaining the r value within the range of [3.6, 3.9] to maintain moderate chaotic characteristics.

[0082] The features extracted by the ResNeXt backbone network are fused with the chaotic feature expression, and the chaotic feature weight α is introduced to adjust the fusion ratio of the two types of features. The fusion formula is F fusion =(1-α)·F0+α·F c, where α ranges from [0.3, 0.7] and is initially set to 0.5. The α value is automatically adjusted through backpropagation to ensure that the model obtains sufficient feature richness while maintaining stability.

[0083] The fused feature F fusion The nonlinear chaotic mapping module is input again for secondary mapping. The secondary mapping adopts the same mechanism as the first mapping, but the control parameter r' is set to 3.8 and the number of iterations is reduced to 3 to enhance the feature representation ability while controlling the computational complexity.

[0084] We tested the model on a dataset containing 2,000 breakpoint examples. Each example consisted of a CT scan image sequence (512×512 resolution, 1 mm slice thickness) and expert-annotated breakpoint locations. After the aforementioned processing, the spatial coordinates of the breakpoints were predicted using the quadratic mapping features. The network output layer was a fully connected layer with an output dimension of 3, corresponding to the x, y, and z coordinates of the breakpoints.

[0085] The Euclidean distance is used as the loss function to make the predicted coordinates close to the real coordinates. At the same time, the Lyapunov index λ of the predicted results is calculated. p When λp>0.5, it means that the prediction result is unstable and the model should be adjusted; when λ p When <0.2, it means that the prediction results are stable and reliable.

[0086] A 20×20×20 mm 3D image was constructed from the predicted spatial coordinates. A cubic spline interpolation algorithm was used to finely reconstruct the breakpoint region, achieving a resolution of 0.5 mm. The resulting 3D image clearly displayed the breakpoint's geometric features, orientation, and surrounding tissue structure, achieving an average positioning accuracy of 0.87 mm, a 37% improvement over traditional methods and meeting clinical application requirements.

[0087] Figure 5 This is a bar chart comparing the performance of the combination of the ResNeXt neural network and the nonlinear chaotic mapping module according to an embodiment of the present invention: This figure compares the performance of three methods, standard ResNeXt, enhanced pooling mapping, and dual pooling mapping, across four key performance metrics. In terms of feature extraction accuracy, standard ResNeXt achieved 65.3%, enhanced pooling mapping improved to 78.1%, and dual pooling mapping reached a peak of 86.2%. In terms of coordinate prediction accuracy, standard ResNeXt achieved 71.4%, enhanced pooling mapping improved to 83.2%, and dual pooling mapping achieved the best performance at 89.5%. In terms of noise immunity, standard ResNeXt achieved 58.7%, enhanced pooling mapping significantly improved to 75.3%, and dual pooling mapping achieved an excellent level of 91.8%. In 3D reconstruction quality assessment, standard ResNeXt achieved 76.2%, enhanced pooling mapping improved to 84.9%, and dual pooling mapping reached the highest level at 93.7%. The data shows that the dual mixed pooling mapping method significantly outperforms the other two methods in all four evaluation indicators, especially in noise resistance and three-dimensional reconstruction quality, which are 33.1 and 17.5 percentage points higher than the standard method respectively, reflecting the significant advantages of this method in feature extraction and performance optimization.

[0088] According to a second aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0089] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0090] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-branch wind turbine blade breakpoint intelligent positioning method based on nanosecond pulses and superconductors, characterized in that: include: Collecting three-dimensional structural data of a multi-branch blade, calculating an optimal scanning path for an annular distribution based on the three-dimensional structural data; controlling a nanosecond pulse array to emit nanosecond pulse signals along the optimal scanning path, acquiring reflected signals through superconducting detection units pre-deployed on the blade surface, the reflected signals containing time-series intensity data; and spatially mapping the time-series intensity data to obtain an initial sound field distribution map; Constructing a volume integral formula of the initial sound field distribution map, and obtaining the sound field distribution by calculating the volume integral formula; Dividing the sound wave propagation path into multiple propagation intervals, calculating the sound wave attenuation coefficient of each propagation interval, wherein the sound wave attenuation coefficient is determined by the attenuation rate of the sound field intensity in the interval; enhancing the sound field signal based on the sound wave attenuation coefficient to obtain enhanced sound field characteristics; Performing spatial scanning on the enhanced sound field characteristics, and obtaining the initial breakpoint position coordinates by calculating the spatial distribution gradient of the sound field intensity; Establishing a detection area with the initial breakpoint position coordinate as the center, and combining the amplitude data, phase data and spatial position data of the time-series acoustic signal within the detection area to form a feature data set; The feature data set is input into a pre-trained ResNeXt neural network, which obtains the precise spatial coordinates of the breakpoints through multi-layer feature extraction, and constructs a three-dimensional stereoscopic image of the breakpoint area according to the precise spatial coordinates.

