Fan bolt phased array ultrasonic artifact removal method based on generative adversarial network

Abstract

The application relates to the technical field of ultrasonic nondestructive testing, and particularly discloses a fan bolt phased array ultrasonic artifact removal method based on an adversarial generation network, which is applied to a control terminal connected with a phased array ultrasonic detection device, and the method comprises the following steps: in response to a starting signal of the phased array ultrasonic detection device, establishing a device reference library containing a probe initial impedance, a coupling agent initial thickness and a probe initial angle, and initializing an adversarial generation network model. In actual nondestructive testing, a cooperative closed-loop mechanism of a front-end detection device and a rear-end adversarial generation network is established, a state deviation value of a device physical state and a reference library is calculated in real time, and the state deviation value is used to synchronously and dynamically adjust acquisition parameters of bottom-layer hardware and constraint weights of an upper-layer software model.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic nondestructive testing technology, and in particular to a method for removing ultrasonic artifacts from phased array wind turbine bolts based on generative adversarial networks. Background Technology

[0002] In the non-destructive testing of wind turbine bolts, phased array ultrasonic testing equipment is often combined with deep learning networks, such as generative adversarial networks, to remove artifacts from ultrasonic images. Existing artifact removal methods are usually static and unidirectional, meaning that the algorithm model is fixed after being trained on an ideal dataset and is only used as a post-processing tool during actual testing. It directly receives the conventional signals collected by the front-end ultrasonic equipment and performs image reconstruction and denoising calculations.

[0003] However, in actual field applications, due to the complex environment of wind turbine sites, physical states such as probe angle deflection, coupling agent loss, and probe wear change in real time, causing dynamic shifts in the characteristic distribution of the underlying ultrasonic signal. Existing fixed models cannot perceive changes in the physical state of the front-end hardware. Forcibly applying static logic to offset signals may lead to the algorithm over-filtering out real, minute defect features or missing dynamically changing artifacts, resulting in a disconnect between hardware physical acquisition and software algorithm processing. Summary of the Invention

[0004] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a method for removing phased array ultrasonic artifacts from wind turbine bolts based on generative adversarial networks, thereby improving the reliability of on-site defect detection and the safety of equipment operation.

[0005] To achieve the above objectives, a first aspect of the present invention proposes a method for removing phased array ultrasonic artifacts from wind turbine bolts based on generative adversarial networks, applied in a control terminal connected to a phased array ultrasonic testing device. The method includes: In response to the start signal of the phased array ultrasonic testing device, a device reference library containing the initial impedance of the probe, the initial thickness of the coupling agent, and the initial angle of the probe is established, and the generative adversarial network model is initialized. Acquire real-time ultrasonic data and real-time status parameters collected by the phased array ultrasonic testing equipment, and calculate the status deviation value between the real-time status parameters and the corresponding parameters in the equipment reference library; The acquisition parameters of the phased array ultrasonic testing device are dynamically adjusted according to the state deviation value in order to obtain target ultrasonic data. The constraint weights are determined based on the state deviation values, and the target ultrasound data is input into the adversarial generative network model to generate an ultrasound image with artifacts removed by combining the constraint weights. Multi-dimensional physical feature verification is performed based on the real-time state parameters and the artifact-removed ultrasound image; If the verification passes, the artifact-free ultrasound image is output, and a reset command is sent to the phased array ultrasound testing device to restore the parameters in the device's reference library.

[0006] To achieve the above objectives, a second aspect of the present invention proposes a wind turbine bolt phased array ultrasonic artifact removal system based on generative adversarial networks, applied in a control terminal connected to a phased array ultrasonic testing device. The system includes: An initialization module is used to respond to the start signal of the phased array ultrasonic testing device, establish a device reference library including the probe initial impedance, the initial thickness of the coupling agent and the probe initial angle, and initialize the generative adversarial network model. The deviation calculation module is used to acquire real-time ultrasonic data and real-time status parameters collected by the phased array ultrasonic testing equipment, and to calculate the status deviation value between the real-time status parameters and the corresponding parameters in the equipment reference library. The parameter adjustment module is used to dynamically adjust the acquisition parameters of the phased array ultrasonic testing device according to the state deviation value in order to obtain target ultrasonic data. The image generation module is used to determine the constraint weights based on the state deviation value, input the target ultrasound data into the generative adversarial network model, and generate an ultrasound image with artifacts removed by combining the constraint weights. The verification and output module is used to perform multi-dimensional physical feature verification based on the real-time status parameters and the artifact-removed ultrasound image; if the verification passes, the artifact-removed ultrasound image is output, and a reset command is sent to the phased array ultrasound detection device to restore the parameters in the device's reference library.

[0007] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein when the computer program is executed by the processor, it implements the above-described method for removing ultrasonic artifacts from phased arrays of wind turbine bolts based on generative adversarial networks.

[0008] The wind turbine bolt phased array ultrasonic artifact removal method based on adversarial generative networks (PGN) of this invention establishes a collaborative closed-loop mechanism between the front-end testing equipment and the back-end GPN during actual non-destructive testing. This mechanism calculates the deviation between the physical state of the equipment and the baseline database in real time, and uses this deviation to dynamically adjust the acquisition parameters of the underlying hardware and the constraint weights of the upper-level software model. This two-way hardware-software feedback mechanism effectively resolves the conflict between static algorithms and dynamic operating conditions, enabling the model to adaptively perceive and compensate for physical data offsets caused by on-site operations. Meanwhile, by performing multi-dimensional physical feature verification in conjunction with real-time status parameters before output, the algorithm model can directly block image forgery behavior that may occur under complex working conditions, ensuring the physical authenticity of the image after removing artifacts, and effectively improving the reliability of on-site defect detection and the safety of equipment operation. Attached Figure Description

[0009] Figure 1 This is a schematic flowchart of the method for removing ultrasonic artifacts from phased array phased array wind turbine bolts based on adversarial generative networks provided by the present invention.

[0010] Figure 2 This is a comparison chart of the consistency verification of ultrasonic amplitude and depth attenuation in the ultrasonic artifact removal method for phased array wind turbine bolts based on adversarial generative networks provided by this invention.

[0011] Figure 3 This invention provides a pixel-level region partitioning thermal bitmap of the ultrasonic image noise interference level in the ultrasonic artifact removal method for phased array wind turbine bolts based on adversarial generative networks.

[0012] Figure 4 This is a dynamic response curve of probe wear and the piecewise increasing coupling thickness correction factor in the ultrasonic artifact removal method for phased array of wind turbine bolts based on adversarial generative networks provided by this invention.

[0013] Figure 5 This is a schematic diagram of the sliding time window monitoring and trend change rate triggering of the comprehensive domain difference factor in the ultrasonic artifact removal method for phased array wind turbine bolts based on adversarial generative networks provided by the present invention.

[0014] Figure 6 This is a comparison of time-domain waveforms showing the signal-to-noise ratio enhancement of the multi-pulse hardware superposition mode in the ultrasonic artifact removal method for phased array phased array wind turbine bolts based on adversarial generative networks provided by this invention.

[0015] Figure 7 This invention provides a method for removing ultrasonic artifacts from phased array ...

[0016] Figure 8 This is a schematic diagram illustrating the implementation of the wind turbine bolt phased array ultrasonic artifact removal system based on adversarial generative networks provided by the present invention.

[0017] Figure 9 This is a schematic diagram of the electronic device provided by the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0019] The following describes, with reference to the accompanying drawings, a method, system, and electronic device for removing ultrasonic artifacts from phased array ...

[0020] Example 1:

[0021] This embodiment provides a method for removing phased array ultrasonic artifacts from wind turbine bolts based on generative adversarial networks. This method is primarily applied in a control terminal connected to a phased array ultrasonic testing device. In this technical scenario, the control terminal is typically an industrial computer or edge computing server with high-performance computing capabilities, internally deployed with a deep learning inference framework and a low-level hardware communication protocol stack. The phased array ultrasonic testing device is the front-end hardware performing physical acoustic transmission and reception, integrating a multi-channel ultrasonic pulse generator, a receiver, and multi-dimensional state-sensing sensors. Specifically, the method in this embodiment includes the following steps: Step 1: System initialization and establishment of device benchmark library.

[0022] In the initial stage of performing the wind turbine bolt inspection task, the system needs to establish a stable physical and algorithmic baseline. The control terminal first responds to the start signal of the phased array ultrasonic testing equipment and executes the system-level initialization process.

[0023] First, a device reference library is established, including the initial impedance of the probe, the initial thickness of the coupling agent, and the initial angle of the probe. At the wind turbine high-altitude operation site, the phased array probe needs to be attached to the end face of the bolt for testing. Since the bolt end face may be rough, have uneven coating, or be slightly tilted, obtaining an initial stable attachment state is crucial. The initial impedance of the probe is defined as the initial stable value of the acoustic-to-electric conversion impedance calculated by the control terminal after the phased array probe elements and the end face of the bolt under test form an acoustic path through the coupling agent at the start of the testing task, by sending a broadband test pulse to the probe and measuring the electrical characteristics of the reflected echo. This value reflects the initial health state and electroacoustic matching efficiency of the probe. The initial thickness of the coupling agent is defined as the initial thickness of the ultrasonic coupling medium layer located between the bottom surface of the probe wedge and the bolt end face when the probe is pressed against the bolt end face to achieve a stable pressure state. This value is usually obtained by measuring the first round-trip flight time of the high-frequency monitoring pulse within the coupling agent layer and combining it with the sound velocity of the coupling agent. The initial angle of the probe is defined as the initial spatial angle between the normal direction of the probe crystal array and the axial direction of the wind turbine bolt. This value can be obtained by a high-precision microelectromechanical system attitude sensor integrated inside the phased array probe housing after the device is stably attached.

[0024] In addition, while establishing the device benchmark library, the control terminal initializes the adversarial generative network model in memory, that is, loads the pre-trained neural network weight parameters into the video memory and allocates computing resources for subsequent inference calculations.

[0025] Step 2: Real-time status parameter sensing and status deviation value calculation.

[0026] During the formal operation of the testing process, the micro-vibration of the tower caused by the operation of the wind turbine and the changes in the hand posture of the testing personnel will cause dynamic deviations in the front-end testing status. Furthermore, the control terminal acquires the real-time ultrasonic data and real-time status parameters collected by the phased array ultrasonic testing equipment, and calculates the state deviation value between the real-time status parameters and the corresponding parameters in the equipment's reference library.

[0027] Specifically, the real-time impedance of the probe, the real-time thickness of the coupling agent, and the real-time angle of the probe are acquired. During this process, the control terminal polls the sensor registers inside the phased array ultrasonic testing equipment in real time to acquire these dynamic physical quantities at a set high-frequency sampling rate. Subsequently, the first state deviation between the real-time impedance of the probe and the initial impedance of the probe, the second state deviation between the real-time thickness of the coupling agent and the initial thickness of the coupling agent, and the third state deviation between the real-time angle of the probe and the initial angle of the probe are calculated respectively.

