Ultrasonic and electromagnetic composite detection method for internal defects of generator rotor forgings

Abstract

The present application relates to the technical field of defect detection, in particular to a kind of generator rotor forge internal defect ultrasonic and electromagnetic composite detection method, first to the surface of forge is adaptively path electromagnetic scanning, utilize multi-frequency eddy current technology and intelligent signal analysis identification and parameterization mark near-surface suspicious area;Subsequently, based on suspicious area information, through dynamic trigger mechanism executes global scanning and local directional precision scanning, generates the defect point cloud dataset containing multidimensional attribute;Finally, through three-level fusion architecture, electromagnetic and ultrasonic data are spatio-temporal alignment and feature-level correlation, and the type of defect is accurately determined and visualized output;The present application realizes the composite detection of the internal defect of generator rotor forge, can simultaneously improve the detection rate, positioning accuracy and type identification ability of defect.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and in particular to a method for combined ultrasonic and electromagnetic detection of internal defects in generator rotor forgings. Background Technology

[0002] As a core component of large power generation equipment, the internal quality of generator rotor forgings directly affects the safety and reliability of the generator. During the forging process, internal defects such as inclusions, cracks, and porosity may occur due to improper metallurgical, forging, or heat treatment processes. Currently, industry primarily employs single non-destructive testing methods for defect detection. For example, ultrasonic testing is sensitive to volumetric defects but has limited ability to detect fine cracks or near-surface defects; while electromagnetic testing (such as eddy current testing) is sensitive to surface and near-surface defects but struggles to detect deep internal defects.

[0003] Existing single detection methods cannot simultaneously meet the requirements of high sensitivity, full depth coverage and accurate classification of internal defects in generator rotor forgings. In particular, it is difficult to effectively distinguish defect types (such as cracks and non-metallic inclusions) and their spatial distribution, resulting in incomplete detection results and affecting the quantitative assessment of defects and maintenance decisions. Summary of the Invention

[0004] The purpose of this invention is to provide a combined ultrasonic and electromagnetic detection method for internal defects in generator rotor forgings, which enables combined detection of internal defects in generator rotor forgings and can simultaneously improve the detection rate, positioning accuracy and type identification ability of defects.

[0005] To achieve the above objectives, the present invention provides a method for combined ultrasonic and electromagnetic detection of internal defects in generator rotor forgings, comprising the following steps:

[0006] Based on the three-dimensional model of the rotor forging, the eddy current effect detection method is used to perform electromagnetic scanning by covering the surface of the forging with an adaptive path. Abnormal signals are identified by dynamic threshold and multi-parameter pattern recognition algorithm, and the coordinates of the regions corresponding to the abnormal signals are marked.

[0007] Based on the information of the marked region, full waveform radio frequency data is collected according to the master-slave detection method of global conventional scanning and local directional scanning. After preprocessing and feature extraction, a defect point cloud dataset containing spatial coordinates and multi-dimensional attributes is generated.

[0008] The marked region is fused with the defect point cloud dataset at the feature level. The defect comprehensive feature vector of each fused defect ID is analyzed to determine the defect type and output a three-dimensional fusion detection report.

[0009] Among them, based on the three-dimensional model of the rotor forging, an eddy current effect detection method is used to perform electromagnetic scanning by covering the surface of the forging with an adaptive path. Abnormal signals are identified through dynamic thresholding and multi-parameter pattern recognition algorithms, and the coordinates of the regions corresponding to the abnormal signals are marked, including:

[0010] Based on the three-dimensional model of the rotor forging, the electromagnetic detection unit is controlled to cover the surface of the forging with an adaptive path;

[0011] A multi-frequency eddy current probe is used to excite an alternating magnetic field, and complex impedance signals are acquired simultaneously.

[0012] The complex impedance signal is processed in real time. Abnormal signals are identified by dynamic threshold and multi-parameter pattern recognition algorithm. Abnormal signal points within a set spatial range are aggregated and parameterized using spatial clustering algorithm to obtain the marked region.

[0013] The method further includes, after obtaining the marked region:

[0014] Simultaneously, the center coordinates, boundaries, defect confidence level, and preliminary type tendency of each marked area are recorded.

[0015] Specifically, the adaptive path planning is as follows:

[0016] The outer cylindrical surface of the forging is unfolded into a two-dimensional plane, and a helical line with equal pitch is planned as the basic scanning path.

[0017] Based on the geometric features identified from the 3D CAD model of the forging, the system automatically switches to encrypted grid scanning in the journal and slot areas to ensure that the probe is vertical and the scan covers the entire area.

