CN118731190A - An automatic detection method for internal defects of lead seals using multi-channel phased array ultrasonic data based on machine learning - Google Patents

Abstract

The invention provides a machine learning-based multichannel phased array ultrasonic data lead sealing internal defect automatic detection method, which comprises the following steps: using phased array equipment, checking cable terminal lead sealing by using tens of channels with different refraction angles, and collecting data by using full matrix capturing and related technologies; all lead-sealed ultrasonic data sets divide available defects and canvases into a training set, a verification set and a test set; making a physical sample for ML training, training a machine learning model, and processing lead-sealed ultrasonic phased array data to realize image classification; manufacturing a physical sample for verifying ML, verifying the performance of a machine learning model, and automatically detecting the lead sealing internal defect; the model was tested against both the inner and outer defects of similar lead sealing geometry used in training and compared to the results of a human inspector.

Description

An automatic detection method for internal defects of lead seals using multi-channel phased array ultrasonic data based on machine learning

Technical Field

The present invention belongs to the technical field of machine learning, and in particular relates to a method for automatically detecting internal defects of lead seals using multi-channel phased array ultrasonic data based on machine learning.

Background Art

Cable is the carrier of power transportation and an important part of the power grid. Lead seal is one of the important accessories in high-voltage cable. Its main function is to seal the cable joint to prevent leakage and moisture intrusion. Lead seal is mainly made of lead-tin alloy, most of which are in outdoor environment for a long time. Lead seal accessories are also prone to deformation, cracking and other problems. Once the accessory has defects, the joint will lose its protective function. There are various types of defects in high-voltage cable lead seal accessories, such as cracking, deformation, surface peeling, scratches, internal lamination, sand holes, etc. These defects exist either inside the lead seal or on the surface of the lead seal accessories, which can easily lead to poor electrical connection, reduced insulation strength, cable leakage, breakdown, cable fire and other problems. Therefore, detecting lead seal defects has important practical significance.

A variety of technologies such as partial discharge detection, infrared detection, electric field distribution detection, and X-ray detection have been widely used in the diagnosis of defects in power cable accessories. These methods have problems such as high requirements for detection conditions, complex operation, long detection time, unclear imaging, and low detection accuracy. At present, ultrasonic non-destructive testing is the main method for detecting internal defects of lead seals. Modern ultrasonic inspection utilizes richer data sets provided by phased array equipment. A typical inspection may include dozens of channels with different refraction angles, which are acquired at high speed. These rich data sets allow highly reliable and effective inspections in complex situations. Convolutional neural networks have recently shown the ability to detect defects with human-level accuracy in ultrasonic signals at the b-scan level. In order to achieve human-level accuracy in automatic defect detection for critical applications, these neural networks need to be developed to take advantage of today's rich phased array data sets.

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 of NDT data analysis, even in noisy and complex situations. Since most inspection data is usually defect-free, ML models can be used to find areas where the seals are defective. After the location of possible defect signs in the cable terminal seals is identified by the machine learning system, the inspector can verify the results and apply expert judgment in the defect assessment. The ability to utilize more and more inspection data allows for seal defect detection at an early stage, as well as more effective monitoring of power systems and defects.

Summary of the invention

In order to solve the shortcomings of the prior art, the present invention provides a multi-channel phased array ultrasonic data lead seal internal defect automatic detection method based on machine learning to solve the technical problems existing in the background;

To achieve the above objectives, the present invention is implemented through the following technical solutions:

A method for automatic detection of internal defects of lead seals using multi-channel phased array ultrasonic data based on machine learning, the method comprising the following steps:

S1; Using phased array equipment, dozens of channels with different refraction angles are used to inspect the cable terminal seals, using full matrix capture and correlation techniques to collect data;

S2; All lead sealing ultrasonic datasets divide the available defects and canvases into training set, validation set and test set;

S3; Prepare physical samples for ML training to train machine learning models and process lead sealing ultrasonic phased array data to achieve image classification;

S4; Make physical samples for ML verification to verify the performance of the machine learning model and automatically detect internal defects of lead seals;

S5;The model was tested with two internal and external defects of similar seal geometry used in training and compared with the results of human inspectors.