[0028] To ensure the accuracy of the calculation, the first state deviation is defined as the ratio of the absolute value of the difference between the real-time impedance of the probe and the initial impedance of the probe to the initial impedance of the probe; the second state deviation is defined as the absolute value of the difference between the real-time thickness of the coupling agent and the initial thickness of the coupling agent; and the third state deviation is defined as the absolute value of the difference between the real-time angle of the probe and the initial angle of the probe.

[0029] For example, let's say This indicates the real-time impedance of the probe. This indicates the initial impedance of the probe. Let the first state deviation be represented, and its calculation algorithm is as follows: ; set up This indicates the real-time thickness of the coupling agent. Indicates the initial thickness of the coupling agent. To represent the second state deviation, its calculation algorithm is as follows: ; set up Indicates the real-time angle of the probe. Indicates the initial angle of the probe. To represent the deviation of the third state, its calculation algorithm is as follows: ; Step 3: Dynamically adjust the underlying acquisition parameters based on the state deviation value.

[0030] In traditional post-processing denoising methods, the front-end device is only responsible for acquiring data with fixed parameters. However, in this embodiment, the acquisition parameters of the phased array ultrasonic testing device need to be dynamically adjusted based on the state deviation value to obtain the target ultrasonic data. By improving signal quality at the physical source, the processing pressure on the back-end generative adversarial network can be effectively reduced.

[0031] Specifically, regarding the first state deviation, if the first state deviation is greater than a preset impedance threshold, the sampling rate and the number of sampling points of the phased array ultrasonic testing device are increased. An abnormal increase in impedance usually indicates minor wear or poor local contact in the probe array elements, leading to a decrease in the transmission efficiency of the high-frequency sound beam and severe attenuation of the high-frequency components of the echo signal. In this case, increasing the sampling rate and the number of sampling points follows the Nyquist sampling theorem, capturing denser waveform details in the time domain and using data redundancy to compensate for the decrease in energy of a single physical pulse.

[0032] Regarding the second state deviation, if it exceeds a preset thickness threshold, the ultrasonic transmission power is reduced, and a couplant replenishment alert signal is generated. An abnormal increase in couplant thickness usually indicates probe lifting or the presence of voids in the couplant. If high transmission power is maintained in this situation, the sound waves will undergo strong multiple reverberations within the thicker couplant layer, generating severe interface clutter. This means that the effective acoustic energy entering the tested bolt decreases, while the acoustic energy oscillating at the interface increases. Therefore, actively reducing the ultrasonic transmission power can effectively suppress this interface reverberation artifact, while the system-generated alert signal guides the operator to perform physical intervention.

[0033] Regarding the third state deviation, if the third state deviation is greater than a preset angle threshold, the device focal length is dynamically adjusted based on the third state deviation. The deflection of the probe angle causes the actual deflection angle and focusing depth of the phased array acoustic beam to deviate from the theoretical design value. Dynamically adjusting the device focal length refers to modifying the emission delay law (focusing law) at the underlying level of the phased array ultrasonic testing equipment to spatially compensate the focusing point of the acoustic beam at the electronic control level, thereby re-aligning it with the expected detection depth inside the bolt.

[0034] Through dynamic interaction and adjustment of the aforementioned underlying parameters, the data output by the phased array ultrasonic testing equipment is called target ultrasonic data, which has been optimized for anti-interference to the greatest extent possible at the physical level.

[0035] Step 4: Training set construction and pre-training of the adversarial generative network model.

[0036] It is important to note that before performing the aforementioned real-time inference calculations, the initialization of the adversarial generative network model includes pre-training the adversarial generative network model using the training set. The performance of data-driven models is highly dependent on high-quality training sets. Since defect data for wind turbine bolts is extremely scarce, and manually annotating complete artifact-free ground truth maps is extremely costly and prone to introducing subjective errors, this embodiment employs a weakly supervised pseudo-label generation and region segmentation strategy that integrates prior physical laws.

[0037] The training set acquisition process first includes: identifying structural artifacts conforming to periodic reflection laws based on the prior geometric features of the bolts, and generating structural artifact labels. Wind turbine bolts have standardized thread structures; when ultrasonic waves propagate to the root of the thread, they generate echoes with strict spatial periodicity and fixed reflection path. The control terminal extracts geometric prior features such as thread spacing by loading the bolt's CAD design model, automatically searches for and defines these periodic echo regions in the unlabeled ultrasonic spectrum, and marks them as structural artifact labels. Secondly, based on the theoretical attenuation law of ultrasound, consistency verification is performed on the multi-frequency dimension ultrasonic data, and regions whose amplitude attenuation curves deviate from the theoretical attenuation law are marked as attenuation anomaly pseudo-labels. In isotropic metallic materials, the amplitude of ultrasonic waves attenuates exponentially with increasing propagation depth.

[0038] For example, let's say This indicates that the ultrasonic wave propagates at a depth of [depth value missing]. The amplitude at that point, Indicates the initial transmission amplitude. Let represent the ultrasonic attenuation coefficient at a specific frequency. Then, the algorithmic formula for the theoretical ultrasonic attenuation law is: ; like Figure 2 The data shows the energy distribution pattern of ultrasonic signals propagating inside the wind turbine bolts. The horizontal axis represents the propagation depth in millimeters, and the vertical axis represents the ultrasonic echo amplitude in millivolts.

[0039] The blue dashed line in the figure represents the theoretical attenuation curve, which smoothly decreases from the initial 100 millivolts and attenuates to about 8 millivolts at a depth of 100 millimeters. This reflects the physical fact that ultrasound in a metallic medium attenuates exponentially with increasing propagation depth.

[0040] The solid red line in the figure represents the actual acquired signal envelope. Within the depth ranges of 0 to 40 mm and 60 to 100 mm, the trend of the actual acquired signal envelope generally conforms to the theoretical attenuation curve. However, within the depth range of 40 mm to 60 mm, the actual acquired signal envelope exhibits significant energy spikes and fluctuations, reaching a peak of approximately 60 millivolts at 50 mm, which clearly deviates from the normal physical attenuation pattern. This deviation range of 40 mm to 60 mm is marked as a gray attenuation anomaly area in the figure.

[0041] By combining the implementation steps of training set construction, the system can automatically lock bands such as 40 mm to 60 mm that do not conform to the physical attenuation law by comparing the deviation of amplitude attenuation curves at different depths with theoretical attenuation curves, and mark them as pseudo-labels of attenuation anomalies. In this way, the physical prior law is transformed into weak supervision guidance information during the pre-training of the adversarial generative network model, thereby improving the accuracy of the network in identifying and suppressing artifacts of complex internal structures of materials.

[0042] If, within a certain depth region, the amplitude of the ultrasonic signal does not follow the aforementioned attenuation rules—for example, if an abnormal local energy surge occurs or the attenuation rate is inconsistent across a wide frequency range, and this region is not located on a known structural reflective surface—then this region is highly likely caused by complex grain boundary scattering or irregular artifacts within the material, and the system automatically marks it as an attenuation anomaly pseudo-label.

[0043] Subsequently, the manually annotated real defect areas, the structural artifact labels, and the attenuation anomaly pseudo-labels are integrated, and the noise interference level of the pixel is calculated based on the texture similarity between the target pixel and the reference pixel.

[0044] In this process, it is necessary to quantify the severity of noise contamination at each pixel. For example, suppose... This represents the texture similarity value between the local window containing the target pixel and the local window of the reference pixel. This indicates the noise interference level of the target pixel. It is calculated using a structural similarity measurement algorithm. The value typically ranges from 0 to 1, with smaller values ​​indicating greater texture differences and more severe interference. The formula for calculating the degree of noise interference is defined as: ;

[0045] Based on the noise interference level calculated using the above formula, the ultrasound image is divided into core interference regions, blurred regions, and non-interference regions according to the noise interference level, thereby constructing the training set with weakly supervised prior information. This division method allows the network to focus its attention on the difficult-to-distinguish blurred regions and core interference regions during training, while the non-interference regions serve as a reference for maintaining background consistency.

[0046] like Figure 3 The figure objectively presents the quantitative distribution of external noise contamination at various spatial locations within the ultrasonic testing cross-section of wind turbine bolts. The horizontal axis represents the horizontal position in millimeters, and the vertical axis represents the depth position in millimeters. The color bar on the right indicates the noise interference level of each pixel, with values ​​ranging from 0 to 1 corresponding to a gradient from dark blue to dark red.

[0047] In the two-dimensional spatial distribution shown in the figure, the large areas of dark blue and light blue correspond to non-interference areas with a calculated noise interference level of less than 0.35, indicating that the ultrasonic reflection texture structure in these areas remains stable and can be used as a benchmark reference for maintaining the consistency of the physical features of the image background during model training.

[0048] The transitional zone between yellow-green and light yellow corresponds to a vague area with a noise interference level between 0.35 and 0.57, reflecting a moderate level of acoustic scattering or attenuation.

[0049] The dark red patches concentrated around 25 mm in the horizontal direction and 40 mm in the depth have a noise interference level significantly greater than 0.57, with a local peak value close to 1. This part corresponds to the core interference area, indicating that the physical morphology of this local area has been obscured by severe interface reverberation or anomalous structural artifacts.

[0050] By using this thermal bitmap partitioning method based on numerical calculation and color mapping, the system provides intuitive spatial region guidance for the construction of the training set of the Generative Adversarial Network. This enables the network to accurately focus its attention on blurred regions and core interference regions with an interference level greater than 0.35 during pre-training, thereby effectively enhancing the model's ability to identify and remove real defects and complex artifacts in harsh detection environments.

[0051] Step 5: Dynamic constraint weight calculation and domain adaptive adjustment.

[0052] After acquiring the target ultrasound data and having a pre-trained model, the next step is an adaptive inference process in the feature domain. Specifically, constraint weights are determined based on the state deviation values. Deep learning, if relying entirely on fixed static weights during inference, will be unable to adapt to the drift of the physical state on site.

[0053] In this embodiment, the loss function of the adversarial generative network model includes a dynamic constraint loss, which comprises an amplitude-preserving loss term and a shape-preserving loss term. The amplitude-preserving loss term constrains the image generated by the network to not deviate from the true physical sound intensity in terms of pixel intensity, while the shape-preserving loss term constrains the image generated by the network to not undergo distortion in the defective topology.