[0018] The multi-parameter pattern recognition algorithm specifically includes:

[0019] The phase angle differences, shape complexity of the impedance plane trajectory, and harmonic components of the complex impedance signal at multiple excitation frequencies are analyzed.

[0020] By matching the extracted feature parameters with the pre-stored defect feature spectrum, the defect confidence level is calculated, and a preliminary tendency judgment is made on the crack and inclusion types.

[0021] Based on the information from the marked regions, full-waveform radio frequency data is acquired using a master-slave detection method that combines global conventional scanning and local directional scanning. After preprocessing and feature extraction, a defect point cloud dataset containing spatial coordinates and multi-dimensional attributes is generated, including:

[0022] Based on the information of the marked region, the ultrasound detection unit is controlled to perform global conventional scanning and local directional scanning.

[0023] When the ultrasound probe enters the dynamic trigger boundary of the marked area, the local directional scanning mode is triggered, and high-density grid scanning or sector scanning is performed.

[0024] During the scanning process, full-waveform radio frequency data is acquired. After time-domain gain compensation and coherent composite noise reduction, the three-dimensional image of the defect is reconstructed by the full-focusing method.

[0025] Extract geometric, texture, and frequency domain features to generate a defect point cloud dataset containing spatial coordinates and multidimensional attributes.

[0026] The method for generating the dynamic trigger boundary is as follows:

[0027] A buffer zone is extended outward from the original boundary of the marked area, and its width is dynamically calculated based on the estimated size of the marked area.

[0028] When the ultrasound probe enters the buffer zone, it is marked as prepositioned. When it further enters the core region of the original boundary, it formally triggers local directional scanning. When exiting, the local directional scanning ends when it completely leaves the buffer zone.

[0029] The generated defect point cloud dataset includes:

[0030] The sparse defect indicator points from the main scan and the 3D reconstructed point cloud from the local directional scan are merged and registered using an iterative nearest point algorithm;

[0031] The detection blind zone is filled using the Poisson surface reconstruction algorithm, and finally structured point cloud data containing spatial coordinates, local normal vectors, scattering intensity and frequency response attributes are generated.

[0032] The process of fusing the marked regions with the defect point cloud dataset at the feature level includes:

[0033] Calculate the planar projection distance between the center of the marked region and the defect point cloud dataset, the angle between the electromagnetic inference orientation and the ultrasonic principal component analysis orientation, and the ratio between the near-surface size and the ultrasonic cross-sectional size;

[0034] When all conditions are met, they are determined to be the same defective entity and assigned a unique fusion defect ID.

[0035] The analysis includes the comprehensive feature vector of each generated fusion defect ID, including:

[0036] Initial screening is conducted based on electromagnetic anisotropy, ultrasonic image texture, and three-dimensional spread ratio features in the defect comprehensive feature vector.

[0037] For defects that fail the initial screening or have low confidence, they are compared with a pre-set historical defect case library. A graph neural network is used to find matching cases, and the final classification result and confidence score are obtained by combining the results.

[0038] This invention discloses a method for combined ultrasonic and electromagnetic detection of internal defects in generator rotor forgings. First, an adaptive path electromagnetic scan is performed on the forging surface. Multi-frequency eddy current technology and intelligent signal analysis are used to identify and parameterize suspicious near-surface regions. Then, based on the suspicious region information, a dynamic triggering mechanism is used to perform global scanning and local directional fine scanning, generating a defect point cloud dataset containing multi-dimensional attributes. Finally, a three-level fusion architecture is used to perform spatiotemporal alignment and feature-level correlation of the electromagnetic and ultrasonic data, and to accurately determine and visualize the defect type. This invention achieves combined detection of internal defects in generator rotor forgings, simultaneously improving defect detection rate, location accuracy, and type identification capability. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0040] Figure 1 This is a schematic diagram of the steps of a method for combined ultrasonic and electromagnetic detection of internal defects in a generator rotor forging according to the first embodiment of the present invention.

[0041] Figure 2 This is a flowchart illustrating a method for combined ultrasonic and electromagnetic detection of internal defects in generator rotor forgings provided by the present invention.

[0042] Figure 3 This is a flowchart of a method for combined ultrasonic and electromagnetic detection of internal defects in generator rotor forgings provided by the present invention. Detailed Implementation

[0043] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0044] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0045] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0046] Please see Figures 1-3 This invention provides a method for combined ultrasonic and electromagnetic detection of internal defects in generator rotor forgings, comprising the following steps:

[0047] S101. Based on the three-dimensional model of the rotor forging, the eddy current effect detection method is used to perform electromagnetic scanning by covering the surface of the forging with an adaptive path. Abnormal signals are identified by dynamic threshold and multi-parameter pattern recognition algorithm, and the coordinates of the area corresponding to the abnormal signal are marked.

