Intelligent detection method for internal defects of lead seal based on full focusing and machine learning
By combining full focusing with machine learning, using ultrasonic phased array equipment for multi-angle scanning and data processing, and combining convolutional neural networks for image classification, the problems of high efficiency and accuracy in detecting internal defects in cable seals were solved, achieving high reliability and high precision detection results.
Patent Information
- Application Number
- CN202510961415.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-07-13
- Filing Date
- 2025-07-11
- Publication Date
- 2025-09-26
AI Technical Summary
Existing cable sealing lead defect detection methods have problems such as high detection condition requirements, complex operation, long detection time, unclear imaging and low detection accuracy, making it difficult to achieve efficient and accurate internal defect detection.
A method combining full focusing and machine learning is adopted, multi-angle scanning is performed using ultrasonic phased array equipment, data is processed through full matrix capture and full focusing technology, image classification is performed using machine learning models, and internal defects of lead seals are detected using convolutional neural networks.
It achieves high reliability and high precision detection of internal defects in lead seals, can adapt to rich ultrasonic data, has high flaw detection performance, and is consistent with the results of human inspectors.
Smart Images
Figure CN120703238A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of nondestructive testing and machine learning, and in particular to an intelligent detection method for internal defects of lead seals based on full focusing and machine learning. Background Art
[0002] Cables are the carriers of electricity and a vital component of the power grid. Lead seals are a crucial accessory in high-voltage cables, sealing cable joints to prevent leakage and moisture intrusion. Made primarily of a lead-tin alloy, these seals are often exposed to outdoor environments for extended periods, making them susceptible to deformation and cracking. Defects in these seals can render the joints ineffective. Defects in high-voltage cable lead seals can occur in a variety of ways, including cracking, deformation, surface peeling, scratches, internal lamination, and pinholes. These defects can occur either within the seal or on its surface, and can easily lead to poor electrical connections, reduced insulation strength, cable leakage, breakdown, and even fires. Therefore, detecting lead seal defects is of great practical significance.
[0003] Various techniques, such as partial discharge detection, infrared detection, electric field distribution detection, and radiographic detection, have been widely used to diagnose defects in power cable accessories. However, these methods suffer from issues such as demanding testing conditions, complex operation, time-consuming testing, unclear imaging, and low detection accuracy. Currently, ultrasonic nondestructive testing is the primary method for detecting internal defects in lead seals. Modern ultrasonic inspection utilizes the richer datasets provided by fully focused equipment. A typical inspection may include dozens of channels with varying refraction angles, acquired at high speeds. These rich datasets enable highly reliable and effective inspections in complex situations. Convolutional neural networks have recently demonstrated the ability to detect defects with human-level accuracy in ultrasonic signals at the b-scan level. To achieve human-level accuracy in automated defect detection for critical applications, these neural networks need to be developed to leverage today's rich fully focused datasets.
[0004] Machine learning (ML) models have proven their effectiveness in various image recognition tasks, so ML models can be used to remove much of the repetitiveness from NDT data analysis, even in noisy and complex situations. Since most inspection data is typically defect-free, ML models can be used to find areas with defective seals. After the machine learning system identifies the locations of possible defect signs on cable termination seals, inspectors can verify the results and apply expert judgment in defect assessment. The ability to leverage increasing amounts of inspection data allows for earlier seal defect detection and more efficient monitoring of power systems and defects. Summary of the Invention
[0005] In order to solve the technical problems existing in the background technology, the present invention proposes an intelligent detection method for internal defects of lead seals based on full focusing and machine learning.
[0006] The present invention proposes an intelligent detection method for internal defects of lead seals based on full focusing and machine learning, which includes the following steps:
[0007] S1. Using ultrasonic phased array equipment, cable terminal seals are used to scan at multiple angles, full matrix capture is used to collect data, and full focusing technology is used to process the data.
[0008] S2, all lead sealing ultrasound datasets divide available defects and canvases into training set, validation set and test set;
[0009] S3. Prepare physical samples for ML training to train machine learning models and process lead sealing full focus data to achieve image classification;
[0010] S4. Produce physical samples for ML verification to verify the performance of the machine learning model and automatically detect internal defects in lead seals;
[0011] S5. The model is tested with two internal and external defects of similar seal geometry used in training and compared with the results of human inspectors.
