Ultrasonic intelligent nondestructive testing method based on pose adjustment of probabilistic convolutional neural network

Abstract

The present application relates to an ultrasonic intelligent nondestructive testing method based on pose adjustment of a probabilistic convolutional neural network, and belongs to the technical fields of mechanical engineering, optical engineering, nondestructive testing and automated detection. The method comprises the following steps: initializing ultrasonic testing parameters; generating a standard scanning path for a mechanical arm; pre-processing ultrasonic waveform signals; ultrasonic imaging and defect identification using a probabilistic convolutional neural network; inputting the ultrasonic imaging results into the probabilistic convolutional neural network for defect identification processing, to obtain a classification result of the presence or absence of a defect, a corresponding predicted uncertainty of the defect and defect size information; and determining the classification result of the presence or absence of a defect and the corresponding predicted uncertainty output by the probabilistic convolutional neural network. By introducing a deep learning model to predict the uncertainty as an online feedback quantity, a scanning control mechanism based on uncertainty determination is established, so that the detection process can dynamically adjust the scanning path and the probe pose according to the reliability of the identification result.

Description

Technical Field

[0001] This invention relates to the fields of mechanical engineering, optical engineering, nondestructive testing, and automated testing technology, and in particular to an ultrasonic intelligent nondestructive testing method based on pose adjustment using a probabilistic convolutional neural network. Background Technology

[0002] With the increasing demand for component manufacturing and service testing in fields such as aerospace and high-end equipment, robotic automated non-destructive testing technology has been widely applied. By using multi-degree-of-freedom robotic arms equipped with ultrasonic probes, automatic scanning, inspection, and imaging of large-sized, curved, or complex geometrically structured components can be achieved. This improves inspection efficiency and reduces reliance on manual labor, gradually becoming an important technical means for in-situ inspection of complex components.

[0003] In existing robotic ultrasonic testing systems, the geometric or reference information of the component under test is typically acquired first to calibrate the coordinates between the robotic arm, the detection sensor, and the component, thereby generating a scanning path covering the target surface. During the testing process, the robotic arm drives the probe to move along a preset path, maintaining the probe's attitude and coupling state through position control, force control, or distance control, while simultaneously acquiring ultrasonic signals and pose information. The acquired ultrasonic echo signals are filtered and processed to generate one-dimensional waveforms, two-dimensional, and three-dimensional images, which are used to characterize the internal structural features of the component. Regarding defect identification and analysis, existing technologies typically combine physical mechanism-based signal feature analysis methods with data-driven identification methods to classify, locate, and evaluate the dimensions of defects in the imaging results. In recent years, deep learning models have been widely applied to the automatic identification of ultrasonic imaging results. By training on two-dimensional or three-dimensional imaging data, functions such as defect presence / absence determination, region segmentation, and size regression are achieved. The identification results are then integrated with pose information to generate an inspection report for component quality assessment and process control.

[0004] Current robotic ultrasonic inspection technologies, when combined with deep learning for defect identification, typically only output deterministic classification or dimensional results, lacking quantitative analysis of the reliability of the predicted results. Because the identification results do not contain uncertain information, the model output may fluctuate or even misclassify when affected by factors such as coupling fluctuations, posture deviations, or signal noise during the inspection process. However, the system cannot determine whether the current identification result is stable and reliable, easily leading to missed or false detections. Furthermore, existing inspection processes mainly follow preset scanning paths, with the identification results serving only as the final interpretation basis. There is a lack of online feedback mechanisms based on the reliability of the identification, making it difficult to adjust the path and correct the posture in a timely manner during inspection. When the quality of the inspection data is abnormal or the identification results are unstable, it is usually necessary to identify the problem through offline analysis after the inspection is completed, followed by rescanning or re-inspection, which affects inspection efficiency. Summary of the Invention

[0005] This invention aims to address the technical problem in existing robotic ultrasonic testing technologies that lack quantitative analysis of the reliability of prediction results when combining deep learning for defect identification, and provides an ultrasonic intelligent non-destructive testing method based on pose adjustment using probabilistic convolutional neural networks.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] An ultrasonic intelligent nondestructive testing method based on pose adjustment using probabilistic convolutional neural networks, the applicable testing system of which includes: a control and testing system and a sample to be tested; the control and testing system is used to test the sample to be tested according to a preset scanning path;

[0008] The control and testing system includes: a host computer, a signal generator / receiver, a robotic arm controller, a robotic arm, a phased array ultrasonic probe, and a coupling agent; the host computer is connected to the signal generator / receiver and the robotic arm controller respectively; the signal generator / receiver is connected to the phased array ultrasonic probe; the robotic arm controller is connected to the robotic arm; the phased array ultrasonic probe is located at the front end of the robotic arm; the lower surface of the phased array ultrasonic probe is provided with a coupling agent;