[0048] Specifically, the electromagnetic detection unit consists of an eddy current probe, a multi-frequency excitation source, an impedance analysis module, a motion control unit, and a data acquisition and processing unit. Its core control unit is the motion control unit, which receives scanning commands from the host computer and converts them into precise control signals for the probe's motion mechanism (usually a high-precision three-dimensional robotic arm or a gantry-type scanning frame).

[0049] Based on the input 3D CAD model, the detection path is automatically planned. For regular cylindrical surfaces, a constant-pitch helical path is used to ensure full coverage (typically 10%-20% of the probe diameter). For areas with abrupt geometric changes such as journals and keyways, the adaptive path planning algorithm automatically switches to a finer grid scanning mode and adjusts the probe posture to maintain perpendicularity to the surface. The motion control unit drives the scanning mechanism equipped with the eddy current probe to move precisely along this path and synchronizes the position coordinates and detection data in real time. During the scanning process, the motion control unit synchronously sends the encoder position coordinates (X, Y, Z, and rotation angle) of the probe motion mechanism to the data acquisition and processing unit in real time, achieving a strict one-to-one correspondence between the detection data and the spatial position.

[0050] The eddy current effect is the physical basis of this detection method. When an eddy current probe carrying a high-frequency alternating current approaches the surface of a conductive rotor forging, the probe coil generates an alternating magnetic field. This magnetic field induces vortex-shaped closed currents, i.e., eddy currents, inside the forging. Defects inside the forging (such as cracks and inclusions) disrupt the normal distribution path of the eddy currents, causing changes in their size, phase, and flow pattern. This change, in turn, affects the impedance (i.e., the combined characteristics of resistance and inductive reactance) of the eddy current probe coil. The impedance analysis module indirectly senses the presence of surface and near-surface defects by accurately measuring the amplitude and phase changes of the probe coil impedance and converting them into a voltage signal output.

[0051] Anomalies are identified using a dynamic threshold comparison method and multi-parameter pattern recognition. The dynamic threshold involves initially acquiring a background signal in a region deemed defect-free during the initial scanning phase, calculating its statistical mean and standard deviation. In subsequent scans, if the amplitude or phase of the real-time signal exceeds "background mean ± N times the standard deviation" (N is typically 3-5, adjustable according to detection standards), it is preliminarily marked as an "anomaly." Simultaneously, multiple characteristic parameters of the signal are analyzed, including the shape of the impedance plane trajectory, the signal's zero-crossing frequency, and phase angle differences at multiple excitation frequencies. For example, deep cracks exhibit specific phase responses at different frequencies, while shallow scratches do not. The system's built-in defect feature library matches these real-time extracted feature parameters with the feature spectra of typical defects (such as cracks, inclusions, and folds), providing a "defect confidence" score and a preliminary type tendency (e.g., suspected crack).

[0052] Subsequently, the spatial clustering algorithm aggregates anomalies that are adjacent in location and have similar signal characteristics, forming independent suspicious regions. Each region is parameterized and labeled, and the resulting suspicious regions are called labeled regions. The region center coordinates (calculated based on the spatial positions of all cluster points), region boundaries (the smallest rectangle or ellipse that can completely enclose the region), and region feature vectors (containing the average amplitude, dominant phase, characteristic frequency response, and calculated defect confidence and preliminary type tendency of the signal within the region) are recorded. A highlighted defect distribution heatmap is then generated in the software interface.

[0053] S102. Based on the information of the marked region, collect full waveform radio frequency data in a master-slave detection method of global conventional scanning and local directional scanning. After preprocessing and feature extraction, generate a defect point cloud dataset containing spatial coordinates and multi-dimensional attributes.

[0054] Specifically, based on the center coordinates, boundaries, defect confidence level, and preliminary type tendency of each marked area recorded in step S101, the ultrasonic probe is controlled to move at a constant speed along a preset axial spiral scanning path covering the entire length of the forging. This is the "main scan," which realizes overall routine inspection. Furthermore, a dynamic trigger boundary (an intelligent buffer zone extending from the original boundary) is set for each suspicious area. When the motion control unit detects that the ultrasonic probe has entered this boundary, it immediately saves the main path interruption point state and triggers the secondary scan mode.