[0012] Further technical solutions; all the lead sealing ultrasonic datasets in S2 will be divided into available defects and canvases into training sets, validation sets and test sets including:
[0013] All datasets contain 50% images with only internal defects and 50% images with only external defects;
[0014] The training set data is divided into 50% images with only internal defects and 50% images with only external defects. The training set data is then preprocessed to remove a large amount of redundant data contained in the multi-angle channels to form the training set data.
[0015] The validation set data is divided into 50% images with only internal defects and 50% images with only external defects. The validation set data is then preprocessed and enhanced using virtual defects to form the validation set data.
[0016] After excluding internal defects or external defects in the training / validation set, the remaining images are divided into 50% images with only internal defects and 50% images with only external defects as the test set data.
[0017] A further technical solution is to pre-process the training data as described in S2 to remove a large amount of redundant data contained in the multi-angle channel, including:
[0018] When one element in the probe array is excited to generate an ultrasonic wave, all elements in the array simultaneously record the received ultrasonic echo. This process is repeated sequentially across all elements in the array, with each element acting as a transmitter in turn, and all elements, including the transmitter, acting as receivers. Ultimately, the system acquires and stores complete ultrasonic waveform data (A-scan sequence) for every possible transmitter-receiver element combination in the probe array.
[0019] Each independent transmitter-receiver chip combination channel (i.e., a complete transmitter and a unique combination of data consisting of a specific receiver chip) is treated as an independent unit for processing.
[0020] First, take the absolute value of the original signal of each channel to obtain the envelope information of the signal.
[0021] The core full-focusing technology process is as follows: For each target pixel on the final imaging plane:
[0022] First, calculate the spatial distance from the pixel point to the current transmitting chip, and then calculate the spatial distance from the pixel point to the current receiving chip.
[0023] Based on the above two distances and the speed of ultrasound in the cable seal, the total propagation time for the ultrasound to propagate from the transmitting array element to the pixel point and then be reflected back to the receiving chip is calculated.
[0024] Based on the calculated total propagation time, the amplitude at the corresponding time point is extracted from the A-scan signal envelope data corresponding to the currently processed transmit-receive combined channel (implemented by interpolation).
[0025] After traversing all relevant channels, the signal amplitudes of the pixel at the corresponding time point are accumulated and summed. This accumulated sum eventually becomes the brightness value of the pixel.
[0026] Obtain the above TFM imaging, consider each frame separately and be rectified, that is, take the absolute value of the signal; 1 / 2λ is half of the positive definite matrix eigenvalue of each frame image window, match half of the positive definite matrix eigenvalue of each frame image window to perform maximum pooling on a single channel, and then store the data in a compressed binary file to facilitate file transfer and accelerate learning.
[0027] Further technical solutions; A common technique for processing limited training data in machine learning is to use data enhancement, including: the use of so-called virtual defects described in S2 can obtain more complex enhancement solutions.
[0028] A further technical solution is to prepare physical samples for ML training based on S3 for training machine learning models, including:
[0029] Use original UT data as physical samples for ML training, i.e. scanned UT data of lead-sealed plates with peeling and scratches on the surface and no internal defects;
[0030] In addition, a single lead seal with internal defects is scanned to obtain data with internal defects. Internal defects do not contain any external peeling or scratches, thus providing signals without external defects, which can be enhanced as necessary to enrich the data set.
[0031] The current setup allows extracting a clean internal defect signal from a sample without external defects and embedding it into a signal with only external defects;
[0032] The above data must be preprocessed before being used to train the machine learning model.
[0033] Further technical solutions: Physical samples based on ML training described in S3 are used to train machine learning models, and lead sealing ultrasound full focus data is processed to achieve image classification, including:
[0034] The DCNN used for image classification tasks can be considered as a YOLO network to train the machine learning model, using small convolutional filter output vectorization to achieve internal defect image classification of lead seals;
[0035] The final model is trained with all available non-test internal and external defects.
[0036] Further technical solutions: Based on the physical samples for ML verification described in S4, to verify the performance of the machine learning model and automatically detect internal defects of the lead seal, including:
[0037] Using samples opposite to those used in training the network, in the validation set, the real defects include both internal and external defects;
[0038] Create a completely independent sample set as a physical sample for verification. After preprocessing and enhancing the data in the verification set, input it into the machine learning model to verify the performance of the machine learning model and automatically detect internal defects in the lead seal.
[0039] Further technical solutions: Based on the model described in S4 and the two internal and external defects with similar sealing geometry used in training, the following are tested:
[0040] Initially, ML model performance was measured using a test dataset, extracted from a dataset containing all available defect sizes. Approximately 50% of the scans had internal defects and 50% did not, to measure the true performance of the model and observe possible overfitting. Results were evaluated based on false call rate and probability of detection (POD) metrics.