[0009] The host computer is used to control the signal generator / receiver to generate excitation waveforms and transmit them to the phased array ultrasonic probe; the host computer is also used to control the robotic arm controller, thereby operating the movement of the robotic arm;

[0010] The method includes the following steps:

[0011] Step 1: Initialize ultrasound detection parameters;

[0012] The parameters for generating the excitation waveform are set via the host computer;

[0013] Step 2: Generate the standard scanning path for the robotic arm;

[0014] The host computer generates a standard scanning path covering the detection area;

[0015] Step 3: Ultrasonic signal preprocessing; processing the raw ultrasonic echo signals acquired during the detection process;

[0016] Step 4: Ultrasonic imaging and probabilistic convolutional neural network to identify defects;

[0017] The ultrasound imaging results are input into a probabilistic convolutional neural network for defect identification processing to obtain the classification results of whether the defect exists, the prediction uncertainty of the defect, and the defect size information.

[0018] Step 5: Determine whether the defects output by the probabilistic convolutional neural network have classification results and their corresponding prediction uncertainties.

[0019] In the above technical solution, step 3 specifically includes:

[0020] First, the raw ultrasonic echo signals collected during the detection process are bandpass filtered.

[0021] Then, the envelope signal of the processed echo signal is extracted by Hilbert transform, and the amplitude of the envelope signal is normalized. Finally, the effective signal segment within the set sound path range is extracted for subsequent ultrasound imaging processing.

[0022] In the above technical solution, step 5 specifically includes:

[0023] Step 5.1: Compare the prediction uncertainty with the preset threshold. When the prediction uncertainty is lower than or equal to the preset threshold, the current detection result is determined to be reliable, and the robotic arm continues to perform the detection task according to the standard scanning path. After the robotic arm completes the global scan of the scanning path, it outputs the spatial coordinate information of the detection point determined to have a defect, the ultrasonic imaging image of the corresponding position, and the defect size information predicted by the probabilistic convolutional neural network.

[0024] Step 5.2: Compare the prediction uncertainty with the preset threshold. When the prediction uncertainty is higher than the preset threshold, it is determined that there is recognition uncertainty at the current detection location. Then, it is further determined whether the number of repeated detections in the detection area has reached the preset upper limit.

[0025] When the maximum number of repeated tests has not been reached, the phased array ultrasonic probe is repeatedly adjusted in orientation. During the probe orientation adjustment process, ultrasonic imaging and recognition processing is performed again, and the corresponding prediction uncertainty is calculated.

[0026] When the maximum number of repeated inspections is reached, the spatial coordinates of the inspection area are recorded, and the robotic arm continues to execute the standard scanning path. After the robotic arm completes the global scan of the scanning path, it outputs the spatial coordinates of the inspection points that are determined to have defects, the ultrasonic imaging image of the corresponding location, and the defect size information predicted by the probabilistic convolutional neural network. At the same time, it records the coordinates of the suspicious points that have reached the maximum number of repeated inspections and the corresponding uncertainty data, thus completing the inspection process.

[0027] In the above technical solution, step 5.2, the probe attitude adjustment includes fine-tuning of the front and rear positions and multi-degree-of-freedom angle adjustment.

[0028] In the above technical solution, the fine-tuning of the front and rear positions specifically refers to the forward and backward movement distance of the phased array ultrasonic probe. ;

[0029] Multi-degree-of-freedom angle adjustment specifically includes:

[0030] Phased array ultrasonic probe surround Total rotation angle of axis rotation Phased array ultrasonic probe surround Total rotation angle of axis rotation Phased array ultrasonic probe surround Total rotation angle of axis rotation ;

[0031] in, The axis is horizontal. The axis is with The horizontal direction perpendicular to the axis The axes are respectively with shaft and The vertical direction perpendicular to the axis.

[0032] In the above technical solution, in step 4, the prediction uncertainty corresponding to the defect is:

[0033] The prediction variance or standard deviation of the output of a probabilistic convolutional neural network based on a random weight perturbation method; or

[0034] The predicted distribution and its corresponding prediction variance, prediction entropy, or confidence interval based on the random deactivation Monte Carlo sampling method; or the predicted distribution and its corresponding prediction variance, prediction entropy, or confidence interval based on the output of a multi-model ensemble network; or

[0035] The predicted distribution based on the output of a Bayesian neural network and its corresponding prediction variance, prediction entropy, or confidence interval.