[0054] For example, let's say Represents dynamic constraint loss. Indicates the loss item. This represents the conformal loss term. Indicates the amplitude preservation weight. Let represent the conformal weights. Then the overall loss calculation algorithm is as follows: ; The step of determining the constraint weights based on the state deviation values ​​includes: dynamically calculating the amplitude preservation weights of the amplitude preservation loss term and the shape preservation weights of the shape preservation loss term based on the noise interference level of the pixel, the comprehensive domain difference factor obtained by weighted fusion of the state deviation values, the probe wear correction factor associated with the first state deviation, and the coupling thickness correction factor associated with the second state deviation.

[0055] The amplitude preservation weight is positively correlated with the noise interference level, while the shape preservation weight is negatively correlated with the noise interference level. The physical meaning is that when the noise interference level in a certain region is very high, the image shape (morphology) is severely contaminated, and forcibly preserving the shape will lead to the noise being treated as valid structure. Therefore, in this case, the shape preservation weight is reduced, and the amplitude preservation weight is increased instead, constraining the network only at the energy amplitude level, allowing the network to perform large-scale reconstruction and denoising at the shape level.

[0056] In addition, in order to cope with the deterioration of the probe's physical condition, the value of the probe wear correction factor increases in segments as the first state deviation increases, and the value of the coupling thickness correction factor increases in segments as the second state deviation increases.

[0057] set up This indicates the probe wear correction factor. This represents the first state deviation. The algorithm logic for piecewise incrementing is defined as follows: 1. When When less than or equal to the first limit constant, Output the initial constant value; 2. When When it is greater than the first limit constant and less than or equal to the second limit constant. Output the value of the first penalty constant; 3. When When it is greater than the second limit constant, Output the value of the second penalty constant, which is greater than the value of the first penalty constant.

[0058] set up This represents the coupling thickness correction factor. This represents the second state deviation, which also follows the corresponding piecewise incremental algorithm logic. These correction factors are added as multiplicative coefficients to the base weight values, so that when the physical state of the device deteriorates significantly, the network will more aggressively activate the loss penalty mechanism.

[0059] like Figure 4 It contains two sub-graphs, one above the other, which objectively demonstrate the segmented feedback mechanism of the system algorithm model in response to the degradation of the underlying physical hardware state.

[0060] The horizontal axis of the upper subgraph represents the probe impedance deviation in percentage, and the vertical axis represents the probe wear correction factor. The blue solid line in the figure represents the probe wear correction curve.

[0061] When the probe impedance deviation is in the range of 0 to 5, the correction factor remains at 0, indicating that the probe is in good condition and no additional penalty is required. When the deviation is greater than 5 and less than or equal to 10, the curve experiences its first step, and the correction factor increases to 0.1, indicating that the system initiates a mild wear penalty. When the deviation is greater than 10, the curve experiences another step, and the correction factor increases to 0.2, indicating that the system initiates a severe wear penalty.

[0062] The horizontal axis of the subplot below represents the coupling agent thickness deviation in millimeters, and the vertical axis represents the coupling thickness correction factor. The solid red line in the figure represents the coupling thickness correction curve.

[0063] When the coupling agent thickness deviation is in the range of 0 to 0.2 mm, the correction factor is 0; when the deviation is greater than 0.2 mm and less than or equal to 0.5 mm, the curve jumps to 0.15; when the deviation is greater than 0.5 mm, the correction factor jumps further to 0.3.

[0064] This stepped, segmented, and progressively increasing curve trend intuitively reflects that as the deviation of the physical state increases across intervals, the system will gradually increase the constraint weights in the loss function according to the set numerical gradient. This guides the deep learning model to adaptively adjust the image reconstruction strategy and effectively suppress false artifacts caused by probe anomalies when the physical detection environment gradually deteriorates.

[0065] To further eliminate data distribution shifts caused by changes in operating conditions, the step of determining constraint weights based on the state deviation value also includes an adaptive adjustment step based on a comprehensive domain difference factor: constructing a three-dimensional domain label including equipment dimension, operating condition dimension, and focusing law dimension, and constructing a domain discriminator in the adversarial generative network model for learning common features between domains. The first state deviation, the second state deviation, and the third state deviation are normalized respectively, and the equipment domain difference value is obtained by weighted calculation.

[0066] For example, let's say This represents the first-state deviation after normalization. This represents the normalized second-state deviation. This represents the deviation of the third state after normalization. , , These represent the corresponding preset weighting coefficients. The formula for calculating the device domain difference value is as follows: ; Subsequently, the preset initial three-dimensional domain difference values ​​are weighted and fused with the device domain difference values ​​to obtain the comprehensive domain difference factor. Let... This represents the preset initial three-dimensional domain difference value. and Indicates the fusion weight. If we represent the difference factor in the comprehensive domain, then the calculation formula is: ; Based on the numerical range of the comprehensive domain difference factor, the adversarial loss weight and the domain adaptive loss weight are dynamically allocated. Specifically, when the comprehensive domain difference factor increases and crosses a preset adjustment range, the adversarial loss weight is decreased while the domain adaptive loss weight is simultaneously increased. The technical significance of this domain adaptive mechanism lies in the fact that when a large inter-domain gap exists between newly collected data and laboratory training data due to deteriorating field conditions, maintaining a high proportion of adversarial loss will cause the model to experience severe inference illusions due to the unfamiliar distribution. In this case, forcibly increasing the weight of the domain adaptive loss forces the network to extract device-independent, universal essential features, improving the model's robustness and generalization ability under extremely harsh conditions.

[0067] Step Six: Target ultrasound data feature encoding and adversarial generative network image generation.

[0068] After determining all constraint weights and adaptive parameters, the step of inputting the target ultrasound data into the adversarial generative network model and generating an ultrasound image with artifacts removed by combining the constraint weights is performed.

[0069] Optionally, the adversarial generative network model includes a generator and a discriminator.

[0070] The process of inputting the target ultrasound data into the generative adversarial network model specifically includes: standardizing and encoding the operating parameters, frequency information, focusing law parameters, and bolt geometric parameters to generate a condition vector. Since ultrasound imaging is highly dependent on the current physical settings, simply inputting the image pixel matrix is ​​insufficient. Standardizing these parameters into a one-dimensional tensor, i.e., a condition vector, provides the network with crucial physical priors.

[0071] Subsequently, the conditional vector is mapped through a fully connected layer and then concatenated with the feature map of the target ultrasound data before being input into the generator. This multimodal concatenation technique allows the generator to not only see the spatial morphology of the acoustic image but also understand the specific hardware state and operating conditions that generated the image, thereby enabling more accurate directional mapping and artifact removal.

[0072] On the other hand, the discriminator adopts a dual-branch structure, where the first discriminant branch is used to determine the authenticity features of the input image, and the second discriminant branch is used to determine the physical temporal consistency features and cross-frequency consistency features of the input ultrasonic signal. Traditional discriminators only score the visual realism of the image, while the second discriminant branch introduced in this embodiment forces the generated image to conform to the acoustic time-of-flight law of ultrasonic waves propagating inside the steel of the wind turbine bolt and the attenuation law under different frequency components when it is inverted back to a one-dimensional ultrasonic A-scan signal. This effectively overcomes the technical defect of deep learning in image generation that generates false textures in pursuit of visual smoothness.

[0073] Step 7: Multi-dimensional physical feature verification.

[0074] After the generator outputs the artifact-free ultrasound image, a multi-dimensional physical feature verification is performed based on the real-time state parameters and the artifact-free ultrasound image. This step is the final safety valve before the system outputs the result, ensuring the accuracy of the final result at the physical level.

[0075] The multi-dimensional physical feature verification includes the following parallel verification steps:

[0076] The first dimension verification involves extracting the depth amplitude data corresponding to the artifact-removed ultrasound image, calculating the attenuation curve, and verifying the deviation of the attenuation curve from the theoretical ultrasound attenuation curve. Specifically, this is done by integrating and projecting the grayscale values ​​of the two-dimensional image along the depth direction to recover the amplitude envelope, and verifying whether it violates the physical principle that sound waves in metallic media exhibit exponential attenuation.

[0077] The second dimension verification involves reversing the focus parameters from the artifact-removed ultrasound image and verifying the matching degree between these focus parameters and the real-time focusing law parameters of the phased array ultrasound detection device. If the generated image alters the phase relationship of the original data, the deduced focus position will drift, indicating that the algorithm has excessively modified the original data, and the verification will fail.

[0078] The third dimension verification involves comparing the attenuation rate trends of high-frequency and low-frequency signals in the target ultrasound data. According to Rayleigh scattering theory, high-frequency sound waves attenuate much more than low-frequency sound waves in a medium. By separating frequency bands and verifying the attenuation trend, it can be further confirmed whether the structures preserved in the image are genuine acoustic echoes.

[0079] The fourth dimension verification involves generating a differential correlation mapping between device state change parameters and image data feature changes to verify the correspondence between artifact distribution patterns and abnormal state parameters. For example, when a probe is partially suspended, leading to deteriorated coupling, signal loss will inevitably occur in the corresponding image area. Verifying this correspondence prevents the algorithm from misjudging blind spots caused by poor coupling as the dense structure of the material itself.

[0080] The system will determine that the multi-dimensional physical feature verification has passed only if the first dimension verification, the second dimension verification, the third dimension verification, and the fourth dimension verification all meet their respective set qualification thresholds.

[0081] Step 8: Confidence level classification, closed-loop control and system reset.

[0082] Finally, based on the results of the multi-dimensional physical verification, the system needs to execute feedback actions. The method also includes confidence leveling and closed-loop control steps based on the verification results.

[0083] Specifically, the overall confidence level is calculated based on the scores of the parallel verification steps.

[0084] For example, let's say , , , These represent the quantitative scores for the first, second, third, and fourth dimensions of verification, respectively. , , , These represent the corresponding confidence level weighting ratios. To represent the overall confidence level, its calculation formula is: ; Extract the preset high-confidence threshold and the medium-confidence threshold below the high-confidence threshold. If the multi-dimensional physical feature verification passes, i.e. If the value is greater than or equal to the high confidence threshold, the artifact-free ultrasound image is output, and a reset command is sent to the phased array ultrasound detection device to restore the parameters in the device's reference library. That is, the parameters such as sampling rate, transmission power, and focal length that have been temporarily changed due to state deviation are restored to the initial safe settings to prepare for the next detection.

[0085] If the overall confidence level is less than the high confidence threshold but greater than or equal to the medium confidence threshold, it indicates that the system believes the currently generated image has some physical and logical flaws, but is not completely invalid. At this time, the control terminal sends a parameter adjustment command to the phased array ultrasonic testing device, instructing the underlying hardware to re-optimize the gain or fine-tune the electronic deflection angle, and triggers a secondary verification process after reacquiring ultrasonic data, forming an automatic error correction closed loop.