[0041] Beneficial effects of the present invention:
[0042] 1. Using ultrasonic testing equipment, dozens of channels with different refraction angles are used to inspect the lead seals at the cable terminals. Full-matrix capture is used to collect data and the full-focus algorithm is used to process the data. All available defects and canvases of the lead seal ultrasonic dataset are divided into training sets, validation sets, and test sets. Physical samples for ML training are prepared to train the machine learning model, and the full-focus data of the lead seal ultrasonic is processed to achieve image classification. Physical samples for ML verification are prepared to verify the performance of the machine learning model and automatically detect internal defects in the lead seal. The model is tested with two internal and external defects of similar lead seal geometry used in training, and the results are compared with those of human inspectors.
[0043] 2. The machine learning model can perform highly reliable internal defect detection on typical multi-channel full-focus data on lead seals. The machine learning model can adapt to rich ultrasonic data and has high flaw detection performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 Flowchart of the detection method of the present invention;
[0045] Figure 2 This is a schematic diagram of the receiving and transmitting status of the array elements when the probe is scanning during full matrix capture of the present invention;
[0046] Figure 3 This is a schematic diagram of the full matrix data captured by the present invention;
[0047] Figure 4 Schematic diagram of the full focusing algorithm of the present invention;
[0048] Figure 5 This is a flow chart of ultrasonic data preprocessing of the present invention;
[0049] Figure 6 Schematic diagram of the present invention using a deep convolutional neural network to estimate ultrasonic scanning defects;
[0050] Figure 7 Schematic diagram of verification data evaluation of the present invention. DETAILED DESCRIPTION
[0051] Reference Figure 1-7 The present invention proposes an intelligent detection method for internal defects of lead seals based on full focusing and machine learning, which is implemented according to the following steps:
[0052] S1. Using ultrasonic phased array equipment, scanning with cable terminal seals, full matrix capture and data collection, and full focusing technology to process the data;
[0053] S2, all lead sealing ultrasound datasets divide available defects and canvases into training set, validation set and test set;
[0054] S3. Create physical samples for ML training to train the machine learning model and process lead sealing ultrasound data for image classification.
[0055] S4. Produce physical samples for ML verification to verify the performance of the machine learning model and automatically detect internal defects in lead seals;
[0056] S5. The model is tested with two internal and external defects of similar seal geometry used in training and compared with the results of human inspectors.
[0057] Using ultrasonic phased array equipment, cable terminal seals are used to scan at multiple angles, full matrix capture is used to collect data, and full focusing technology is used to process the data. The specific process includes:
[0058] The cable seal is scanned at multiple angles by the probe, the echo data of the cable seal is captured in a full matrix, and the obtained ultrasonic data is processed using a full focusing algorithm to obtain defect images of the cable seal.
[0059] Specifically, the ultrasonic data acquisition process is as follows: First, a 64-element probe with a center frequency of 5MHz, an element spacing of 0.6mm, and a total aperture of 38.4mm is used to collect data. The probe is mounted on a wedge made of a lead alloy with a sound velocity of 2300m / s to ensure that the sound velocity matches that of the lead sealing material. Before starting the test, sound velocity calibration must be performed: the sound wave propagation time is measured on a lead-tin alloy calibration block of known thickness, and the sound velocity parameters of the equipment are adjusted so that the measured thickness error is less than 0.1mm. The final sound velocity is determined to be 2300±10m / s. At the same time, a high-viscosity ultrasonic coupling agent is used and evenly applied to the test area to achieve good acoustic coupling.
[0060] Data acquisition utilizes full-matrix capture mode: the probe linearly scans the lead seal surface with a 0.5mm step, performing 64×64 transmit-receive cycles at each scan position. Specifically, when array element i transmits a periodic pulse, the remaining 63 elements receive it synchronously. The received signal is recorded at a 50MHz sampling rate for 30μs (1500 sampling points), covering a depth of 30mm.
[0061] The single point data volume is determined by V data =N tx ×N rx ×N sample ×2;
[0062] Calculated to be 12.3MB (Ntx=64, Nrx=64, Nsample=1500, 16-bit sampling;
[0063] Create a 300×300 pixel grid in the imaging area (physical size 30mm×30mm, pixel resolution 0.1mm)
[0064] The grayscale value I(x,z) of each pixel (x,z) is calculated using the following formula:
[0065]
[0066] Where: A-scan signal excited by transmitting element i and recorded by receiving element j;
[0067] v = 2300 m / s is the speed of ultrasound in lead, fs = 50 MHz is the sampling rate, and round() is the rounding function.