[0055] In "from scan" mode, the motion control unit first fine-tunes the probe's orientation to ensure that its acoustic beam central axis is precisely aligned with the center coordinates of the suspected area. Subsequently, the probe automatically performs a high-density, small-area, localized fine scan over this area. This scan typically employs either a fan-shaped scan or a grid scan mode: a fan-shaped scan uses electronic deflection or mechanical oscillation to illuminate the interior of the suspected area at different angles, which is beneficial for detecting cracks with different orientations; a grid scan performs a dense array of probes above the area to precisely delineate the planar projection shape of the defect.

[0056] However, for suspected areas that are deep (>100mm) or large in area, a single fixed focus cannot guarantee detection resolution across the entire defect volume. The following focusing strategy should be used:

[0057] Segmented focusing detection:

[0058] This technique is suitable for defects that extend significantly in the depth direction. The entire suspected area is divided into several consecutive focusing segments along the depth direction. For example, for an area extending from a depth of 50mm to 200mm, it can be divided into three focusing segments: 50-100mm, 100-150mm, and 150-200mm. At each spatial point where a fine local scan is performed on this area, the ultrasonic testing unit emits three times sequentially. With each emission, the electronic focusing delay law of the multi-band focusing probe is reconfigured, precisely setting the acoustic beam focus at the center depth of the current focusing segment. The A-scan signals received from the three emissions are stored independently, and their corresponding effective focusing depth segments are marked. Subsequent signal processing and imaging (such as TFM) will automatically retrieve the data from the corresponding focusing segment for high-resolution analysis based on the depth of the defect echo, thereby ensuring that the defect is clearly imaged throughout its entire depth range.

[0059] Multifocal focusing technology:

[0060] This technology is more suitable for areas with significant depth and horizontal reach, aiming to achieve uniform sensitivity across a greater depth range in a single transmission. Through sophisticated electronic aperture control and transmission waveform design, the multi-band focusing probe simultaneously generates multiple acoustic beams with different focal points in a single pulse excitation. For example, it can be configured to simultaneously generate three acoustic beams with focal lengths of 80mm, 140mm, and 200mm. To achieve signal separation, depth-coded transmission technology is employed. This involves assigning unique coding characteristics (such as specific frequency modulation or phase coding) to the transmission pulses at different focal points. The receiver processes the entire waveform signal using a digital decoder, effectively separating the scattered echoes from different focal points in the time or frequency domain by utilizing the correlation of different codes, thereby obtaining independent signal channel data corresponding to different depth focal points.

[0061] Once the scan mode is triggered, the motion control unit immediately executes:

[0062] Main path interruption and status saving: accurately records the coordinates of the interruption point on the current main scanning path and the probe's movement status (speed, direction), and smoothly decelerates and stops.

[0063] Fine scanning path planning: Automatically select the optimal local scanning mode based on the shape, size, and initial type tendency of the suspicious area.

[0064] Fan-shaped scanning mode: suitable for suspected linear defects (such as cracks). The probe is fixed above the center of the area and scans the sound beam in small angular steps within a preset angle range (such as ±30°) through electronic deflection or precision mechanical oscillation mechanism.

[0065] High-density grid scanning mode: suitable for volumetric defects with complex shapes. The probe is planned to scan line by line on a planar grid covering the area, with the grid point spacing much smaller than the main scan point spacing (e.g., 5 mm for the main scan and 1 mm for the grid scan).

[0066] The motion control unit then drives the probe to move quickly and precisely to the starting point of the local scan, loading the preset dynamic focusing scheme (segmented or multi-focus) and frequency parameters for that area. The probe moves along the planned path, emitting ultrasonic waves and receiving echoes at each point. All raw RF data, corresponding precise spatial coordinates, and transmit / receive parameters are simultaneously acquired, packaged, and labeled with the corresponding fusion defect ID.

[0067] After a local scan is completed, the system does not simply return the probe to the spatial coordinates of the interruption point in a straight line. Because the main scan is a continuous motion process, directly returning to the coordinate point might cause inconsistencies in the probe's posture and direction of motion. The motion control unit calculates the optimal smooth transition path from the current local scan end point back to the main path based on the saved motion state (velocity vector) of the interruption point and the mathematical model of the main scan path. This path is typically a spline curve tangent to the main path or with continuous curvature, ensuring that the probe can reintegrate into the main scan flow with the appropriate speed and direction. The probe moves at high speed along the calculated transition path, and its speed and direction are adjusted to be completely consistent with those at the time of the interruption when it approaches the predicted position of the interruption point. Subsequently, the system immediately resumes the transmission of the main scan command from the interruption point and continues the overall routine scan. The entire process is smooth and efficient. In the main scan data stream, high-density local scan data blocks are inserted only in the time period corresponding to the region, achieving seamless integration of "macroscopic survey" and "microscopic detailed survey" at the data and motion levels.