[0086] If the overall confidence level is less than the median confidence threshold, it indicates that the noise interference or physical degradation in the current environment has exceeded the limits of hardware and software collaborative compensation. Continuing to output deep learning images will bring a significant risk of missed detections or misjudgments. At this time, the system automatically triggers a comprehensive device calibration alarm signal, prompting the operator to reapply the coupling agent or replace the probe, and reverts to the original ultrasound data mode that includes risk area markings. Experienced human experts then take over the evaluation to ensure the safety of non-destructive testing.

[0087] In general, traditional deep learning image denoising solutions often suffer from the black box effect. Their actual application performance is highly dependent on the consistency between the field data and the training library. When encountering complex and dynamic operating conditions at wind turbine sites, they are prone to network crashes or image spoofing.

[0088] The overall technical solution described in this embodiment organically connects the physical state perception of the front-end ultrasonic probe, the dynamic adjustment of the underlying signal acquisition parameters, and the adaptive adjustment of the constraint weights within the adversarial generative network model, constructing a two-way interactive closed loop from the hardware source to the algorithm terminal. This enables the model not only to passively process images but also to actively perceive the physical drift of the detection environment. This eliminates severe artifact interference while maximizing the preservation of the physical authenticity of minute defect features, significantly improving the intelligence level and reliability of non-destructive testing of wind turbine bolts under complex and harsh working conditions.

[0089] Example 2:

[0090] In Example 1, the system adapts to normal physical state deviations by dynamically adjusting the constraint weights of the algorithm. However, in certain extreme situations during actual operation, the wind turbine tower may experience violent high-frequency swaying due to strong gusts, or the operator's probe pressing posture may experience momentary violent shaking due to physical exhaustion, or even severe deep pitting corrosion on the bolt end face. In these extreme scenarios, the aforementioned comprehensive domain difference factor may undergo drastic numerical jumps or exceed the physical tolerance limit in a very short time. If the system still relies on the unrestricted dynamic weight allocation logic in Example 1, it will cause the adversarial generative network model to fall into gradient explosion or mode collapse, making the model not only unable to remove artifacts but also erasing real, subtle defect features.

[0091] To resolve the conflict between this physical limit and the algorithmic logic, this embodiment, after dynamically allocating the adversarial loss weight and the domain adaptive loss weight, also includes a hardware-software collaborative intervention step for extreme degradation detection environments. This collaborative intervention step is not simply an adjustment of algorithm parameters, but rather establishes a complete closed loop from circuit breaker protection at the algorithm level to physical signal compensation at the hardware level.

[0092] For example, for the identification and monitoring of extreme degradation detection environments, the system first constructs a sliding time window, calculates the rate of change of the comprehensive domain difference factor within the sliding time window in real time, compares the comprehensive domain difference factor with a preset extreme degradation threshold, and compares the rate of change with a preset oscillation threshold.

[0093] Specifically, the extreme degradation detection environment in the above steps is defined as the boundary state where the physical coupling state, probe orientation, or hardware impedance of the phased array ultrasonic testing equipment deviates from the equipment's baseline parameters to a degree exceeding the boundary state that the algorithm can effectively compensate for through conventional software mapping. To accurately capture this state, the control terminal allocates a first-in-first-out queue storage area in memory to construct a sliding time window. This sliding time window has a fixed time length and capacity, used to cache the continuously input comprehensive domain difference factor values ​​in a time sequence. As new comprehensive domain difference factor values ​​are continuously input, the sliding time window automatically discards the oldest historical values, thus maintaining dynamic monitoring of the latest physical state.

[0094] After acquiring the data sequence within the sliding time window, the system performs a step of real-time calculation of the rate of change of the comprehensive domain difference factor within the sliding time window. To avoid misleading the calculation of the rate of change due to single-point impulse noise, this embodiment uses a linear regression algorithm based on the least squares method to calculate the rate of change, that is, to characterize the rate of state deterioration by fitting the overall trend slope of the data within the window.

[0095] For example, let's say This indicates the total number of data sampling points contained within the sliding time window. Indicates the first The timestamps corresponding to each data sampling point Indicates that at timestamp The system calculates and stores the comprehensive domain difference factor values ​​within the sliding time window. This represents the average of all timestamps within the sliding time window. This represents the average value of all composite domain difference factors within the sliding time window. This represents the rate of change of the calculated comprehensive domain difference factor within the sliding time window. The formula for calculating the linear regression slope of this rate of change is: ; The significance of this formula lies in the fact that its calculation result represents the instantaneous trend gradient of the comprehensive domain difference factor over time. If... The absolute value is very large, indicating that the physical state of the detection environment is undergoing drastic changes or high-frequency oscillations.

[0096] like Figure 5This demonstrates the system's decision-making process for triggering the parameter circuit breaker protection mechanism when facing extremely degraded detection environments. The horizontal axis in the figure represents time in seconds, the left vertical axis represents the comprehensive domain difference factor (corresponding to the blue broken line and scatter points in the figure), and the right vertical axis represents the trend change rate (corresponding to the orange bars in the figure).

[0097] The graph shows two horizontal dashed lines, representing the extreme degradation threshold of 0.8 on the left ordinate and the oscillation threshold of 0.15 on the right ordinate. During the typical detection period of 0 to 6 seconds, the comprehensive domain difference factor steadily increases between 0.2 and 0.65, and the corresponding trend change rate histogram remains below 0.14, indicating that the front-end physical detection status is within a healthy range that the algorithm can adaptively adjust.

[0098] When the time progressed to 6.5 seconds, the value of the comprehensive domain difference factor instantly jumped to 0.85, breaking through the set extreme degradation threshold. At the same time, the trend change rate histogram within this time window also increased sharply to 0.4, far exceeding the oscillation threshold of 0.15.

[0099] This dual over-limit characteristic is manifested in the graph as a simultaneous abrupt change in the line graph and the bar chart. Based on this, the system determines that the front-end phased array ultrasonic testing equipment has encountered extreme and harsh conditions such as severe probe detachment or strong mechanical vibration. At this time, the system immediately triggers a parameter circuit breaker command, stops the logic of dynamically allocating loss weights and freezes the parameters of the feature extraction layer, and simultaneously sends a multi-pulse hardware compensation command to the underlying hardware. This effectively prevents abnormal noise data from contaminating the deep learning model and ensures the reliability of the system's detection under harsh conditions.

[0100] It is also important to note that after calculating the rate of change, the system executes a dual-dimensional conditional judgment logic. The control terminal compares the latest calculated value of the comprehensive domain difference factor with a preset extreme degradation threshold at the absolute value level; simultaneously, the control terminal compares the calculated rate of change with a preset oscillation threshold at the trend level.

[0101] The extreme degradation threshold is defined as the lower limit of the comprehensive domain difference factor recorded during the offline calibration phase, simulating failure conditions such as the probe completely detaching from the coupling agent or being severely tilted to the point of total ultrasonic reflection. The oscillation threshold is defined as the maximum reasonable fluctuation slope of the comprehensive domain difference factor allowed by the system per unit time.

[0102] Based on the above dual comparisons, the system establishes the following triggering logic: if the difference factor of the synthesis domain is greater than the extreme degradation threshold, or the rate of change is greater than the oscillation threshold, then the parameter circuit breaker protection mechanism is triggered. This OR logic gate design aims to achieve comprehensive security. As long as the absolute value of the difference in the synthesis domain exceeds the physical failure limit (indicating that the currently acquired ultrasonic data has become pure noise), or its rate of change exceeds the high-frequency oscillation limit (indicating severe mechanical vibration interference in the environment, and the algorithm can no longer stably optimize), the system will immediately suspend its normal operation.

[0103] Optionally, after the system confirms that it has encountered an extremely severe working condition and triggered the corresponding signal, the parameter circuit breaker protection mechanism includes: stopping the execution of the allocation instruction to reduce the adversarial loss weight and increase the domain adaptive loss weight, forcibly locking the adversarial loss weight and the domain adaptive loss weight to the weight ratio when the parameter circuit breaker protection mechanism is triggered, and freezing the parameter update action of the feature extraction layer in the adversarial generative network model.

[0104] Specifically, the parameter circuit breaker mechanism is a defensive isolation measure proactively taken to prevent catastrophic forgetting and gradient contamination in deep learning models when they face extreme out-of-distribution data shocks. In the conventional logic of Example 1, as the domain difference factor increases, the system continuously decreases the adversarial loss weights and increases the domain adaptive loss weights. However, in extremely degraded environments, the input data no longer possesses domain adaptive value, but is instead filled with random noise or physical blocking signals. If the domain adaptive loss weights are further increased at this time, the adversarial generative network model will be forced to fit these extreme erroneous noise features, resulting in severe contamination of the model weight matrix.

[0105] Therefore, the parameter-based circuit breaker protection mechanism first intercepts the dynamic allocation logic at the instruction level. The system forcibly locks the adversarial loss weight and the domain adaptive loss weight to the weight ratio required to trigger the parameter-based circuit breaker protection mechanism.

[0106] For example, let's say This represents the adversarial loss weight variable at the current moment. Let the domain adaptive loss weight variable at the current time be... This represents the precise timestamp that triggers the circuit breaker protection mechanism. Indicates in The adversarial loss weight values ​​stored within the system at any given time. Indicates in The system stores the domain adaptive loss weight values ​​at any given time. Then, at any time after the circuit breaker is triggered... (in Greater than The mathematical execution logic for weight locking is as follows: ; ; Through the above locking operation, the composition of the model's loss function is fixed at the relatively healthy state at the last moment before the environment deteriorates to an extreme degree.

[0107] More importantly, in addition to locking the external weights of the loss function, the system also initiates a freezing intervention within the model. The system freezes the parameter update actions of the feature extraction layer in the generative adversarial network (GAN) model. The feature extraction layer is typically defined as a series of convolutional network layers at the front end of the GAN generator. Its main responsibility is to extract basic physical features such as edges, phase, and amplitude from the underlying ultrasound A-scan or C-scan images. Because these basic physical features are extremely sensitive to the signal-to-noise ratio of the input data, if severely contaminated and extremely degraded data is input, the gradient calculation results of the feature extraction layer will exhibit drastic randomness.

[0108] For example, let's say Let represent the parameter vector containing all convolutional kernel weights and bias terms in the feature extraction layer. Further, let This represents the feature extraction layer parameter vector before parameter updates (or in the previous iteration). This represents the feature extraction layer parameter vector after parameter updates (or in the current iteration step); This represents the gradient vector of the loss function with respect to the feature extraction layer, calculated using the backpropagation algorithm. This represents the learning rate of the network optimizer.