Further technical solution: All lead sealing ultrasonic data sets in S2 divide the available defects and canvases into training sets, verification sets and test sets including:

All datasets contain 50% images with only internal defects and 50% images with only external defects;

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.

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 the virtual defect method to form the validation set data.

After excluding the 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.

A further technical solution is to pre-process the training data described in S2 to remove a large amount of redundant data contained in the multi-angle channel, including:

The phased array probe scans at an angle of 40°-70° with a step size of 1°. It scans once every one degree and gets waveform data at 31 angles. Each of the 31 complete waveform channels is considered separately and each frame is rectified, that is, the absolute value of the signal is taken; 1/2λ is half of the eigenvalue of the positive definite matrix of each frame image window, and half of the eigenvalue of the positive definite matrix of each frame image window is matched to perform maximum pooling on a single channel. Then the data is stored in a compressed binary file to facilitate file transfer and accelerate learning.

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 schemes.

A further technical solution; based on the physical sample for ML training described in S3, for training the machine learning model, includes:

Use the original UT data as the physical sample for ML training, that is, scan the UT data of the lead-sealed plate sample with peeling and scratches on the surface and no internal defects;

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.

The current setup allows to extract a clean internal defect signal from a sample without external defects and embed it into the signal with only external defects;

All the above data must be preprocessed before being used to train the machine learning model.

Further technical solutions: Physical samples based on ML training described in S3 are used to train machine learning models, and lead sealing ultrasonic phased array data are processed to achieve image classification, including:

The DCNN 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 lead seal internal defect image classification;

The final model is trained with all available non-test internal and external defects.

Further technical solutions: Based on the physical samples for verifying ML described in S4, to verify the performance of the machine learning model, automatically detect the internal defects of the lead seal, including:

Using samples opposite to those used in training the network, in the validation set, the real defects include both internal and external defects;

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.

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 test includes:

Initially the 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 have internal defects 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.

Using phased array equipment, dozens of channels with different refraction angles are used to inspect the lead seals at the cable terminals, and data is collected using full matrix capture and related technologies. All lead seal ultrasonic data sets divide available defects and canvases into training sets, validation sets, and test sets. Physical samples for ML training are made to train machine learning models, and lead seal ultrasonic phased array data is processed to achieve image classification. Physical samples for ML verification are made to verify the performance of machine learning models and automatically detect internal defects in lead seals. The model is tested with two internal and external defects of similar lead seal geometry used in training, and compared with the results of human inspectors.

The machine learning model can perform high-reliability internal defect detection on typical multi-channel phased array 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

Fig. 1 is a flow chart of the detection method of the present invention;

FIG2 is a schematic diagram of a probe and a wedge device of the present invention;

FIG3 is a flow chart of ultrasonic data preprocessing according to the present invention;

FIG4 shows the present invention using a deep convolutional neural network to estimate ultrasonic scanning defects.

FIG. 5 is a diagram of the verification data evaluation of the present invention.

DETAILED DESCRIPTION

Embodiment:

As shown in FIG1 , this embodiment proposes a method for automatically detecting internal defects of lead seals using multi-channel phased array ultrasonic data based on machine learning, and the method steps include:

S1; Using phased array equipment, dozens of channels with different refraction angles are used to inspect the cable terminal seals, using full matrix capture and correlation techniques to collect data;

S2; All lead sealing ultrasonic datasets divide the available defects and canvases into training set, validation set and test set;

S3; Prepare physical samples for ML training to train machine learning models and process lead sealing ultrasonic phased array data to achieve image classification;

S4; Make physical samples for ML verification to verify the performance of the machine learning model and automatically detect internal defects of lead seals;

S5;The model was tested with two internal and external defects of similar seal geometry used in training and compared with the results of human inspectors.

Using phased array equipment, dozens of channels with different refraction angles are used to inspect the seals, and full matrix capture and correlation techniques are used to collect data. The specific process includes:

In the transceiver longitudinal device, the raw data were scanned using a dual-matrix phased array probe with a frequency of 2.25MHz. Each probe has a total of 28 elements, arranged as 7×4 elements, and an effective aperture of 19×12mm. The probes are placed on a reflector wedge with 18.9 and 0◦ vertex angles, and a 6mm spacing between the probes. The focus law is set to 40◦ to 70◦ angles, with a 1◦ step size. The scanning resolution is 1 mm, the sound path is set to 3.46 ~ 27.75µs, and the sound path resolution is 0.01µs, so the total data volume of a single step scan is 2429×31 samples. The focus point is set inside the lead seal, in the middle of the probe, without a tilt angle. The probe is positioned so that the 55◦ angle will be centered inside the lead seal, because only one scan line is recorded. The schematic diagram of the device is shown in Figure 1.