[0068] Main scan data stream: A series of ultrasonic A-scan waveforms are acquired at a low spatial sampling rate (e.g., 5 mm axially, 2° circumferentially), with each waveform corresponding to a perpendicular incident detection at a fixed point on the forging surface by the probe. The data is stored in continuous data blocks, accompanied by timestamps and spatial coordinates.

[0069] From the scan data stream: a dense ultrasound dataset is formed by high-density spatial sampling (e.g., 1mm × 1mm grid) or multi-angle acquisition within the suspicious area. Each data point contains not only the A-scan waveform, but also beam angle information, focusing parameters, and emission coding markers.

[0070] The preprocessing process includes:

[0071] Temporal gain compensation: Depth-adaptive gain curves are applied to the raw RF signals of both scanning modes. These curves are pre-generated based on the acoustic attenuation characteristics of the material and fine-tuned during scanning according to the actual signal-to-noise ratio to ensure that defects with the same reflection intensity at different depths exhibit similar amplitudes.

[0072] Noise suppression processing:

[0073] Main scan data: Adaptive wavelet threshold denoising is used to effectively suppress grass-like noise caused by microscopic scattering of materials while preserving the echoes of sudden defects.

[0074] From the scanned data: For multi-angle acquisition, coherent composite noise reduction technology is employed. Signals from the same spatial point but with different incident angles are phase-aligned and then averaged, significantly improving the signal-to-noise ratio and enhancing the detection capability for tilting defects.

[0075] Encoded signal separation (for multi-focusing): For signals using deep coding transmission technology, a matched filter decoder separates the independent signal channels corresponding to different focal points.

[0076] For the preprocessed master scan data, a dual-threshold comparison algorithm is applied to each A scan. First, a lower sensitivity threshold is used to initially screen for potential defects. Then, an intelligent confirmation algorithm based on waveform complexity analysis is applied to the initially screened signals to eliminate false defect signals. The following features are extracted from the confirmed defect signals:

[0077] Depth location: calculated through precise acoustic time measurement and sound velocity calibration.

[0078] Peak amplitude: The peak value of the echo after standardization.

[0079] Equivalent size: The estimated equivalent flat-bottom hole diameter for defects, obtained by comparing with the distance-amplitude correction curve.

[0080] Three-dimensional density clustering is performed on spatially adjacent and depth-similar defect indicator points to form preliminary defect clusters. Each cluster records its approximate spatial extent and the number of defect points.

[0081] For the preprocessed scan data, a full-matrix capture method is used to record all transmit-receive combinations for the raster scan data. A full-focusing algorithm is used to reconstruct the 3D scattering intensity distribution map of the suspected area, forming TFM 3D volumetric data. Anisotropic diffusion filtering is applied to the TFM volumetric data to enhance the defect boundaries. An adaptive level set segmentation algorithm is used to extract the complete 3D contour of the defect from the background, overcoming the shortcomings of traditional threshold segmentation for weak boundary defects. Feature extraction:

[0082] Geometric features: Calculate the volume, surface area, aspect ratio, principal axis direction, etc. of the defect based on the segmentation results.

[0083] Texture features: Analyze the distribution characteristics of scattering intensity inside defects, such as uniformity, contrast, and correlation.

[0084] Frequency response characteristics: Analyze the spectral characteristics of defect echoes. Different material defects (such as inclusion types) have different frequency response characteristics.

[0085] Defect indicator points discovered in the main scan are converted into sparse point clouds, with each point containing coordinates (x, y, z) and basic features (amplitude, confidence level). The defect surface contours extracted from the TFM imaging are discretized into dense surface point clouds, with each point containing coordinates and local normal information. Inverse scattering analysis is performed on multi-angle acquired data to extract strong scattering centers within the defects, forming an internal structure point cloud. The coordinates of all data sources are unified to the global detection coordinate system. Iterative nearest-point registration is performed between the sparse point cloud from the main scan and the dense point cloud from the secondary scan to ensure spatial consistency.

[0086] A feature-based point cloud fusion algorithm is employed to combine the broad coverage of sparse point clouds with the detailed precision of dense point clouds. For partially occluded or detection blind spots on defective surfaces, a Poisson surface reconstruction algorithm is applied for appropriate completion, generating a complete defective surface model. Each point cloud point is assigned a multi-dimensional attribute vector, including:

[0087] Spatial properties: 3D coordinates, local curvature, normal vector.