[0068] Thus, a fully focused image dataset of each ultrasound is obtained.
[0069] For each frame of the above TFM-processed image, each frame is considered separately and rectified, that is, the absolute value of the signal is taken;
[0070] 1 / 2λ is half of the positive definite matrix eigenvalue of each frame image window, matching half of the positive definite matrix eigenvalue of each frame image window to perform maximum pooling on a single channel, and then storing the data in a compressed binary file to facilitate file transfer and accelerate learning
[0071] The framework performs max pooling with a window size of 0.5λ. This has the effect of making the envelope of the data computationally efficient. The data size is reduced from 48×1020 to 48×34=1632 samples.
[0072] The data is then stored in compressed binary files to facilitate file transfer and accelerate learning. During training, the data is decompressed and converted from the original 16-bit integers to 32-bit floating point numbers and scaled to 0...2.0 (most data is in the range of 0...1.0);
[0073] A common technique for dealing with limited training data in machine learning is to use data augmentation, including: More sophisticated augmentation schemes can be obtained using so-called virtual defects as described in S2;
[0074] Physical samples for making ML training based on S3 are used to train machine learning models, including:
[0075] Use original UT data as physical samples for ML training, i.e. scanned UT data of lead-sealed plates with peeling and scratches on the surface and no internal defects;
[0076] In addition, a single lead seal with internal defects is scanned to obtain data with internal defects. Internal defects do not contain any external peeling or scratches, thus providing signals without external defects, which can be enhanced as necessary to enrich the data set.
[0077] The current setup allows extracting a clean internal defect signal from a sample without external defects and embedding it into a signal with only external defects;
[0078] The above data must be preprocessed before being used to train the machine learning model.
[0079] The physical samples based on the ML training described in S3 are used to train the machine learning model to process the lead-sealed ultrasound full-focus data for image classification, including:
[0080] The DCNN used for image classification tasks can be considered as a YOLO network to train the machine learning model, using small convolutional filter output vectorization to achieve internal defect image classification of lead seals;
[0081] The final model is trained with all available non-test internal and external defects.
[0082] S4-based physical samples are used to verify the performance of machine learning models and automatically detect internal defects in lead seals, including:
[0083] Using samples opposite to those used in training the network, in the validation set, the real defects include both internal and external defects;
[0084] Create a completely independent sample set as a physical sample for verification. After preprocessing and enhancing the data in the verification set, input it into the machine learning model to verify the performance of the machine learning model and automatically detect internal defects in the lead seal.
[0085] Based on S4, the model was tested with two internal and external defects of similar sealing geometry used in training, including:
[0086] Initially, ML model performance was measured using a test dataset, extracted from a dataset containing all available defect sizes. Approximately 50% of the scans had internal defects and 50% did not, to measure the true performance of the model and observe possible overfitting. Results were evaluated based on false call rate and probability of detection (POD) metrics.
[0087] More complex enhancement schemes can be obtained using so-called virtual defects, including:
[0088] 1. Each seal with only external defects is used as a perfect canvas. The full A-scan is cropped to the region of interest of the seal's external defects to minimize excess data:
[0089] 2. For each canvas, generate a set of 500,000 samples (divided into 50 batches of 10,000 samples each):
[0090] A number is randomly drawn to select a sample that is perfect or has internal defects.
[0091] If the sample is designated as free of internal defects:
[0092] A window of 48 A-movies is randomly selected and added to the result data.
[0093] 3. Each a-scan is moved by a random walk offset to simulate possible probe jitter during the scan, further enhancing the data.
[0094] If a sample is designated as internally defective:
[0095] A defect is randomly picked from the population and embedded into a random position in the file.
[0096] The defect amplitude decreases with the random factor in the range of 0.5 to 1.0.
[0097] 4. Each a-scan is moved by random walk offset to simulate possible probe jitter during the scan, thereby enhancing the defect.
[0098] After embedding, 48 scanning windows are randomly selected so that the defect is completely contained within the window.
[0099] The data was further enhanced by shifting each a-scan with a random walk offset, simulating possible probe jitter during the scan.