[0109] In conventional incremental training and updates of a network, the algorithm for updating the parameter vector is as follows: parameter vector before update Subtract the product of the learning rate and the gradient. When performing the freeze operation, the system truncates the gradient calculation channel of the feature extraction layer within the underlying computational graph framework, thus... Forced to be assigned a zero vector, the mathematical logic of parameter updates in the frozen state then manifests as follows: ;

[0110] The physical meaning of this freeze action is that the system determines that the input data in the current environment is unreliable, and therefore strictly prohibits the model from using this unreliable data to modify its valuable feature extraction capabilities learned from a large amount of high-quality benchmark data. The high-level logical reasoning layer of the model still maintains a certain degree of flexible computation, but the low-level feature extractor is subjected to absolute physical isolation, effectively preventing the model from losing its basic ability to identify real defect features due to data deterioration.

[0111] It is important to note that simply implementing circuit breaker isolation at the software and algorithm level can only prevent damage to the model, but it cannot solve the problem of not being able to obtain effective detection data. Therefore, the system must initiate proactive self-rescue measures at the physical front end. Simultaneously with triggering the parameter circuit breaker protection mechanism, the system sends a multi-pulse hardware compensation command to the phased array ultrasonic testing equipment to force the phased array ultrasonic testing equipment to switch to multi-pulse hardware superposition mode.

[0112] Specifically, the multi-pulse hardware compensation instruction is a special high-priority hardware control command generated by the control terminal and sent to the main control chip of the phased array device via the underlying communication protocol stack. The function of this instruction is to force the phased array ultrasonic testing equipment to abandon the single-pulse transmission and reception logic used in conventional scanning.

[0113] The multi-pulse hardware superposition mode is defined as follows: for the same spatial focusing detection point or the same detection sound beam scanning line inside the bolt, the phased array probe continuously triggers multiple repeated physical pulse transmissions within a very short time interval without changing any electronic deflection and focusing parameters, and performs direct time-domain accumulation and averaging of these multiple received echo signals in the hardware register or field-programmable gate array buffer at the front end of the analog-to-digital converter.

[0114] The physical acoustic principle of this mode lies in leveraging the fundamental difference between signal coherence and noise incoherence to forcibly improve the signal-to-noise ratio under harsh operating conditions. In extremely degraded detection environments, such as strong attenuation due to poor coupling or clutter interference caused by mechanical vibration, each received single ultrasonic pulse echo signal contains two parts:

[0115] One part consists of real reflected signals from the internal structure of the wind turbine bolts. In multiple consecutive transmissions, the phase and arrival time of this part of the signal are strictly locked, i.e., it is coherent. The other part consists of random scattered noise or system thermal noise generated by poorly contacted interfaces. In each transmission, the phase and amplitude of this part of the noise are randomly distributed, i.e., it is incoherent.

[0116] For example, let's say This indicates the total number of pulses that will be continuously and repeatedly transmitted as set in the multi-pulse hardware superposition mode. The index variable represents the current transmitted pulse, and The value range is 1 to Let be a positive integer. This indicates that the phased array ultrasonic testing equipment is in the first... The original time-domain electrical signal received after the next transmission consists of real physical reflection components and random noise components. Let... This indicates the actual physical reflection signal components inside the wind turbine bolts. Indicates the first The signal is mixed with random noise of zero mean during each reception. Therefore, a single received signal can be expressed as: ; Since the target location remains unchanged for each launch, In all The received waveform is a constant and unchanging waveform. This represents the compensated ultrasound data sequence output after averaging and summing at the hardware level in multi-pulse hardware superposition mode. The core algorithm formula for this hardware compensation calculation is: ; Substitute the components of a single received signal into the above formula and expand: ; ; According to statistical principles, for a sequence of random noise variables with a mean of zero and independent distributions, the variance of its cumulative mean will increase with the number of accumulations. The energy of the random noise component decreases significantly with the increase in the number of pulses. Specifically, after multi-pulse hardware superposition processing, the energy of the random noise component is attenuated to its original value. This means that the signal-to-noise ratio of the underlying physical signal is forcibly increased. times.

[0117] like Figure 6 The diagram, consisting of two sub-graphs, visually illustrates the changes in the underlying acoustic signals of a phased array ultrasonic testing device before and after activating the multi-pulse hardware overlay mode under extremely degraded conditions.

[0118] The upper sub-figure illustrates the waveform state under conventional single-pulse transmit and receive logic. The horizontal axis represents time in microseconds, and the vertical axis represents signal amplitude in millivolts. The red waveform in the figure represents the raw electrical signal received in a single transmission. Within the time interval of 0 to 10 microseconds, the waveform exhibits violent oscillations with extremely large amplitudes, and the peak random noise exceeds 100 millivolts. This causes the actual physical reflection signal component around 5 microseconds to be overwhelmed by interface scattering noise and system thermal noise.

[0119] The sub-figure below shows the waveform state after executing the multi-pulse hardware compensation command, with the horizontal and vertical axes defined in the same way as the sub-figure above. The blue waveform in the figure represents the multi-pulse hardware superposition compensation data. By directly accumulating and averaging the received echoes transmitted multiple times in a very short time interval in the time domain at the underlying hardware level, the incoherent random noise components are significantly attenuated, and the noise floor is effectively suppressed to within 20 millivolts.

[0120] Meanwhile, due to the strict coherence of the actual physical reflection signal inside the wind turbine bolts during continuous reception, its high-frequency packet characteristics around 5 microseconds are clearly highlighted. This waveform comparison objectively reflects the practical physical significance of the hardware compensation step in the hardware-software collaborative intervention mechanism. Specifically, it utilizes the principle of acoustic coherent accumulation to improve the signal-to-noise ratio (SNR) from the hardware source, providing input data with a high initial physical SNR for the subsequent adversarial generative network model, which is in a weight-locked and parameter-frozen state. This ensures the reliability of the model's deep feature mapping and image reconstruction under harsh operating conditions.

[0121] It is also important to note that when applying this formula, the system must control the time interval between pulse transmissions. This time interval is set to be strictly greater than the maximum physical round-trip time of the ultrasonic wave propagating in the wind turbine bolts. This setting is to avoid the late multiple reflections of the previous pulse interfering with the reception and acquisition of the next new pulse, thereby ensuring that the random noise received in each pulse according to the above formula is minimized. To achieve optimal noise suppression, they are statistically independent of each other.

[0122] Specifically, through the aforementioned physical means that consume extremely high hardware resources and time, the system extracts effective compensation signals from extremely poor working conditions. Finally, the control terminal acquires the compensated ultrasound data in the multi-pulse hardware superposition mode and inputs the compensated ultrasound data into the weighted generative adversarial network model for artifact removal processing.

[0123] At this point, the data input into the generative adversarial network model is the data calculated using the formula described above. In this processing stage, despite the extremely harsh external physical environment, on the one hand, the data fed into the model has already been purified and enhanced at the hardware level using multiple pulse coherent superposition principles, resulting in a high initial physical signal-to-noise ratio; on the other hand, the generative adversarial network model is in a protected state with weight locking and feature extraction layer parameter freezing, preventing the model from making aggressive and destructive parameter adaptive modifications due to extreme anomalies in the current environmental state. Therefore, the model with locked weights can safely and stably utilize its artifact removal and inference capabilities learned in the previous healthy state to perform deep feature mapping and image reconstruction on the hardware-compensated signal.

[0124] In light of current technology, when faced with extreme and harsh conditions such as sudden high-frequency vibrations or severe coupling failures during on-site inspection of wind turbine bolts, traditional machine learning algorithms often attempt to cope unilaterally by widening threshold boundaries or enhancing software filtering parameters. However, purely software-based generative adversarial networks are highly susceptible to mode collapse and gradient tearing when faced with severely abrupt signals that have lost their physical meaning. This results in outputting a falsely normal image that appears smooth and flat but actually lacks all the minute, real defects, posing a serious threat to industrial safety.

[0125] The technical solution in this embodiment accurately captures the moment of crossing physical limits by real-time monitoring of the mutation rate of differential factors and setting dual trigger thresholds. Subsequently, the solution solves the problem of separation between software and hardware data processing and acquisition. When the software faces the risk of crashing, the system decisively executes a parameter circuit breaker mechanism to protect the knowledge memory of the generative adversarial network by locking weights and freezing feature extraction layer updates.

[0126] Simultaneously, the system reverses and takes over the physical pulse transmission logic of the underlying phased array probe, forcibly activating a multi-pulse hardware superposition mode. Utilizing the principle of acoustic coherent accumulation, it compensates for signal-to-noise ratio loss at the hardware source. This strong hardware-software collaborative intervention mechanism—software failure followed by hardware activation and compensation, then handing over to the protected software—significantly enhances the survivability of deep learning systems in the field of non-destructive testing when facing unforeseen interference. The overall solution avoids the defect erasure problem caused by relying solely on algorithms to forcibly fit noise, ensuring that the inspection data of wind turbine bolts under the harshest service environment not only achieves artifact removal in visual images but also possesses extremely high security and authenticity in terms of underlying physical logic and feature preservation.

[0127] Example 3:

[0128] During high-altitude operation and long-term service of wind turbines, irregular fatigue microcracks are prone to initiation in stress concentration areas such as the root of bolt threads. These microcracks are often accompanied by intergranular corrosion, exhibiting physical characteristics such as star-shaped branching, extremely rough surfaces, and polyhedral spatial distribution. When ultrasonic waves are incident on these areas, strong waveform conversion, phase cancellation, and multiple diffuse reflections occur. In the adversarial generative network models described in Examples 1 and 2, this complex physical scattering phenomenon leads to severe discriminative divergence in the dual-branch structure within the discriminator, thereby causing conflicts in gradient calculation directions.

[0129] To address the potential image distortion and defect detection risks that may arise when deep learning models handle complex physical scattering, this embodiment provides a gradient conflict suppression and hardware matrix capture step, which includes the following:

[0130] The system first performs the step of extracting the true / false discrimination loss value output by the first discrimination branch and the physical consistency discrimination loss value output by the second discrimination branch in real time, and calculating the relative difference between the true / false discrimination loss value and the physical consistency discrimination loss value.

[0131] Specifically, in the discriminator's dual-branch architecture of the generative adversarial network model, the first discriminator branch is primarily responsible for evaluation from the two-dimensional image texture dimension of computer vision, while the second discriminator branch is primarily responsible for evaluation from the one-dimensional acoustic-physical dimension of the ultrasonic time-series signal. When faced with the aforementioned star-shaped fatigue microcracks, due to strong phase cancellation and diffuse reflection, conventionally synthesized ultrasonic images in this region will exhibit scattered, fragmented, low-contrast, and spot-like clusters lacking the clear outlines of typical defects. After extracting such scattered visual features, the first discriminator branch tends to classify them as unrealistic image noise and outputs a very large penalty loss value.