[0036] In the above technical solutions, the scanning path is: serpentine scanning, grid scanning, or circular arc scanning;

[0037] The serpentine scanning method is as follows: after the phased array ultrasonic probe completes a straight line scan in the first direction, it moves a preset step distance in the second direction and then performs the next straight line scan in the opposite direction. This process is repeated to form a continuous zigzag scanning path to cover the entire detection area.

[0038] Grid scanning involves dividing the area to be detected into several regular grid points, and the robotic arm moves to each grid point in a preset row and column order, and completes the acquisition of ultrasonic signals at each grid point.

[0039] Circular scanning involves a robotic arm driving a phased array ultrasonic probe along a circular arc trajectory based on a preset center point or structural contour, performing ultrasonic detection at multiple sampling positions along the circular arc path.

[0040] The present invention has the following beneficial effects:

[0041] This invention presents an ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment. By introducing a deep learning model (probabilistic convolutional neural network) to predict uncertainty as an online feedback quantity, a scanning control mechanism based on uncertainty determination is established. This allows the testing process to dynamically adjust the scanning path and probe posture according to the reliability of the identification results. When the predicted uncertainty increases, local key scanning and posture fine-tuning are automatically triggered. When the uncertainty decreases below a threshold, standard scanning is resumed. Points with high uncertainty in the deep learning model are recorded and reported, thereby reducing unnecessary manual rescanning, lowering the risk of missed detections, improving detection efficiency and result reliability, and achieving a balance between detection accuracy and efficiency. Attached Figure Description

[0042] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 This is a schematic diagram of an ultrasound image in full matrix capture mode.

[0044] Figure 2 This is a schematic diagram of a probabilistic convolutional neural network structure based on random weight perturbation.

[0045] Figure 3 This is a schematic diagram of the detection system applicable to the ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment of the present invention.

[0046] Figure 4 This is a schematic diagram of the probe attitude adjustment process, which involves adjusting the position of the phased array ultrasonic probe along the scanning direction.

[0047] Figure 5 For phased array ultrasonic probes A schematic diagram of the probe attitude adjustment process for adjusting the axis angle.

[0048] Figure 6 For phased array ultrasonic probes A schematic diagram of the probe attitude adjustment process for adjusting the axis angle.

[0049] Figure 7 This is a flowchart illustrating the ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment according to the present invention.

[0050] The reference numerals in the figure are:

[0051] 1-Control and testing system; 2-Sample under test;

[0052] 11-Host computer; 12-Signal generator / receiver; 13-Robot arm controller; 14-Robot arm;

[0053] 15-Phase array ultrasound probe; 16-Couplet; 17-Scanning path;

[0054] 21-Sample defect; 22-Suspicious area;

[0055] For the phased array ultrasonic probe along the edge of the sample being tested directional movement distance;

[0056] For phased array ultrasonic probes to rotate around the sample under test Adjust the total rotation angle by rotating the axis;

[0057] For phased array ultrasonic probes to rotate around the sample under test Adjust the total rotation angle by rotating the axis;

[0058] For phased array ultrasonic probes to rotate around the sample under test Adjust the total rotation angle by rotating the axis;

[0059] , , These are the coordinate axes. The axis is horizontal. The axis is with The horizontal direction perpendicular to the axis The axes are respectively with shaft and The vertical direction perpendicular to the axis. Detailed Implementation

[0060] The inventive concept of this invention is as follows:

[0061] The purpose of this invention is to construct an ultrasonic intelligent non-destructive testing method based on probabilistic convolutional neural network pose adjustment. During the scanning process, the uncertainty output by the probabilistic convolutional neural network is used as an online detection quality index. The scanning path and probe posture are dynamically adjusted according to the uncertainty, so as to realize local key rescanning and in-process correction, thereby reducing the risk of missed detection and improving detection efficiency and result reliability.

[0062] The core innovation of this invention lies in using the prediction uncertainty output by the deep learning model (probabilistic convolutional neural network) as an online feedback quantity in the detection process. This uncertainty is compared with a preset threshold, and the robot scanning strategy is dynamically adjusted based on the comparison result, thereby establishing a closed-loop detection mechanism based on uncertainty feedback. Unlike traditional detection methods, this invention allows the detection process to be adjusted in real time according to the reliability of the model's predictions.

[0063] The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment of this invention uses the uncertainty of a defect presence / absence classification task as the trigger for rescanning. When the classification prediction uncertainty exceeds a threshold, the reliability of the current detection result is deemed insufficient, triggering a focused scanning strategy. When the uncertainty is below the threshold, the standard scanning path continues to be executed.