2. The method according to claim 1, characterized in that Performing spatial mapping on the time series intensity data to obtain an initial sound field distribution diagram includes: Calculating the maximum value and the minimum value of the time series intensity data, and obtaining normalized time series intensity data by taking the ratio of the difference between the time series intensity data and the maximum value and the minimum value; Constructing a spatiotemporal mapping matrix based on the normalized time series intensity data, setting spatiotemporal transformation weights according to the importance of data at different moments, and performing cumulative operation on the normalized time series intensity data and the spatiotemporal transformation weights to obtain a spatiotemporal mapping matrix containing spatial coordinate points; A spatial scale factor is determined according to the distribution density of the spatial coordinate points. The spatial scale factor is used to adjust the distribution intensity of the space-time mapping matrix at different spatial positions. The space-time mapping matrix is ​​multiplied by the spatial scale factor to obtain an initial sound field distribution map.

3. The method according to claim 1, characterized in that Constructing a volume integral formula of the initial sound field distribution map, and obtaining the sound field distribution by calculating the volume integral; Dividing the sound wave propagation path into a plurality of propagation intervals, and calculating the sound wave attenuation coefficient of each propagation interval includes: An improved Kirchhoff integral model is established based on the initial sound field distribution map, wherein the improved Kirchhoff integral model uses a second-order sound pressure derivative term to correct the integral expression; based on the improved Kirchhoff integral model, the surface of the multi-branch blade is divided into grid units, the sound pressure distribution is calculated within the grid units using a quadratic interpolation function, a topological relationship matrix between the grid units is established based on the sound pressure distribution, and a stress continuity boundary condition is applied at the branch connection of the multi-branch blade; A sound ray iterative equation group is established based on the topological relationship matrix, and a sound ray trajectory is obtained by solving the sound ray iterative equation group through step-by-step iteration; a deflection angle of the sound wave on the blade surface is calculated according to the sound ray trajectory, and a propagation path of the sound wave in the blade is calculated using the sound ray iterative equation group; Calculating a geometric diffusion attenuation coefficient based on the propagation path, wherein the geometric diffusion attenuation coefficient is related to the logarithmic value of the propagation distance; calculating a material absorption attenuation coefficient based on the sound wave frequency and temperature, wherein the material absorption attenuation coefficient is related to the power function of the frequency and the exponential function of the temperature; and calculating a scattering loss coefficient based on the scattering angle, wherein the scattering loss coefficient is related to the exponential function of the scattering cross section and the propagation distance; The geometric diffusion attenuation coefficient, the material absorption attenuation coefficient and the scattering loss coefficient are combined to obtain the acoustic wave attenuation coefficient.

4. The method according to claim 3, characterized in that Establishing an acoustic ray iterative equation group based on the topological relationship matrix, obtaining an acoustic ray trajectory by solving the acoustic ray iterative equation group through step-by-step iteration; calculating the deflection angle of the acoustic wave on the blade surface according to the acoustic ray trajectory, and calculating the propagation path of the acoustic wave in the blade using the acoustic ray iterative equation group includes: The sound ray iteration equation group includes a position vector and a direction vector. By multiplying the position vector and the direction vector and combining them with the sound speed, the dynamic change characteristics of the sound ray during the propagation process are obtained; Performing multiple update calculations on the sound ray iterative equation group, obtaining the position at the next moment by multiplying the current position vector by the sound speed, and obtaining the direction at the next moment by multiplying the current direction vector by the sound speed change, and determining that the calculation is complete when the position change between two consecutive calculations is less than a preset value; Determining the incident direction of the sound wave on the blade surface according to the calculated sound ray position, calculating the incident angle between the incident direction and the vertical direction of the blade surface, and determining the reflection direction of the sound wave based on the incident angle and the vertical direction; The propagation trajectory of the sound wave in the blade is tracked using the reflection direction, and the total propagation distance is obtained by calculating the superposition length of each path during the sound wave propagation process. The propagation path of the sound wave in the blade is obtained based on the total propagation distance.