[0132] However, despite the visual clutter of the image, the arrival time and attenuation rate of the high-frequency components of these echoes in the underlying time-domain signal sequence still strictly follow the dynamic and kinematic laws of ultrasonic wave propagation inside the steel bolt. After extracting these physical temporal features, the second discriminant branch tends to classify them as real physical interface reflection signals, thus outputting a very small reward loss value.

[0133] To quantify this internal discrepancy caused by complex scattering, the system needs to extract these two loss values ​​in real time and calculate their relative difference. Let... Let the true / false loss value be the output of the first discrimination branch for the current input data packet. Let represent the physical consistency discrimination loss value output by the second discrimination branch for the same input data packet. This represents the calculated relative difference value, which characterizes the degree of gradient tearing between the two branches during backpropagation. This represents a correction factor for the smallest normal quantity introduced to prevent division by zero. The algorithm for calculating the relative difference value is then defined as follows: ; The result obtained by the above formula The value ranges between 0 and 1. The closer the value is to 1, the more contradictory the conclusions given by the two decision branches are, and the more severe the gradient conflict within the network.

[0134] For example, after completing the real-time calculation of the relative difference value, the system performs a step of determining whether the relative difference value is greater than a preset conflict threshold; if the relative difference value is greater than the conflict threshold, it is confirmed that the system is currently in a complex scattering state, and a weight dynamic takeover action is performed.

[0135] Specifically, the preset conflict threshold in the above steps is a critical value statistically derived after extensive scanning tests using artificial defect comparison blocks with standard irregular surfaces during the system's offline calibration phase. Let... This indicates the preset conflict threshold, which the logic processing unit of the control terminal will calculate in real time. and Perform numerical comparisons. When Less than or equal to When the gradient divergence is considered to be a minor deviation caused by normal training fluctuations or normal image background noise, the adversarial generative network model continues to maintain the established joint loss weight allocation ratio and performs normalized parameter updates.

[0136] It is also important to note that when Strictly greater than At this point, the system qualitatively confirms at the logical level that the current ultrasonic detection area is in a complex scattering state. In this state, if the original loss weight allocation is maintained, due to... The numerical value is enormous, and the gradient vector it generates during backpropagation will dominate. This will force the generator of the adversarial generative network model to forcibly remove those broken spots that represent real microcracks as artifacts in order to cater to the visual smoothness aesthetic of the first discriminant branch, ultimately leading to extremely dangerous false negatives.

[0137] Therefore, in response to the confirmation of this complex scattering state, the system immediately triggers a dynamic weight takeover action. The core technical means of this action is to reduce the loss weight corresponding to the first discrimination branch and simultaneously increase the loss weight corresponding to the second discrimination branch, so as to force the generator to retain the signal that conforms to the ultrasonic physical time sequence characteristics.

[0138] Specifically, set This represents the initial value of the image discrimination loss weight corresponding to the first discrimination branch stored in the system before the dynamic weight takeover action is triggered. Let it be... This represents the initial value of the physical discrimination loss weight corresponding to the second discrimination branch stored in the system before the action is triggered. Let it be... This represents the new loss weight of the first discriminant branch after dynamic takeover adjustment, let's assume... This represents the new loss weight of the second discriminant branch after dynamic takeover adjustment, let's assume... This represents the preset visual weight attenuation coefficient, let's assume... This represents the preset physical weight amplification factor. The mathematical calculation logic for dynamic weight takeover is then defined as follows: ; ; As can be seen from the above formula, the larger the relative difference value, the more severely the weight of the first discriminant branch is compressed exponentially, while the weight of the second discriminant branch receives a corresponding compensatory boost. Through this forced tilting of the algorithm weights, the system constructs an extremely deep physically consistent local optimum on the loss function surface, forcibly guiding the generator's gradient descent direction away from visual smoothness and towards physical reality. This mechanism effectively suppresses the smooth erasure of complex scattering image features, ensuring that the generator no longer sacrifices the underlying acoustic realism in order to output a smooth detection image that conforms to conventional human vision.

[0139] Alternatively, while reducing the weight of the first discrimination branch solely at the algorithmic level can prevent real, complex defects from being obscured, the limited amount of physical information acquired by conventional phased array ultrasonic testing equipment in fixed-focus scanning mode means the second discrimination branch still lacks sufficient multi-angle physical evidence to accurately reconstruct these irregular microcracks. Therefore, simultaneously with executing the aforementioned weight dynamic takeover action, the system issues a full-matrix capture switching command to the phased array ultrasonic testing equipment.

[0140] Specifically, the full-matrix capture switching command is a hardware control message with the highest interrupt priority sent from the control terminal to the phased array front-end host. This full-matrix capture switching command is used to force the phased array ultrasonic testing equipment to stop the current conventional focusing law scan and start the full-matrix capture mode in which all array elements transmit and receive independently in sequence for the target depth range that generates the physical timing consistency signal.

[0141] In conventional focusing law scanning, such as sector scanning or linear scanning, multiple elements of a phased array probe simultaneously emit ultrasonic waves under the control of a set delay law to form a focused sound beam at a specific location in space. Subsequently, multiple elements receive signals according to a similar delay law and synthesize a single radio frequency signal. Although this mode offers high detection speed, the synthesized signal loses the independent phase and amplitude spatial distribution information between each element. When encountering microcracks with polyhedral morphology, conventional synthesized sound beams are prone to irreversible physical energy cancellation.

[0142] The full-matrix capture mode is an exhaustive low-level data acquisition technology that does not rely on any preset physical focus. Let the phased array ultrasonic testing equipment's probe contain a total of [number] array elements. Each individual piezoelectric crystal, in full-matrix capture mode, the system controls the first array element to emit a broadband ultrasonic pulse independently, while the others... Each array element simultaneously activates its receiving channel and independently records the echo signal; subsequently, the second array element is controlled to transmit independently, while the others... Each array element receives the signal simultaneously again; this cycle repeats until the [number]th array element is received. Each array element completes its individual transmission and is received by all array elements. Through one complete capture cycle, the hardware layer will acquire a total of A vast dataset of independent time-series acoustic signals.

[0143] It should also be noted that, due to the extremely large amount of data generated by the full matrix capture mode, if the entire effective detection depth of the wind turbine bolts is captured along the entire time axis, it will lead to communication bandwidth blockage and subsequent deep learning network memory overflow.

[0144] Therefore, the switching instruction in this embodiment explicitly limits the activation of this mode to the target depth range that generates the physical timing consistency signal. Let... This represents the initial arrival time of the signal with physical consistency characteristics identified by the second discriminant branch. Let... Indicates the termination time of this signal cluster, let... This represents the preset time redundancy to ensure a complete envelope of the scattered wave train. Therefore, the data acquisition window opening time inside the phased array device is reconfigured to... The data acquisition window closing time has been reconfigured to By employing this hardware-level time-domain truncation within a depth range, the system can acquire full matrix data containing complete spatial scattering information while significantly reducing the amount of invalid background data transmitted.

[0145] For example, after the phased array ultrasonic testing equipment completes the exhaustive acoustic transceiver at the lower level, the system performs the step of acquiring the lower level acoustic matrix data generated by the full matrix capture mode and inputting the lower level acoustic matrix data into the second discrimination branch for auxiliary verification.

[0146] Specifically, the underlying acoustic matrix data not only contains rich physical information, but also requires specific reconstruction algorithms to extract its deep features before it can be effectively input into the second discriminant branch. Let... The spatial index variable represents the transmitting array element. This represents the spatial index variable of the receiving array element, where and The value range is from 1 to A positive integer. Let... Indicates the first Each element is launched, the first The time variable received by each array element The one-dimensional time-domain ultrasonic signal waveform. The underlying acoustic matrix data acquired by the full matrix capture mode can then be represented as a time-domain... The three-dimensional tensor basis is the independent variable.

[0147] To perform high-resolution physical inversion verification of complex scattering signals, the system incorporates a fully focused ultrasonic reconstruction algorithm in the front-end preprocessing module of the second discrimination branch. This algorithm utilizes the acquired full matrix data to coherently superimpose and reconstruct the acoustic intensity of arbitrary spatially discrete pixels within the target depth range. Let... This represents the two-dimensional coordinates of the spatial physical pixel point to be verified within the target depth range, where Indicates the horizontal position coordinates. Represents the position coordinates in the depth direction. Let... Indicates the first The spatial lateral center coordinates of each transmitting element on the probe surface are set as follows: Indicates the first The spatial lateral center coordinates of each receiving array element on the probe surface. Let... This represents the sound velocity constant that represents the propagation of ultrasonic waves within the steel matrix of the fan bolt.

[0148] In the full-focus reconstruction algorithm, the distance from the u-th emission element to the target pixel is first calculated. The theoretical acoustic flight time of the signal reflected back to the v-th receiving element. Let... This represents the theoretical acoustic time of flight. Based on the Euclidean distance formula and the laws of wave dynamics propagation, the algorithm for calculating this time of flight is defined as follows: ; Subsequently, based on the calculated theoretical acoustic time of flight, the underlying time-domain signal waveform was analyzed. Extract the instantaneous amplitude value at the corresponding moment, and then combine all... The amplitude values ​​of the transmit / receive combined link are coherently superimposed. Let... Indicates at the target pixel point The value of the focused physical acoustic intensity after full matrix coherent synthesis is given. The formula for calculating this value is: ;

[0149] Through the matrix operations described above, the system converts the discrete bottom-level time-domain signal into an acoustic intensity tensor spectrum with extremely high spatial resolution and fidelity within a specified depth range. Because this coherent superposition algorithm considers the precise geometric distance from all array elements within the probe aperture to the target point in its time delay calculation, it can effectively refocus the acoustic energy that is dissipated and misaligned during conventional scanning due to surface roughness and polyhedral reflections onto the actual spatial location of the microcrack, significantly improving the signal-to-noise ratio of complex defects in the physical dimension.

[0150] like Figure 7The figure presents the acoustic field energy distribution of the star-shaped fatigue microcracks after high-resolution physical inversion within the target depth range. The horizontal axis represents the lateral position in millimeters, ranging from 10 to 30; the vertical axis represents the depth position in millimeters, increasing downwards along the depth direction, ranging from 40 to 60. The color bars on the right indicate the acoustic intensity values ​​after coherent superposition and reconstruction, with the color changing from dark blue to dark red representing a gradual increase in acoustic intensity.

[0151] In the illustrated two-dimensional physical space, the dark blue and light blue areas represent the normal structural background of the bolt matrix and weak grain boundary scattering noise. At the center, 20 mm laterally and 50 mm deep, an irregular, bright energy accumulation zone composed of dark red and orange hues is observed. This bright area is not a single, regular geometric patch, but rather extends outwards with multiple secondary bright branches, objectively reproducing the star-shaped branching and polyhedral spatial distribution morphology formed by microcracks accompanied by intergranular corrosion.