[0064] The present invention relates to an ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment, which combines the pose adjustment of a phased array ultrasonic probe, including the phased array ultrasonic probe along the edge of the sample being tested. directional movement distance (i.e., the forward and backward movement distance), the phased array ultrasonic probe rotates around the sample being tested. Adjust the total rotation angle by rotating the axis (i.e., in-plane rotation angle), the phased array ultrasonic probe rotates around the sample being tested. Adjust the total rotation angle by rotating the axis (i.e. around (axis rotation angle) and the phased array ultrasonic probe rotating around the sample under test Adjust the total rotation angle by rotating the axis (i.e. around By using the rotation angle of the phased array ultrasonic probe (the angle of rotation of the axis), the coupling state between the phased array ultrasonic probe and the sample under test (i.e. the component under test) can be improved by multi-angle incident illumination, thereby improving the ultrasonic imaging quality and reducing the uncertainty of model prediction.

[0065] The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment of the present invention has an uncertainty reduction judgment and exit mechanism. If the prediction uncertainty is reduced to below a threshold after pose adjustment and repeated testing, the current test result is determined to be reliable, the key scanning area is exited, and the standard scanning path is resumed. If it is still higher than the threshold and the upper limit of repeated testing is reached, the coordinates of the testing area are recorded and the standard scan continues.

[0066] The ultrasonic intelligent nondestructive testing method based on pose adjustment of probabilistic convolutional neural network of the present invention uses probabilistic convolutional neural network to simultaneously output defect classification results, uncertainty and defect size information, so as to realize the integrated output of identification results and reliability assessment.

[0067] The present invention will now be described in detail with reference to the accompanying drawings.

[0068] The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment of the present invention is applicable to detection systems such as... Figure 3 As shown, it includes: a control and testing system 1 and a sample 2 to be tested; the control and testing system 1 is used to test the sample 2 according to a preset scanning path 17;

[0069] The control and testing system 1 includes: a host computer 11, a signal generator / receiver 12, a robotic arm controller 13, a robotic arm 14, a phased array ultrasonic probe 15, and a coupling agent 16. The host computer 11 is connected to both the signal generator / receiver 12 and the robotic arm controller 13; the signal generator / receiver 12 is connected to the phased array ultrasonic probe 15; the robotic arm controller 13 is connected to the robotic arm 14; the phased array ultrasonic probe 15 is located at the front end of the robotic arm 14; and the lower surface of the phased array ultrasonic probe 15 is provided with the coupling agent 16.

[0070] The host computer 11 is used to control the signal generator / receiver 12 to generate an excitation waveform and transmit it to the phased array ultrasonic probe 15; the host computer 11 is also used to control the robotic arm controller 13, thereby operating the robotic arm 14 to move.

[0071] The sample 2 under test has a sample defect 21 and a suspicious area 22. The sample defect 21 can be of various types, such as regular cracks, holes, polygonal defects, and irregular defects, such as corrosion areas or structural anomalies generated during processing; the defect is located inside the sample 2 under test.

[0072] The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment of the present invention, such as... Figure 7 As shown (only a summary of the steps is shown in the figure), the steps include:

[0073] Step 1: Initialize ultrasound detection parameters;

[0074] The parameters for generating the excitation waveform by the signal generator / receiver 12 are set by the host computer 11.

[0075] In a specific embodiment of the present invention, 100 Taking a negative square wave as an example, the center frequency is set to 5. The sampling frequency is 50. .

[0076] Step 2: Generate the standard scanning path for the robotic arm;

[0077] The host computer 11 generates a standard scanning path 17 covering the detection area.

[0078] In a specific embodiment of the present invention, a serpentine scanning path is taken as an example. The path coordinates are transformed to the robot arm base coordinate system by the robot arm controller 13, and the robot arm 14 is driven to move point by point according to the generated scanning path 17 to scan the sample 2 under test and perform the ultrasonic testing task.

[0079] Step 3: Ultrasonic signal preprocessing;

[0080] The raw ultrasonic echo signals acquired during the detection process are processed; the specific steps are as follows:

[0081] First, the raw ultrasonic echo signals collected during the detection process are bandpass filtered.

[0082] Then, the envelope signal of the processed echo signal is extracted by Hilbert transform, and the amplitude of the envelope signal is normalized.

[0083] Finally, the effective signal segment within the set sound path range is extracted for subsequent ultrasound imaging processing.

[0084] In a specific embodiment of the present invention, taking the full-matrix capture mode full-focusing imaging method for converting waveform signals into two-dimensional ultrasound images as an example, the converted full-focusing ultrasound image (dB) is as follows: Figure 1 As shown in the figure, defects, detection dead zones, and bottom waves can be seen.