5. The method according to claim 1, wherein Performing enhancement processing on the sound field signal based on the sound wave attenuation coefficient to obtain enhanced sound field characteristics; Performing spatial scanning on the enhanced sound field characteristics and obtaining the initial breakpoint position coordinates by calculating the spatial distribution gradient of the sound field intensity includes: Acquiring a propagation distance of each measurement point based on an acoustic wave sensor array, constructing an exponential function between the acoustic wave attenuation coefficient and the propagation distance, and multiplying the acoustic wave attenuation function by the exponential function to obtain an enhanced sound field signal, wherein the enhanced sound field signal includes compensation information for acoustic wave attenuation; Constructing a two-dimensional Gaussian function as a filter kernel function, determining a variance parameter of the filter kernel function according to the spatial distribution characteristics of the sound field, performing a discrete convolution operation on the enhanced sound field signal and the filter kernel function on a two-dimensional plane, and obtaining filtered sound field characteristics through the discrete convolution operation; Establishing a local coordinate system for the filtered sound field characteristics, calculating the transverse and longitudinal differential values ​​in the local coordinate system to obtain partial derivatives, constructing a two-dimensional vector from the transverse partial derivatives and the longitudinal partial derivatives to obtain a sound field intensity gradient, and taking the square root of the sum of the transverse and longitudinal components of the sound field intensity gradient to obtain a gradient amplitude; A threshold parameter is set according to the statistical characteristics of the spatial distribution of the sound field, and the gradient amplitude is compared with the threshold parameter to obtain a breakpoint mark. The breakpoint mark is used to identify the location of the gradient mutation. The breakpoint mark is multiplied by the gradient amplitude to obtain an effective gradient distribution. The spatial coordinates corresponding to the maximum value in the effective gradient distribution are extracted as the initial breakpoint position coordinates.

6. The method according to claim 1, characterized in that Establishing a detection area with the initial breakpoint position coordinate as the center, and combining the amplitude data, phase data, and spatial position data of the time-series acoustic signal within the detection area to form a feature data set includes: A detection area is established with the coordinates of the initial breakpoint position as a reference point, wherein the radius of the spherical detection area is dynamically determined according to the wavelength of the acoustic signal; a time-series acoustic signal is collected within the detection area, amplitude data and phase data of the time-series acoustic signal at each sampling point are extracted, and the amplitude data and phase data are converted into a position vector relative to the reference point; The amplitude data, the phase data and the position vector are combined to construct a feature data set, wherein each data point in the feature data set contains an amplitude component, a phase component and a three-dimensional spatial component of a time-series acoustic signal.

7. The method according to claim 1, characterized in that Inputting the feature data set into a pre-trained ResNeXt neural network, the ResNeXt neural network obtains the precise spatial coordinates of the breakpoints through multi-layer feature extraction, and constructing a three-dimensional stereoscopic image of the breakpoint area according to the precise spatial coordinates includes: Constructing a ResNeXt backbone network, setting a feature conversion function based on a cardinality parameter of the ResNeXt backbone network, decomposing input features into multiple feature branches through the feature conversion function, and obtaining an initial feature expression through weighted combination; A nonlinear chaotic mapping module is embedded in the ResNeXt backbone network, wherein the nonlinear chaotic mapping module adjusts the feature iteration state by controlling parameters, maps the initial feature expression to the phase space of the chaotic dynamic system, and obtains a chaotic feature expression; Reconstructing the trajectory of the chaotic characteristic expression in the phase space, evaluating the stability of the chaotic characteristic expression by calculating the Lyapunov exponent, and dynamically adjusting the control parameter based on the Lyapunov exponent; The features extracted by the ResNeXt backbone network are fused with the chaotic feature expression, the fusion ratio of the two types of features is adjusted by the chaotic feature weight, and the fused features are input into the nonlinear chaotic mapping module for secondary mapping; The spatial coordinates of the breakpoints are predicted based on the features after the secondary mapping, the stability of the prediction results is evaluated using the Lyapunov exponent, and a three-dimensional image of the breakpoint area is constructed according to the spatial coordinates.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images