[0088] Acoustic properties: scattering intensity, frequency response characteristics, signal attenuation coefficient.

[0089] Geometric relationship attributes: distance to the defect center, distance to the defect surface (for internal points).

[0090] The final generated defect point cloud dataset is a structured data object containing:

[0091] Point cloud geometric data:

[0092] Surface point cloud: tens of thousands to millions of points, accurately describing the three-dimensional surface morphology of defects.

[0093] Internal point cloud: A set of points that describes the internal structure of a defect, suitable for porous defects.

[0094] Point cloud attribute data:

[0095] Each point contains spatial coordinates (x, y, z), reflection intensity value, and local normal vector (nx, ny, nz).

[0096] Additional attributes include material acoustic impedance variation estimation and scattering type identification (specular / diffuse reflection).

[0097] Global characteristics of the defect:

[0098] Bounding box parameters: minimum / maximum coordinates, center position.

[0099] Equivalent ellipsoid parameters: the length and orientation of the three axes obtained through principal component analysis.

[0100] Volume and surface area: precise values ​​obtained through point cloud computing.

[0101] Orientation distribution diagram: The statistical distribution of the surface normals of the defect, reflecting the overall orientation characteristics of the defect.

[0102] Quality assessment indicators:

[0103] Point cloud integrity score: a quality assessment based on detection coverage and interpolation ratio.

[0104] Confidence distribution map: Heatmap of detection confidence in each region of the point cloud.

[0105] S103. Perform feature-level fusion of the marked region and the defect point cloud dataset, analyze the defect comprehensive feature vector of each generated fused defect ID, determine the defect type, and output a three-dimensional fusion detection report.

[0106] Specifically, the coordinates of the suspicious areas from electromagnetic detection and the defect point cloud dataset from ultrasonic detection are both recorded in the same physical coordinate system of the forging (usually with the center of the rotor end face as the origin). The data fusion unit first performs coordinate transformation verification to ensure that all data points are located in a unified three-dimensional spatial reference system. Due to the non-synchronous scanning, the system aligns all detection events (electromagnetic signal sampling points, ultrasonic A-scan trigger points) to a unified time axis based on the high-precision time scale provided by the motion control unit, thereby establishing a temporal correlation of cross-modal data. Heterogeneous feature parameters from the two technologies are mapped to a unified metric scale. For example, the impedance change amplitude of the electromagnetic signal and the echo amplitude of the ultrasonic signal are normalized to a dimensionless relative response intensity between 0 and 1 through calibration coefficients; angular and frequency domain parameters such as phase and frequency are also transformed to a standardized parameter space.

[0107] The fusion process is not a simple superposition, but a progressive analysis process from the data level to the decision level, establishing a three-level fusion architecture of "feature matching-spatial correlation-decision verification".

[0108] Level 1: Feature-level fusion and correlation matching. The goal of this level is to correlate the surface / near-surface features of electromagnetic detection with the internal volume features of ultrasonic detection to confirm whether they originate from the same physical defect.

[0109] Spatial Correlation: For each suspicious area of ​​an electromagnetic detection marker, the data fusion unit extends its boundary outward in three-dimensional space by a preset "correlation tolerance range" (e.g., extending to 20 mm in the depth direction). The system searches for all ultrasonic defect point clouds falling within this range in space.

[0110] Feature consistency calculation: If associated ultrasonic point clouds are found, the system will calculate a set of feature consistency indices:

[0111] Positional consistency: The distance between the center of the electromagnetic region and the centroid of the ultrasonic point cloud on the plane projection.

[0112] Orientation consistency: The angle between the main orientation of the defect inferred from the anisotropic characteristics of the electromagnetic signal and the main orientation obtained from ultrasonic TFM reconstruction or point cloud principal component analysis.

[0113] Size trend consistency: The proportional relationship between the near-surface opening size of the defect estimated by electromagnetic signals and the near-surface cross-sectional size of the defect detected by ultrasonic testing.

[0114] Association Determination: When all the above consistency indicators meet the preset thresholds (e.g., distance < 5mm, included angle < 30°, size ratio between 0.5 and 2), the system determines that the electromagnetically suspicious area is highly correlated with the ultrasonic defect point cloud, belongs to the same defect entity, and generates a unique fusion defect ID. Otherwise, it is considered an independent anomaly and recorded separately.

[0115] Level Two: Decision-level fusion and defect classification. For each successfully associated fusion defect ID, the system enters the core classification and discrimination stage. A dual-drive classification engine based on rules and case-based reasoning is employed.