[0100] DCNN for image classification tasks can be thought of as a YOLO network to train machine learning models, including:
[0101] The DCNN architecture used is similar to the VGG16 network with three convolutional blocks. Each block consists of two consecutive convolutional layers with rectified linear unit (ReLU) activations. This is followed by a batch normalization (BN) layer that normalizes the input distribution to the next block, increasing the robustness of the network by reducing internal covariate shift. The convolutional blocks are followed by vectorization and densely connected layers with ReLU activations, whose units correspond to the digital filters of the last convolution. Finally, the weights of the densely connected layers converge to a single classification unit with a sigmoid activation, indicating the presence or absence of internal features. The loss function applied is binary cross entropy. During backpropagation, the adaptive moment estimation (ADAM) method is used to calculate the new weights. The TensorFlow library is used to preprocess and filter the data stream, and the Keras high-level API is used to build the DCNN.
[0102] like Figure 7As shown, to verify the performance of the machine learning model, including:
[0103] Data from a separate sample of seals with internal defects were run through the trained machine learning model as follows: Each file was segmented into a set of individually evaluated data frames corresponding to the selected training model input data size (96 rows). Frames were interpreted by moving windows of the aforementioned size across the data with 50% overlap, i.e., the first frame contained rows 1-96, the second contained rows 49-144, and so on. For each frame, all 31 channels were evaluated separately. If any (even one) of these frames was designated as having an internal defect, the frame location was considered to have an internal defect. If a frame containing an internal defect was identified as having an internal defect, this was considered a true hit; if it was identified as not having an internal defect, it was considered a miss. If a frame was identified as having an internal defect but did not contain an internal defect, this was considered a false call. The data did not contain any cases where an internal defect would fall partially on a frame. Overall, the data segmented this way contained 32 separate data frames, with 11 hit / miss opportunities for the near end, 11 hit / miss opportunities for the far end, and 10 false calls.
[0104] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.
Claims
1. An intelligent detection method for internal defects of lead seals based on full focusing and machine learning, characterized in that: include: S1. Use ultrasonic phased array equipment to perform multi-angle scanning on the cable terminal lead seal, collect data through full matrix capture, and process the data using full focusing technology; S2. Divide the lead sealing ultrasound dataset into a training set, a validation set, and a test set; S3, create physical samples to train machine learning models and process full-focus data for image classification; S4. Create physical samples for verification, verify model performance, and automatically detect internal defects; S5. Test the model and compare the results with those of human inspectors.
2. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 1 is characterized in that: The data set in S2 is divided as follows: The dataset contains 50% images with only internal defects and 50% images with only external defects; The training set, validation set, and test set are all divided into 50% internal defect images and 50% external defect images.
3. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 2 is characterized in that: The preprocessing of the training set data in S2 includes: Take the absolute value of the transmit-receive chip combined channel data captured by the full matrix to obtain the signal envelope; Calculate the brightness value of the target pixel based on the full focusing technology and generate TFM imaging; The TFM images are processed with max pooling and stored as compressed binary files.
4. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 1 is characterized in that: The data enhancement in S2 adopts virtual defect technology, including: The external defect area of the lead seal without internal defects is used as a canvas and the randomly selected internal defect signals are embedded; Perform a random walk offset on the A-scan data to simulate probe jitter.
5. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 1 is characterized in that: The methods for producing ML training physical samples in S3 include: Collect UT data of lead seals with peeling and scratches on the surface but no internal defects; Collect lead-sealed UT data containing only internal defects and no external defects, and embed the internal defect signals into the external defect signals.
6. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 1 or 5, characterized in that: The S3 uses the YOLO network as a deep learning model, outputs vectorized results through a small convolution filter, and realizes the classification of internal defect images of lead seals.
7. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 1 is characterized in that: The verification physical samples in S4 are samples independent of the training set, and contain real internal defects and external defects. During verification, the preprocessed and enhanced data are input into the model.
8. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 1 is characterized in that: The test data set in S5 contains 50% scans with internal defects and 50% scans without internal defects. The model performance is evaluated based on the false call rate and the detection probability.
9. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 1 is characterized in that: The ultrasonic phased array equipment parameters in S1 are: 64-element probe, center frequency 5 MHz, element spacing 0.6 mm, wedge sound velocity 2300 m / s, scanning step 0.5 mm, sampling rate 50 MHz.
10. The intelligent detection method for internal defects of lead seals based on total focusing and machine learning according to claim 3 is characterized in that: The process of processing data with the full focusing technology is as follows: Calculate the spatial distance from the target pixel to the transmitting chip and the receiving chip, and determine the ultrasonic propagation time based on the sound speed; The amplitude at the corresponding time point is extracted from the A-scan signal envelope and accumulated to obtain the pixel brightness value to generate the TFM image.