Figure 1 Schematic diagram of the probe and wedge setup. The phased array is focused into the lead seal from 40° to 70° in azimuth. The probe motion is recorded by a single encoder while a single scan line is manually scanned along the defect inside the lead seal. The contact medium used is water sprayed by hand from a spray can.

Preprocess the data to remove a large amount of redundant data contained in the multi-angle channels, including:

The phased array probe scans at an angle of 40°-70° with a step size of 1°, that is, it scans once every one degree, so waveform data at 31 angles are obtained. Each of the 31 angle complete waveform channels is considered separately and each frame is rectified, that is, the absolute value of the signal is taken;

1/2λ half of the eigenvalue of the positive definite matrix of each frame image window is matched to perform maximum pooling on a single channel, and then the data is stored in a compressed binary file to facilitate file transfer and accelerate learning

The framework performs max pooling with a window matching 0.5λ. The effect of this is to make the envelope of the data computationally efficient. The data size is reduced from 48 × 1020 to 48 × 34 (=1632) samples;

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 0…1.0);

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;

Physical samples for making ML training based on S3 are used to train machine learning models, including:

Use the original UT data as the physical sample for ML training, that is, scan the UT data of the lead-sealed plate sample with peeling and scratches on the surface and no internal defects;

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.

The current setup allows to extract a clean internal defect signal from a sample without external defects and embed it into the signal with only external defects;

All the above data must be preprocessed before being used to train the machine learning model.

The physical samples based on the ML training described in S3 are used to train the machine learning model to process the lead sealing ultrasonic phased array data for image classification, including:

The DCNN 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 lead seal internal defect image classification;

The final model is trained with all available non-test internal and external defects.

Based on S4, physical samples for ML verification are made to verify the performance of machine learning models and automatically detect internal defects of lead seals, including:

Using samples opposite to those used in training the network, in the validation set, the real defects include both internal and external defects;

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.

Based on S4, the model was tested with two internal and external defects of similar sealing geometry used in training, including:

Initially the 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 have internal defects 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.

More complex enhancement schemes are possible using so-called virtual defects, including:

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:

2. For each canvas, generate a set of 500,000 samples (divided into 50 batches of 10,000 samples each):

A number is randomly drawn to select a sample that is either perfect or has internal defects.

If the sample is designated as free of internal defects:

A window of 48 a-movies is randomly selected and added to the result data.

3. Each a-scan is moved by a random walk offset to simulate possible probe jitter during the scan, further enhancing the data.

If the sample is designated as having internal defects:

A defect is randomly picked from the population and embedded into a random position in the file.

The defect amplitude decreases with the random factor in the range of 0.5 to 1.0.

4. Each a-scan is moved by a random walk offset to simulate possible probe jitter during the scan, thereby enhancing the defect.

After embedding, 48 scanning windows are randomly selected so that the defect is completely contained within the window.

The data was further enhanced by moving each a-scan with a random walk offset, simulating possible probe jitter during the scan.

DCNN for image classification tasks can be thought of as a YOLO network to train machine learning models, including:

The DCNN architecture used is similar to the VGG16 network with 3 convolutional blocks. Each block contains two consecutive convolutional layers with rectified linear unit (ReLU) activation. 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 activation, 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 an internal. The loss function applied is binary cross entropy. During the back-propagation process, the adaptive moment estimation (ADAM) is used to calculate the new weights. The network used is shown in Figure 3. The TensorFlow library is used to preprocess and filter the data stream, and the Keras high-level API is used to build the DCNN.