[0116] Multi-dimensional feature vector construction: A unified comprehensive feature vector for each fused defect ID is constructed. This vector encompasses:

[0117] Electromagnetic vectors: normalized impedance amplitude, characteristic phase angle, multi-frequency phase difference, and signal harmonic components.

[0118] Ultrasonic subvectors: depth, -6dB length and width, echo peak dynamic range, waveform complexity, TFM image texture features (such as contrast, uniformity), and the spread ratio (length / width) of the 3D point cloud.

[0119] Related derivative features: such as the various consistency indicators mentioned above.

[0120] Preliminary screening by the rule engine: The classification engine first runs a set of hard physical rules for rapid initial screening. For example:

[0121] Rule 1 (Crack Indication): If the electromagnetic signal exhibits high anisotropy (large differences in response in different scanning directions) and high harmonic components, and the ultrasound TFM image is elongated, deep, and oriented in accordance with the electromagnetic inference, then the "crack" high confidence indicator is triggered.

[0122] Rule 2 (Inclusion Indication): If the electromagnetic response is weak or absent, and the ultrasonic point cloud is clustered, the three-dimensional distribution ratio is close to 1, and the echo waveform is relatively simple, then the "non-metallic inclusion" indicator is triggered.

[0123] Rule 3 (Loose Indicator): If there are no electromagnetic anomalies, but ultrasound detects a group of spatially adjacent points with moderate echo amplitude, small individual size, and a diffuse point cloud distribution, then it may be "loose".

[0124] For complex cases where the rule engine cannot clearly classify the defect, or where features contradict each other, the system activates a case-based reasoning engine. This engine has a built-in, continuously learning historical defect feature case library, which stores a large number of defect cases verified by metallographic dissection and their complete comprehensive defect feature vectors. The engine uses the nearest neighbor algorithm or a more advanced graph neural network matching to find several historical cases in the case library that are most similar to the current defect feature vector to be judged. The final classification result integrates the known types of these similar cases and provides a classification confidence score based on similarity.

[0125] Level 3: Results Integration and 3D Visualization Output

[0126] Comprehensive report generation: Each fused defect ID is ultimately assigned a clear defect type (e.g., crack, non-metallic inclusion, porosity, fold, etc.), a comprehensive confidence score (integrating rule triggering strength and case matching degree), and precise three-dimensional spatial dimension information (length, width, depth, volume, main orientation).

[0127] 3D Visualization Fusion: A 3D map of internal defects in the rotor forging is generated on the interactive interface of the data fusion unit. This map displays the forging as a 3D model, where:

[0128] Different types of defects are rendered with different colors and icons (e.g., red stripes represent cracks, and blue spheres represent inclusions).

[0129] The map supports rotation, zooming, and cross-sectional viewing. You can click on any defect to view its comprehensive feature vector, classification criteria, and confidence level.

[0130] Meanwhile, the original electromagnetic C-scan image and the ultrasound TFM / point cloud slice image can be displayed as overlay layers, enabling the fusion process and results to be traceable and verifiable.

[0131] By combining ultrasonic and electromagnetic testing methods, a composite inspection of internal defects in generator rotor forgings has been achieved, simultaneously improving the defect detection rate, location accuracy, and type identification capability. The ultrasonic testing unit can detect deep volumetric defects, while the electromagnetic testing unit is sensitive to surface and near-surface cracks. After data fusion, a comprehensive assessment of defect depth, size, orientation, and nature can be achieved, significantly improving the reliability and accuracy of the inspection.