[0085] Step 4: Ultrasonic imaging and probabilistic convolutional neural network to identify defects;

[0086] The ultrasound imaging results are input into a probabilistic convolutional neural network for defect identification processing, which yields the classification results of whether a defect exists, the prediction uncertainty of the defect, and the defect size information.

[0087] In a specific embodiment of the present invention, a probabilistic convolutional neural network based on random weight perturbation is used as an example to output the prediction variance or standard deviation as an uncertainty index. The structure diagram of this probabilistic convolutional neural network can be found in [reference needed]. Figure 2 including the input layer ( Figure 2 The model uses "input" to represent the input layer, a feature extraction module, a random weight perturbation module, a classification output branch, and a regression output branch. The input layer receives ultrasound images and inputs them to the feature extraction module. The feature extraction module includes three convolutional modules: convolutional module one, convolutional module two, and convolutional module three. Each convolutional module consists of a 3×3 convolution, batch normalization, an activation function, and a max-pooling layer. The output feature sizes of each convolutional module are H / 2×W / 2×32, H / 4×W / 4×64, and H / 8×W / 8×128, respectively. Here, H and W represent the height and width of the input ultrasound image, H / 2, H / 4, H / 8, and W / 2, W / 4, W / 8 represent the progressively downsampled spatial dimensions after max-pooling, and 32, 64, and 128 represent the number of feature channels in the corresponding convolutional module. The features extracted by each convolutional module are flattened and input into a fully connected layer to obtain a high-dimensional feature representation. The random weight perturbation module includes a Bayesian fully connected layer with weight probability distribution and a random weight perturbation layer. It is used to randomly perturb and sample the network weights to obtain a feature representation with uncertainty expression capabilities. Its input dimension is 128-dimensional, and its output dimension is 64-dimensional. The classification output branch is used to determine whether a defect exists, including a fully connected layer (64-dimensional). 2) and a classification function (normalized exponential function), combined with a random sampling deactivation mechanism, are used for multiple forward propagations to obtain the probability output of whether a defect exists or not. The mean, standard deviation, or variance is further calculated as the classification prediction uncertainty. The regression output branch is used to output defect size information and mainly includes a fully connected layer (64... 2) Used to predict parameters such as defect length, width, and area. Through the above structure, the probabilistic convolutional neural network can simultaneously output defect classification results, classification uncertainty, and defect size information, achieving integrated output of detection results and reliability assessment.

[0088] Step 5: Determine whether the defects output by the probabilistic convolutional neural network have classification results and their corresponding prediction uncertainties. The prediction uncertainty threshold is determined based on the quality of the training data and the model training effect. This invention uses the prediction standard deviation threshold (hereinafter referred to as the preset threshold). Let's take an example to illustrate. The specific steps include:

[0089] Step 5.1: Compare the prediction uncertainty with a preset threshold. A comparison is made when the prediction uncertainty is lower than or equal to a preset threshold. If the current detection result is deemed reliable, the robotic arm 14 continues to perform the detection task according to the standard scanning path 17. After the robotic arm 14 completes the global scan of the scanning path 17, it outputs the spatial coordinate information of the detection points determined to have defects (i.e., defect location coordinate information), the ultrasonic imaging image of the corresponding location (i.e., ultrasonic image data), and the defect size information predicted by the probabilistic convolutional neural network (i.e., coordinates of the suspicious area).

[0090] Step 5.2: Compare the prediction uncertainty with a preset threshold. A comparison is made when the prediction uncertainty exceeds a preset threshold. If it is determined that there is a high degree of recognition uncertainty at the current detection location, then the number of repeated detections in that detection area is further determined. Has the preset upper limit been reached?

[0091] When the upper limit of repeated detection is not reached, the phased array ultrasonic probe 15 is repeatedly adjusted in probe posture; during the probe posture adjustment process, ultrasonic imaging and recognition processing is performed again, and the corresponding prediction uncertainty is calculated.

[0092] When the maximum number of repeated inspections is reached, the spatial coordinates of the suspicious area 22 are recorded, and the robotic arm 14 continues to execute the standard scanning path 17. After the robotic arm 14 completes the global scan of the scanning path 17, it outputs the spatial coordinates of the inspection points determined to have defects, the ultrasonic imaging images of the corresponding locations, and the defect size information predicted by the probabilistic convolutional neural network. At the same time, it records the coordinates of the suspicious points that have reached the maximum number of repeated inspections and the corresponding uncertainty data, thus completing this inspection process.