Please read Figure 5 to verify the performance of the machine learning model, including:

Data from a separate internally defective seal verification sample was run through the trained machine learning model as follows: Each file was split 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 above sizes with 50% overlap across the data, 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 is designated as having an internal defect, the frame position is considered to have an internal defect. If a frame containing an internal defect is identified as having an internal defect, this is considered to be a true hit. If it is identified as not having an internal defect, it is considered to be a miss. If a frame is indicated as having an internal defect but does not contain an internal defect, it is considered to be a false call. The data does not contain any cases where an internal defect will partially fall on the frame. In total, the data divided like this contains 32 separate data frames, with 11 hit/miss chances for the near end, 11 hit/miss chances for the far end, and 10 chances for false calls.

Claims (8)

1. A method for automatic detection of internal defects of lead seals based on multi-channel phased array ultrasonic data based on machine learning, characterized in that the method steps include: S1; Using phased array equipment, dozens of channels with different refraction angles are used to inspect the cable terminal seals, using full matrix capture and correlation techniques to collect data; S2; All lead sealing ultrasonic datasets divide the available defects and canvases into training set, validation set and test set; S3; Prepare physical samples for ML training to train machine learning models and process lead sealing ultrasonic phased array data to achieve image classification; S4; Make physical samples for ML verification to verify the performance of the machine learning model and automatically detect internal defects of lead seals; S5;The model was tested with two internal and external defects of similar seal geometry used in training and compared with the results of human inspectors. 2. According to the method of automatic detection of internal defects of multi-channel phased array ultrasonic data seals based on machine learning in claim 1, it is characterized in that: All the lead sealing ultrasonic datasets in S2 divide the available defects and canvases into training sets, validation sets and test sets including: All datasets contain 50% images with only internal defects and 50% images with only external defects; 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. 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 the virtual defect method to form the validation set data. After excluding the 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. 3. According to the method of automatic detection of internal defects of lead seals based on multi-channel phased array ultrasonic data based on machine learning in claim 2, it is characterized in that the pre-processing of the training data described in S2 to remove a large amount of redundant data contained in the multi-angle channels includes: The phased array probe scans at an angle of 40°-70° with a step size of 1°, and waveform data at 31 angles are obtained. Each of the 31 complete waveform channels is considered separately, and each frame is rectified, that is, the absolute value of the signal is taken; 1/2λ is half of the eigenvalue of the positive definite matrix of each frame image window, and half of the eigenvalue of the positive definite matrix of each frame image window is matched to perform maximum pooling on a single channel, and then the data is stored in a compressed binary file to facilitate file transfer and accelerate learning. 4. According to a method for automatic detection of internal defects of lead seals based on multi-channel phased array ultrasonic data based on machine learning in claim 1, it is characterized in that the common technology for processing limited training data in machine learning is the use of data enhancement, including: the use of so-called virtual defects described in S2 can obtain more complex enhancement schemes. 5. According to the method of automatic detection of internal defects of lead seals based on multi-channel phased array ultrasonic data based on machine learning in claim 1, it is characterized in that the physical samples for making ML training described in S3 are used to train the machine learning model, including: Use the original UT data as the physical sample for ML training, that is, scan the UT data of the lead-sealed plate sample with peeling and scratches on the surface and no internal defects; 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. The current setup allows to extract a clean internal defect signal from a sample without external defects and embed it into the signal with only external defects; All the above data must be preprocessed before being used to train the machine learning model. 6. The method for automatic detection of internal defects of lead seals based on multi-channel phased array ultrasonic data based on machine learning according to claim 1 is characterized in that the physical samples based on the ML training described in S3 are used to train the machine learning model, and the lead seal ultrasonic phased array data are processed to realize image classification, including: The DCNN 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 lead seal internal defect image classification; The final model is trained with all available non-test internal and external defects. 7. According to the method of automatic detection of internal defects of lead seals based on multi-channel phased array ultrasonic data based on machine learning in claim 1, it is characterized by automatically detecting the internal defects of lead seals based on the physical samples for verifying ML described in S4 to verify the performance of the machine learning model, including: Using samples opposite to those used in training the network, in the validation set, the real defects include both internal and external defects; 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. 8. The method for automatic detection of internal defects of lead seals based on multi-channel phased array ultrasonic data based on machine learning according to claim 1 is characterized in that the model described in S4 and the internal and external defects of similar lead seal geometry used in training are tested, comprising: Initially the 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 have internal defects 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.