[0132] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0133] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for combined ultrasonic and electromagnetic detection of internal defects in generator rotor forgings, characterized in that, The process includes the following steps: based on a three-dimensional model of the rotor forging, an eddy current effect detection method is used to perform electromagnetic scanning by covering the surface of the forging with an adaptive path; abnormal signals are identified by a dynamic threshold and a multi-parameter pattern recognition algorithm; and the coordinates of the regions corresponding to the abnormal signals are marked. Based on the information of the marked region, full waveform radio frequency data is collected according to the master-slave detection method of global conventional scanning and local directional scanning. After preprocessing and feature extraction, a defect point cloud dataset containing spatial coordinates and multi-dimensional attributes is generated. The marked region is fused with the defect point cloud dataset at the feature level. The defect comprehensive feature vector of each fused defect ID is analyzed to determine the defect type and output a three-dimensional fusion detection report. Based on the three-dimensional model of the rotor forging, the electromagnetic scanning of the forging surface is carried out by using the eddy current effect detection method with an adaptive path. Abnormal signals are identified by dynamic threshold and multi-parameter pattern recognition algorithm, and the coordinates of the region corresponding to the abnormal signals are marked. The method includes: based on the three-dimensional model of the rotor forging, controlling the electromagnetic detection unit to cover the forging surface with an adaptive path. A multi-frequency eddy current probe is used to excite an alternating magnetic field, and complex impedance signals are acquired simultaneously. The complex impedance signal is processed in real time. Abnormal signals are identified by dynamic threshold and multi-parameter pattern recognition algorithm. Abnormal signal points within a set spatial range are aggregated and parameterized by spatial clustering algorithm to obtain the marked region. The adaptive path is specifically as follows: the outer cylindrical surface of the forging is unfolded into a two-dimensional plane, and a helical line with equal pitch is planned as the basic scanning path; based on the geometric features identified by the three-dimensional CAD model of the forging, the scanning path is automatically switched to a dense grid in the journal and slot areas to ensure that the probe is vertical and the scanning is fully covered. The marked region is fused with the defect point cloud dataset at the feature level, including: calculating the planar projection distance between the center of the marked region and the defect point cloud dataset, the angle between the electromagnetic inference orientation and the ultrasonic principal component analysis orientation, and the ratio between the near-surface size and the ultrasonic cross-sectional size; when all of them meet the preset threshold, they are determined to be the same defect entity and a unique fused defect ID is assigned.

2. The ultrasonic and electromagnetic combined detection method for internal defects in generator rotor forgings as described in claim 1, characterized in that, After obtaining the marked regions, the method further includes: simultaneously recording the center coordinates, boundaries, defect confidence level, and preliminary type tendency of each marked region.

3. The ultrasonic and electromagnetic combined detection method for internal defects in generator rotor forgings as described in claim 1, characterized in that, The multi-parameter pattern recognition algorithm specifically includes: analyzing the phase angle difference of complex impedance signals at multiple excitation frequencies, the shape complexity of impedance plane trajectories, and signal harmonic components; calculating the defect confidence level by matching the extracted feature parameters with the pre-stored defect feature spectrum, and making a preliminary tendency judgment on crack and inclusion types.

4. The ultrasonic and electromagnetic combined detection method for internal defects in generator rotor forgings as described in claim 1, characterized in that, Based on the information of the marked regions, full waveform radio frequency data is collected in a master-slave detection mode of global conventional scanning and local directional scanning. After preprocessing and feature extraction, a defect point cloud dataset containing spatial coordinates and multi-dimensional attributes is generated, including: controlling the ultrasonic detection unit to perform global conventional scanning and local directional scanning based on the information of the marked regions. When the ultrasonic probe enters the dynamic trigger boundary of the marked area, the local directional scanning mode is triggered, and high-density grid scanning or sector scanning is performed. During the scanning process, full waveform radio frequency data is acquired, and after time-domain gain compensation and coherent composite noise reduction, the three-dimensional image of the defect is reconstructed by the full focusing method. Extract geometric, texture, and frequency domain features to generate a defect point cloud dataset containing spatial coordinates and multidimensional attributes.

5. The ultrasonic and electromagnetic combined detection method for internal defects in generator rotor forgings as described in claim 4, characterized in that, The method for generating the dynamic trigger boundary is as follows: a buffer zone is extended outward from the original boundary of the marked area, and its width is dynamically calculated based on the estimated size of the marked area. When the ultrasound probe enters the buffer zone, it is marked as prepositioned. When it further enters the core region of the original boundary, it formally triggers local directional scanning. When exiting, the local directional scanning ends when it completely leaves the buffer zone.

6. The ultrasonic and electromagnetic combined detection method for internal defects in generator rotor forgings as described in claim 4, characterized in that, The generated defect point cloud dataset includes: sparse defect indicator points from the main scan and 3D reconstructed point clouds from local directional scans, which are then registered using an iterative nearest-point algorithm; The detection blind zone is filled using the Poisson surface reconstruction algorithm, and finally structured point cloud data containing spatial coordinates, local normal vectors, scattering intensity and frequency response attributes are generated.

7. The ultrasonic and electromagnetic combined detection method for internal defects in generator rotor forgings as described in claim 1, characterized in that, The comprehensive feature vector of each generated fusion defect ID is analyzed, including: initial screening based on electromagnetic anisotropy, ultrasonic image texture, and three-dimensional spread ratio features in the comprehensive feature vector; for defects that fail the initial screening or have low confidence, they are compared with a preset historical defect case library, and matching cases are found through graph neural network matching, and the final classification result and confidence score are obtained by combining the results.

Images