[0093] Taking a preset upper limit of 5 as an example, if If the preset upper limit is not reached (<5), the current detection position is set as the key scanning position. A suspicious area 22 is formed according to the key scanning path generated by the host computer 11, and the robotic arm 14 is driven by the robotic arm controller 13 to execute the key scanning path to improve the coupling status of the coupling agent 16.

[0094] The focused scanning process involves locally intensified scanning near the original detection point, combined with end-effector attitude adjustments for targeted detection. During the focused scanning process, such as... Figure 4 As shown, the phased array ultrasonic probe 15 is adjusted back and forth along the scanning direction, and the moving distance is adjusted accordingly. ,Pick For example. And in Angular rotation adjustment in a plane (i.e., around) (axis rotation adjustment), total rotation angle is .like Figure 5 As shown, the phased array ultrasonic probe is rotated 15 times. The axis is rotated slightly to make fine adjustments, with a total rotation angle of [value missing]. .like Figure 6 As shown, the phased array ultrasonic probe is rotated 15 times. The axis is adjusted in angle, with a total rotation angle of [value missing]. Total rotation angle , , It depends on the accuracy of the robotic arm's 14-position control; here, we take an accuracy of 1° as an example.

[0095] This invention presents an ultrasonic intelligent nondestructive testing method based on pose adjustment using probabilistic convolutional neural networks. It constructs a closed-loop detection control mechanism based on deep learning prediction uncertainty. Unlike traditional detection methods that execute according to a fixed preset path, this invention uses the prediction uncertainty output by the deep learning model (probabilistic convolutional neural network) as online feedback information during the detection process, comparing it with a preset threshold and adjusting the robot's detection strategy in real time based on the comparison result. In this way, the detection system can dynamically optimize the scanning process based on the reliability of the current identification results, thereby improving the overall intelligence level and reliability of the detection.

[0096] In its implementation, this invention uses the prediction uncertainty corresponding to the defect classification task as the trigger for focused rescanning. When the uncertainty corresponding to the classification result exceeds a preset threshold... If the uncertainty is below a preset threshold, it indicates that the current detection location's identification reliability is insufficient. Based on this, the system triggers a focused scanning strategy, performing further rescanning and optimized detection on that area. If the current detection result is deemed to have high reliability, the robotic arm 14 continues to perform subsequent detection tasks along the standard scanning path 17. In this way, the present invention achieves adaptive rescanning decision based on classification uncertainty. For suspicious regions 22 identified as having high uncertainty, the present invention further proposes a multi-pose local scanning adjustment strategy. During focused scanning, local focused scanning is performed near the original detection position, such as... Figure 4 , 5 As shown in Figure 6, in conjunction with the probe attitude adjustment of the phased array ultrasonic probe 15, fine-tuning of the front and rear positions and multi-degree-of-freedom angle adjustment are performed on the phased array ultrasonic probe 15, including: the total in-plane rotation angle. , around Total rotation angle of the shaft and around Total rotation angle of the shaft By changing the incident position and incident angle of the phased array ultrasonic probe 15, the coupling state between the phased array ultrasonic probe 15 and the sample 2 under test can be effectively improved, thereby improving the quality of ultrasonic signal acquisition and imaging clarity, and thus reducing the uncertainty of model prediction.

[0097] This invention establishes a determination and exit mechanism based on uncertainty changes. Specifically, after local attitude adjustment and repeated detection, the system re-images and intelligently identifies the acquired results, and recalculates the corresponding uncertainty. When the prediction uncertainty decreases to a preset threshold... If the current detection result is deemed reliable, the system exits the key scanning area and resumes the standard scanning path; if the prediction uncertainty still exceeds the preset threshold after multiple repeated detections... If the preset limit for repeated detections has been reached, the spatial coordinates of the area are recorded, and subsequent standard scanning tasks continue. This mechanism ensures that the detection process can focus on optimizing difficult areas while reducing the impact of repeated local detections on overall detection efficiency. This invention uses a probabilistic convolutional neural network to achieve integrated output of detection results and uncertainty information. This network can not only output the classification results of the presence or absence of defects, but also simultaneously provide the uncertainty of the corresponding prediction results, and further output parameter information such as defect size. By unifying the identification results, size representation, and reliability assessment into the same model output framework, this invention achieves the organic integration of detection, judgment, and reliability evaluation, providing a direct basis for subsequent scanning strategy adjustments and suspicious area marking.