[0152] This high-contrast energy mapping effect shows that the system drives the underlying hardware to perform exhaustive transmission and reception of all array elements, and uses a full-focus reconstruction algorithm to perform precise time delay calculation and coherent superposition of acoustic intensity on discrete pixels within a specified depth range. This overcomes the phase cancellation and energy divergence problems caused by conventional fixed-focus law scanning when facing microcracks with polyhedral morphology, and accurately reconstructs the originally scattered physical reflection signals to the real spatial location of the microcrack.

[0153] The reconstructed high-resolution acoustic intensity spectrum, as supplementary physical evidence, is input into the second discriminant branch. This guides the network model to reconstruct the image based on underlying objective physical laws, thereby improving the detection reliability and morphological accuracy of complex stress cracks in ultrasonic nondestructive testing.

[0154] Subsequently, the acoustic intensity tensor spectrum obtained from the above calculation and the original... The underlying signal set, serving as supplementary physical evidence, is input into the second discriminant branch, which is in a high-weight state, for secondary physical verification. At this point, the second discriminant branch no longer relies solely on the one-dimensional signal generated by the initial conventional scan, but instead receives exhaustive acoustic field feature support from the lowest level of the hardware. Guided by the strong gradient dominated by the second discriminant branch, the generator is forced to reconstruct the image of the region based on this detailed physical evidence. This results in a final ultrasound image that not only eliminates clutter artifacts but also realistically reproduces the morphology of minute stress cracks with complex scattering characteristics, thus forming a robust and realistic physical preservation loop.

[0155] Analysis of existing technologies reveals a polarized technical bottleneck in current artifact removal methods for industrial nondestructive testing: some methods rely solely on signal processing circuits embedded in front-end hardware, using simple threshold filtering or spatial averaging algorithms for noise reduction. This purely physical approach often struggles to distinguish noise from genuine defect signals when faced with coupling interference from wind turbine bolt ends or severe structural artifacts, resulting in a persistently low image signal-to-noise ratio.

[0156] Another set of newer methods directly incorporates deep learning computer vision techniques, such as conventional generative adversarial networks (GANs). However, these purely vision-driven algorithms treat ultrasonic data as equivalent to ordinary optical images, lacking a deep understanding of the underlying acoustic physics within their networks. When encountering microscopic cracks that produce complex diffuse reflections, visual algorithms tend to perceive these irregular, scattered spots as noise that disrupts image smoothness, leading to pattern collapse and highly destructive image forgery and defect erasure.

[0157] The technical solution of this embodiment cleverly overcomes the technical barrier between simple software algorithms and underlying physical hardware. By designing an independent visual and physical dual-branch architecture within the discriminator of the neural network and introducing a real-time difference divergence monitoring mechanism, the system can keenly detect the precursors of the algorithm's impending over-smoothing and fraud. Once the relative difference value exceeds the limit, the system does not passively stop calculation, but actively executes a weight takeover action, stripping the visual branch of its aesthetic dominance and forcibly pulling the evaluation criteria back to the objective dimension of physical acoustics.

[0158] More notably, this technical solution directly uses the conflicting features of the loss function within deep learning as the triggering condition for higher-order functions of the underlying hardware. When the performance of the algorithm is limited in processing complex data, the system forces the phased array ultrasonic hardware to activate a full-matrix capture mode, acquiring the most original and comprehensive exhaustive matrix data from the physical sound source. Subsequently, a full-focus reconstruction algorithm is used to feed back the massive amount of physical evidence collected by the hardware to the physical discrimination branch, which is in a high-weight state. This deep fusion and closed-loop verification design, which triggers full-matrix capture of the hardware through algorithm divergence warning and feeds back exhaustive evidence of the hardware to the algorithm for faithful generation, greatly suppresses the hallucination generation behavior of generative adversarial networks when processing complex non-stationary data.

[0159] The overall solution not only significantly suppresses and removes various artifacts on wind turbine bolts under harsh working conditions, but also safeguards the morphological authenticity of dangerous defects such as irregular fatigue microcracks with rigorous physical logic, fundamentally improving the industrial reliability and engineering application value of AI-assisted ultrasonic non-destructive testing technology.

[0160] Example 4:

[0161] like Figure 8As shown, this embodiment provides a phased array ultrasonic artifact removal system for wind turbine bolts based on generative adversarial networks. As can be seen from the background technology, existing ultrasonic artifact removal methods are usually static and unidirectional. The fixed deep learning models are only used as post-processing tools in practical applications, exhibiting technical shortcomings such as being unable to perceive and adapt to the dynamic degradation of the physical detection conditions of the front-end hardware.

[0162] Specifically, the wind turbine bolt phased array ultrasonic artifact removal system based on generative adversarial networks described in this embodiment is applied to a control terminal connected to a phased array ultrasonic testing device. In the actual physical architecture of a wind turbine high-altitude operation site, the phased array ultrasonic testing device is typically an industrial-grade portable non-destructive testing host containing a multi-channel ultrasonic pulse generator and receiving circuit. The front end of this host is connected to a phased array ultrasonic probe via a cable. The probe integrates an attitude microelectromechanical sensor and an electrical impedance monitoring circuit for monitoring the fit status. The control terminal is typically a ruggedized industrial computer or edge computing server equipped with a high-performance graphics processing unit (GPU) and a multi-core central processing unit (CPU). A low-latency data communication link is established between the phased array ultrasonic testing device and the control terminal via industrial Ethernet (such as TCP / IP or UDP communication protocols) or a high-speed USB bus. The system includes: an initialization module, a deviation calculation module, a parameter adjustment module, an image generation module, and a verification and output module.

[0163] First, the system includes an initialization module, used to respond to the start signal of the phased array ultrasonic testing equipment by establishing a device reference library containing the initial impedance of the probe, the initial thickness of the coupling agent, and the initial angle of the probe, and initializing the generative adversarial network model. At the beginning of the on-site testing cycle, when the testing personnel attach the phased array probe to the end face of the bolt to be tested inside the wind turbine hub and start the testing system, the front-end phased array ultrasonic testing equipment sends a start signal containing hardware handshake information to the control terminal. Upon receiving this signal, the initialization module in the control terminal allocates dedicated storage space in the memory or solid-state drive as the device reference library.

[0164] Simultaneously, by reading the probe's electrical impedance value currently reported by the sensor, the coupling medium layer thickness calculated based on high-frequency acoustic time measurement feedback, and the probe's spatial tilt angle obtained by the built-in gyroscope, these values ​​are used as the health status benchmark for this detection task and permanently written into the storage space. Concurrently, the initialization module loads the pre-trained adversarial generative network model weight parameters from non-volatile storage media into the GPU memory of the control terminal, completing the construction of the tensor computation graph and the allocation of system inference resources.

[0165] Secondly, the system includes a deviation calculation module, used to acquire real-time ultrasonic data and real-time status parameters collected by the phased array ultrasonic testing equipment, and calculate the state deviation value between the real-time status parameters and the corresponding parameters in the equipment reference library. During the continuous scanning of wind turbine bolts, the physical state of the probe will undergo dynamic displacement due to the influence of tower wind load vibration or human operation fatigue. The background monitoring process of the control terminal continuously reads the real-time acoustic raw data matrix (i.e., real-time ultrasonic data) uploaded by the front-end equipment through the communication bus, and simultaneously acquires the real-time sensor readings (i.e., real-time status parameters) attached to the header of the communication data packet. The deviation calculation module calls the central processing unit to perform the difference calculation and normalization processing between the real-time status parameters and the fixed parameters in the equipment reference library, and outputs the quantified state deviation value. This state deviation value serves as an objective quantitative indicator of the degree of environmental degradation within the system.

[0166] Furthermore, the system includes a parameter adjustment module for dynamically adjusting the acquisition parameters of the phased array ultrasonic testing equipment based on the state deviation value to obtain target ultrasonic data. Traditional denoising systems passively receive data in only one direction, while the parameter adjustment module in this embodiment possesses downlink control hardware capabilities. When the state deviation value changes significantly, the parameter adjustment module compiles specific low-level hardware control messages and sends them to the front-end phased array ultrasonic testing equipment. For example, after receiving the instruction, the field-programmable gate array (FPGA) of the phased array host dynamically adjusts the transmission voltage amplitude of the ultrasonic pulse, increases the sampling frequency, or modifies the internal focusing delay rule of the hardware according to the instruction content. Through this closed-loop intervention at the physical level, the front-end hardware can actively compensate for acoustic energy loss caused by adverse environmental conditions, and then output target ultrasonic data with a higher physical signal-to-noise ratio to the host computer after optimized compensation.

[0167] After acquiring the underlying hardware compensation data, the system enters the upper-level algorithm processing stage, including an image generation module. This module determines constraint weights based on the state deviation values ​​and inputs the target ultrasound data into the Generative Adversarial Network (GAN) model. The constrained weights are then combined to generate an ultrasound image with artifacts removed. The logic unit within the image generation module extracts the state deviation values ​​output from the preceding steps and, based on a preset mathematical mapping function, recalculates and allocates the amplitude-preserving and shape-preserving constraint weights in the deep learning model's loss function. Subsequently, the target ultrasound data is converted into a multidimensional tensor and input into the GAN model loaded in GPU memory for forward propagation inference. Because this model incorporates constraint weights that are dynamically adjusted according to external physical conditions during inference, its algorithmic behavior can adaptively compensate for data distribution drift caused by probe offset or poor coupling, thereby objectively preserving the true physical characteristics of minor defects and suppressing misjudgments of dynamic variation artifacts.

[0168] Finally, to ensure the security of industrial testing results, the system includes a verification and output module, used to perform multi-dimensional physical feature verification based on the real-time state parameters and the artifact-removed ultrasonic image. If the verification passes, the artifact-removed ultrasonic image is output, and a reset command is sent to the phased array ultrasonic testing equipment to restore the parameters in the equipment's reference library. This module, independent of the generative adversarial network model, acts as a physical safety net for the final result. After the deep learning model outputs its results, this module calls a pre-set acoustic rule library to extract indicators such as attenuation rate and sound path distribution from the output image, and performs rigorous acoustic logic verification against the current physical-level real-time state parameters.

[0169] This step effectively prevents image distortion that may occur in pure data-driven algorithms under complex working conditions. When all physical verifications meet the safety thresholds, the graphical interface of the control terminal will render, display, and archive the artifact-free ultrasonic image. At the same time, the module will automatically encapsulate and send a reset control command to the front-end phased array ultrasonic testing equipment, causing the various acquisition parameters in its internal registers to revert to the initial reference values, so as to ensure that the physical starting point of the next wind turbine bolt inspection cycle has consistency and comparability.