[0098] Regarding the uncertainty acquisition method, this invention uses a probabilistic convolutional neural network based on random weight perturbation as an example to output the prediction variance or standard deviation as an uncertainty index. In other specific embodiments, any neural network structure capable of outputting the quantitative result of model prediction reliability can be used as an alternative, such as a Monte Carlo-based random deactivation network, a multi-model ensemble network, a Bayesian neural network, or other network structures capable of outputting prediction distribution, prediction entropy, or confidence interval, all of which can be used to obtain uncertainty and participate in loop closure determination.

[0099] Regarding the attitude adjustment strategy, this invention utilizes the forward and backward movement and rotation of the phased array ultrasonic probe 15. axis, Fine-tuning by rotating the axis at a small angle improves the coupling state. In other specific embodiments, any posture or contact condition adjustment method that can improve the quality of ultrasonic imaging data can be used, such as: adjusting the downward pressure of the phased array ultrasonic probe 15, changing the gap distance between the phased array ultrasonic probe 15 and the sample 2 under test, changing the incident angle, or using force control or impedance control to stabilize the coupling state. As long as it can improve the imaging quality and reduce the uncertainty of model prediction, it is an alternative implementation of the present invention.

[0100] Regarding the judgment criteria, this invention compares the prediction variance or prediction standard deviation with a preset threshold. In other specific embodiments, other model prediction reliability indicators can also be used as the judgment criteria, such as prediction entropy, confidence interval width, distance between output probability and decision boundary, or other statistics reflecting the uncertainty of the model, and compared with the set threshold to trigger a key scan or exit mechanism.

[0101] Regarding the detection path, this invention uses a serpentine scanning path 17 as an example. In other specific embodiments, the standard detection path can be a serpentine scan, a grid scan, a circular arc scan, or other comprehensive scanning methods. Serpentine scanning refers to the phased array ultrasonic probe 15 completing a straight-line scan along a first direction, then moving a preset step distance along a second direction, and then performing the next straight-line scan in the opposite direction, repeating this process to form a continuous zigzag scanning path to cover the entire detection area, such as... Figure 3 As shown. Grid scanning refers to dividing the area to be inspected into several regular grid points. The robotic arm 14 moves sequentially to each grid point according to a preset row and column order, and completes ultrasonic signal acquisition at each grid point, thereby achieving discrete point-matrix coverage of the inspection area. Circular arc scanning refers to the robotic arm 14 driving the phased array ultrasonic probe 15 to move along an arc trajectory based on a preset center point or structural contour, performing ultrasonic detection at multiple sampling positions along the arc path. It is suitable for detecting circular, arc-shaped, or curved structural areas.

[0102] In terms of signal processing and ultrasound imaging, this invention uses filtering, envelope extraction, normalization processing, and a full-matrix capture mode for full-focus imaging as examples. In other specific embodiments, signal processing can employ other noise reduction methods, time-frequency analysis methods, or normalization methods; imaging methods can employ phased array sector scanning imaging, linear scanning imaging, plane wave imaging, synthetic aperture focusing technology, or other ultrasound imaging methods that can generate two-dimensional or three-dimensional imaging results. As long as imaging data that can be generated for deep learning models to recognize and the corresponding uncertainty can be output, these are all alternative implementations of this invention.

[0103] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An ultrasonic intelligent nondestructive testing method based on pose adjustment using a probabilistic convolutional neural network, wherein the applicable testing system includes: Control and testing system (1), sample under test (2); the control and testing system (1) is used to test the sample under test (2) according to the preset scanning path (17); The control and testing system (1) includes: a host computer (11), a signal generator / receiver (12), a robotic arm controller (13), a robotic arm (14), a phased array ultrasonic probe (15), and a coupling agent (16); the host computer (11) is connected to the signal generator / receiver (12) and the robotic arm controller (13) respectively; the signal generator / receiver (12) is connected to the phased array ultrasonic probe (15); the robotic arm controller (13) is connected to the robotic arm (14); the phased array ultrasonic probe (15) is located at the front end of the robotic arm (14); the lower surface of the phased array ultrasonic probe (15) is provided with a coupling agent (16); The host computer (11) is used to control the signal generator / receiver (12) to generate an excitation waveform and transmit it to the phased array ultrasonic probe (15); the host computer (11) is also used to control the robotic arm controller (13), thereby operating the robotic arm (14) to move; Its features are, The method includes the following steps: Step 1: Initialize ultrasound detection parameters; The parameters for generating the excitation waveform by the signal generator / receiver (12) are set by the host computer (11); Step 2: Generate the standard scanning path for the robotic arm (14); The host computer (11) generates a standard scanning path (17) covering the detection area. Step 3: Ultrasonic signal preprocessing; The raw ultrasonic echo signals collected during the detection process are processed; Step 4: Ultrasonic imaging and probabilistic convolutional neural network to identify defects; The ultrasound imaging results are input into a probabilistic convolutional neural network for defect identification processing to obtain the classification results of whether the defect exists, the prediction uncertainty of the defect, and the defect size information. Step 5: Determine whether the defects output by the probabilistic convolutional neural network have classification results and their corresponding prediction uncertainties; Step 5 specifically includes: Step 5.1: Compare the prediction uncertainty with the preset threshold. When the prediction uncertainty is lower than or equal to the preset threshold, the current detection result is determined to be reliable. The robotic arm (14) continues to perform the detection task according to the standard scanning path (17). After the robotic arm (14) completes the global scan of the scanning path (17), it outputs the spatial coordinate information of the detection point that is determined to have a defect, the ultrasonic imaging image of the corresponding position, and the defect size information predicted by the probabilistic convolutional neural network. Step 5.2: Compare the prediction uncertainty with the preset threshold. When the prediction uncertainty is higher than the preset threshold, it is determined that there is recognition uncertainty at the current detection location. Then, it is further determined whether the number of repeated detections in the detection area has reached the preset upper limit. When the upper limit of repeated detection is not reached, the phased array ultrasonic probe (15) is repeatedly adjusted in probe posture; during the probe posture adjustment process, ultrasonic imaging and recognition processing is performed again, and the corresponding prediction uncertainty is calculated. When the upper limit of repeated detection is reached, the spatial coordinate information of the detection area is recorded, and the robotic arm (14) continues to execute the standard scanning path (17). After the robotic arm (14) completes the global scan of the scanning path (17), it outputs the spatial coordinate information of the detection point that is determined to have a defect, the ultrasonic imaging image of the corresponding position, and the defect size information predicted by the probabilistic convolutional neural network. At the same time, it records the coordinates of the suspicious point that has reached the upper limit of repeated detection and the corresponding uncertainty data, thus completing the detection process.

2. The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment according to claim 1, characterized in that, Step 3 specifically includes: First, the raw ultrasonic echo signals collected during the detection process are bandpass filtered. Then, the envelope signal of the processed echo signal is extracted by Hilbert transform, and the amplitude of the envelope signal is normalized. Finally, the effective signal segment within the set sound path range is extracted for subsequent ultrasound imaging processing.

3. The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment according to claim 1, characterized in that, In step 5.2, the probe attitude adjustment includes fine-tuning of the front and rear positions and multi-degree-of-freedom angle adjustment.

4. The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment according to claim 3, characterized in that, The fine-tuning of the front and rear positions specifically refers to the forward and backward movement distance of the phased array ultrasonic probe (15). ; Multi-degree-of-freedom angle adjustment specifically includes: Phased array ultrasonic probe (15) around Total rotation angle of axis rotation ; Phased array ultrasonic probe (15) around Total rotation angle of axis rotation ; Phased array ultrasonic probe (15) around Total rotation angle of axis rotation ; in, The axis is horizontal. The axis is with The horizontal direction perpendicular to the axis The axes are respectively with shaft and The vertical direction perpendicular to the axis.

5. The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment according to any one of claims 1-4, characterized in that, In step 4, the prediction uncertainty corresponding to the defect is: The prediction variance or standard deviation of the output of a probabilistic convolutional neural network based on a random weight perturbation method; or The predicted distribution output by the random deactivation Monte Carlo sampling method, along with its corresponding prediction variance, prediction entropy, or confidence interval; or Based on the predicted distribution of the multi-model ensemble network output, and its corresponding prediction variance, prediction entropy, or confidence interval; or The predicted distribution based on the output of a Bayesian neural network, and its corresponding prediction variance, prediction entropy, or confidence interval.

6. The ultrasonic intelligent nondestructive testing method based on probabilistic convolutional neural network pose adjustment according to any one of claims 1-4, characterized in that, The scanning path (17) is: serpentine scan, grid scan, or circular scan; The serpentine scanning is as follows: After the phased array ultrasonic probe (15) completes a straight line scan along the first direction, it moves a preset step distance along the second direction and then performs the next straight line scan in the opposite direction. This process is repeated to form a continuous zigzag scanning path to cover the entire detection area. The grid scanning is as follows: the area to be detected is divided into several regular grid points, and the robotic arm (14) moves to each grid point in the preset row and column order, and completes the ultrasonic signal acquisition at each grid point; The circular arc scanning is as follows: the robotic arm (14) drives the phased array ultrasonic probe (15) to move along the circular arc trajectory based on the preset center point or structural contour, and performs ultrasonic detection at multiple sampling positions on the circular arc path.

Images