[0170] Combining the aforementioned modular architecture with the actual hardware operation mechanism, the wind turbine bolt phased array ultrasonic artifact removal system based on adversarial generative networks provided in this embodiment establishes a two-way collaborative closed loop within the control terminal, from front-end hardware physical perception to back-end software algorithm constraints. Through data flow and signaling interaction between modules, the system effectively resolves the contradiction between traditional static algorithms and dynamic adverse working conditions. In field applications, this architecture not only achieves real-time linkage between equipment physical data acquisition and artificial intelligence inference, but also constructs a security defense line for physical feature verification at the output terminal, improving the stability and reliability of removing complex structural artifacts and external noise interference, and possessing high engineering implementation value and industrial applicability.

[0171] Example 5:

[0172] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0173] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0174] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0175] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0176] The memory 103 stores a computer program corresponding to the method for removing phased array ultrasonic artifacts from wind turbine bolts based on generative adversarial networks according to the above embodiments of the present invention. This computer program is executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0177] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0178] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for removing phased array ultrasonic artifacts from wind turbine bolts based on generative adversarial networks, characterized in that, The method, applied in a control terminal connected to a phased array ultrasonic testing device, includes: In response to the start signal of the phased array ultrasonic testing device, a device reference library containing the initial impedance of the probe, the initial thickness of the coupling agent, and the initial angle of the probe is established, and the generative adversarial network model is initialized. The system acquires real-time ultrasonic data and real-time status parameters collected by the phased array ultrasonic testing device, calculates the status deviation value between the real-time status parameters and the corresponding parameters in the device's reference library, and dynamically adjusts the acquisition parameters of the phased array ultrasonic testing device based on the status deviation value to acquire target ultrasonic data. The constraint weights are determined based on the state deviation values, and the target ultrasound data is input into the adversarial generative network model to generate an ultrasound image with artifacts removed by combining the constraint weights. Multi-dimensional physical feature verification is performed based on the real-time state parameters and the artifact-removed ultrasound image; If the verification passes, the artifact-free ultrasound image is output, and a reset command is sent to the phased array ultrasound testing device to restore the parameters in the device's reference library.

2. The method according to claim 1, characterized in that, The step of acquiring real-time ultrasonic data and real-time status parameters collected by the phased array ultrasonic testing equipment, and calculating the status deviation value between the real-time status parameters and the corresponding parameters in the equipment reference library, includes: Obtain the real-time impedance of the probe, the real-time thickness of the coupling agent, and the real-time angle of the probe. Calculate the first state deviation between the real-time impedance of the probe and the initial impedance of the probe, the second state deviation between the real-time thickness of the coupling agent and the initial thickness of the coupling agent, and the third state deviation between the real-time angle of the probe and the initial angle of the probe, respectively. The step of dynamically adjusting the acquisition parameters of the phased array ultrasonic testing equipment based on the state deviation value includes: If the first state deviation is greater than the preset impedance threshold, then the sampling rate and the number of sampling points of the phased array ultrasonic testing device are increased. If the second state deviation is greater than the preset thickness threshold, the ultrasonic emission power is reduced and a coupling agent replenishment prompt signal is generated. If the deviation of the third state is greater than the preset angle threshold, the focal length of the device is dynamically adjusted according to the deviation of the third state.

3. The method according to claim 1, characterized in that, The initialization of the adversarial generative network model includes pre-training the adversarial generative network model using a training set; The steps for obtaining the training set include: Based on the prior geometric features of bolts, structural artifacts that conform to the periodic reflection law are identified, and structural artifact labels are generated. Based on the theoretical attenuation law of ultrasound, the consistency of ultrasound data in multiple frequency dimensions is checked, and the regions where the amplitude attenuation curve deviates from the theoretical attenuation law of ultrasound are marked as pseudo-labels of attenuation anomalies. The system integrates manually annotated real defect areas, structural artifact labels, and attenuation anomaly pseudo-labels, and calculates the noise interference level of the pixel based on the texture similarity between the target pixel and the reference pixel. Based on the degree of noise interference, the ultrasound images are divided into core interference regions, blurred regions, and non-interference regions, which are then used to construct the training set.

4. The method according to claim 3, characterized in that, The loss function of the adversarial generative network model includes a dynamic constraint loss, which includes an amplitude-preserving loss term and a shape-preserving loss term. The step of determining the constraint weights based on the state deviation values ​​includes: Based on the noise interference level of the pixel, the comprehensive domain difference factor obtained by weighted fusion of the state deviation values, the probe wear correction factor associated with the first state deviation, and the coupling thickness correction factor associated with the second state deviation, the amplitude preservation weight of the amplitude preservation loss term and the shape preservation weight of the shape preservation loss term are dynamically calculated. The amplitude preservation weight is positively correlated with the noise interference level, while the shape preservation weight is negatively correlated with the noise interference level. The value of the probe wear correction factor increases segmentally as the first state deviation increases, and the value of the coupling thickness correction factor increases segmentally as the second state deviation increases.

5. The method according to claim 2, characterized in that, The step of determining the constraint weights based on the state deviation values ​​further includes an adaptive adjustment step based on the comprehensive domain difference factor: A three-dimensional domain label is constructed, which includes equipment dimension, operating condition dimension and focusing law dimension, and a domain discriminator is constructed in the adversarial generative network model to learn common features between domains; The first state deviation, the second state deviation, and the third state deviation are normalized respectively, and the device domain difference value is obtained by weighted calculation. The preset initial three-dimensional domain difference value is weighted and fused with the device domain difference value to obtain the comprehensive domain difference factor; Based on the numerical range of the comprehensive domain difference factor, dynamically allocate adversarial loss weights and domain adaptive loss weights; Specifically, when the comprehensive domain difference factor increases and crosses a preset adjustment range, the adversarial loss weight is reduced and the domain adaptive loss weight is increased simultaneously.

6. The method according to claim 1, characterized in that, The adversarial generative network model includes a generator and a discriminator; The step of inputting the target ultrasound data into the generative adversarial network model includes: The working condition parameters, frequency information, focusing law parameters, and bolt geometric parameters are standardized and encoded to generate a condition vector; The conditional vector is mapped through a fully connected layer, concatenated with the feature map of the target ultrasound data, and then input into the generator; The discriminator adopts a dual-branch structure, wherein the first discriminant branch is used to determine the authenticity features of the input image, and the second discriminant branch is used to determine the physical temporal consistency features and cross-frequency consistency features of the input ultrasound signal.

7. The method according to claim 1, characterized in that, The multi-dimensional physical feature verification based on the real-time state parameters and the artifact-removed ultrasound image includes the following parallel verification steps: First dimension verification: Extract the depth amplitude data corresponding to the artifact-removed ultrasound image, calculate the attenuation curve, and verify the deviation of the attenuation curve from the theoretical ultrasound attenuation curve. Second dimension verification: Based on the ultrasound image with artifacts removed, the focus parameters are deduced and the matching degree between the focus parameters and the real-time focusing law parameters of the phased array ultrasound detection device is verified. Third-dimensional verification: Compare the attenuation rate trends of high-frequency signals and low-frequency signals in the target ultrasound data; Fourth dimension verification: Generate differential correlation mapping between device status change parameters and image data feature changes to verify the correspondence between artifact distribution patterns and abnormal status parameters; When the first dimension verification, the second dimension verification, the third dimension verification, and the fourth dimension verification all meet their respective set qualification thresholds, the multi-dimensional physical feature verification is determined to be passed.

8. The method according to claim 7, characterized in that, The method also includes confidence grading and closed-loop control steps based on the verification results: Calculate the overall confidence level based on the scores of the parallel verification steps; Extract the preset high confidence threshold and the medium confidence threshold that is lower than the high confidence threshold; If the overall confidence level is less than the high confidence threshold and greater than or equal to the medium confidence threshold, then a parameter adjustment command is sent to the phased array ultrasonic testing device, and a secondary verification process is triggered after the ultrasonic data is reacquired. If the overall confidence level is less than the medium confidence level threshold, a full calibration alarm signal is triggered, and the system reverts to the original ultrasound data mode that includes risk area markings.

9. The method according to claim 5, characterized in that, Following the dynamic allocation of adversarial loss weights and domain adaptive loss weights, a hardware-software collaborative intervention step is also included: A sliding time window is constructed, the rate of change of the comprehensive domain difference factor within the sliding time window is calculated in real time, and the comprehensive domain difference factor is compared with a preset extreme degradation threshold, and the rate of change is compared with a preset oscillation threshold. If the comprehensive domain difference factor is greater than the extreme degradation threshold, or the rate of change is greater than the oscillation threshold, then the parameter circuit breaker protection mechanism is triggered. The parameter circuit breaker protection mechanism includes: stopping the execution of the allocation instruction to reduce the adversarial loss weight and increase the domain adaptive loss weight, forcibly locking the adversarial loss weight and the domain adaptive loss weight to the weight ratio when the parameter circuit breaker protection mechanism is triggered, and freezing the parameter update action of the feature extraction layer in the adversarial generative network model. Simultaneously with triggering the parameter fuse protection mechanism, a multi-pulse hardware compensation command is sent to the phased array ultrasonic testing equipment to force the phased array ultrasonic testing equipment to switch to multi-pulse hardware superposition mode. Acquire compensated ultrasound data in the multi-pulse hardware superposition mode, and input the compensated ultrasound data into the weight-locked adversarial generative network model for artifact removal processing.

10. A phased array ultrasonic artifact removal system for wind turbine bolts based on generative adversarial networks, characterized in that, The system, used in a control terminal connected to a phased array ultrasonic testing device, includes: An initialization module is used to respond to the start signal of the phased array ultrasonic testing device, establish a device reference library including the probe initial impedance, the initial thickness of the coupling agent and the probe initial angle, and initialize the generative adversarial network model. The deviation calculation module is used to acquire real-time ultrasonic data and real-time status parameters collected by the phased array ultrasonic testing equipment, and to calculate the status deviation value between the real-time status parameters and the corresponding parameters in the equipment reference library. The parameter adjustment module is used to dynamically adjust the acquisition parameters of the phased array ultrasonic testing device according to the state deviation value in order to obtain target ultrasonic data. The image generation module is used to determine the constraint weights based on the state deviation value, input the target ultrasound data into the generative adversarial network model, and generate an ultrasound image with artifacts removed by combining the constraint weights. The verification and output module is used to perform multi-dimensional physical feature verification based on the real-time status parameters and the artifact-removed ultrasound image; if the verification passes, the artifact-removed ultrasound image is output, and a reset command is sent to the phased array ultrasound detection device to restore the parameters in the device's reference library.

Images