A method for identifying welding corrosion of copper braids in cable accessories based on deep learning

Abstract

The present invention discloses a method for identifying corrosion in copper braided cable accessory welds based on deep learning. The method comprises the following steps: collecting multiple actual corrosion defects at the welds of copper braided cable accessory welds; artificially simulating and creating multiple corrosion defects at the welds of copper braided cable accessory welds; performing non-destructive detection on the collected and simulated corrosion defects at the welds of copper braided cable accessory welds to obtain different types of corrosion defect echo samples; obtaining multiple signal features in the time domain, frequency domain, and time-frequency domain from the different types of corrosion defect echo samples, labeling the corrosion defect types, and creating a data set; building a corrosion defect recognition model for corrosion defect identification; training the corrosion defect recognition model using the generated data set; and using the trained corrosion defect recognition model to detect the welds of copper braided cable accessory welds in use to obtain detection results. The present invention has the advantages of low cost, high recognition efficiency, and high accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a method for identifying welding corrosion of copper braided belts of cable accessories based on deep learning. Background Art

[0002] The cable accessory grounding system primarily consists of two components: 1. The connection between the cable accessory tail tube and the cable metal sheath (i.e., the lead seal location). This location can be electrically connected using a lead seal or a welded copper braid; and 2. The grounding wire or coaxial cable. Corrosion defects often occur at the connection between the cable accessory tail tube and the cable metal sheath, and some severe defects can damage the cable's primary insulation.

[0003] In recent years, cable accessory failures due to corrosion in cable grounding systems have become common. However, there is currently a lack of effective means to prevent and detect these defects, and they can only be discovered when the defects become very serious. Therefore, how to detect and identify grounding system corrosion is imminent.

[0004] Faced with the severe safety situation of cable accessory grounding systems, there is an urgent need for new inspection methods for the copper braid welding points of cable accessories in use. Non-destructive testing can protect cable accessories in use to the greatest extent. Non-destructive testing of cable accessories is the process of measuring the status characteristics of high-voltage cable accessories and evaluating them according to certain criteria by applying certain inspection technologies and analysis methods without destroying the use status of high-voltage cables. Currently, the main methods used for non-destructive testing of high-voltage cables are magnetic testing, eddy current testing, and infrared testing. For cable accessories, the magnetic particle testing method needs to be carried out in ferromagnetic materials and is not suitable for the inspection of high-voltage cable accessories. Eddy current technology mainly focuses on several skin depths and has a low inspection depth. Therefore, it is usually combined with other technologies to perform near-surface corrosion inspection. The infrared detection method has low penetration ability and poor anti-interference ability, making it unsuitable for detecting internal cable damage. At the same time, it has low sensitivity for identifying early defects and weak damage.

[0005] Furthermore, currently, after obtaining the detection signal characteristics, manual identification is still required, which is inefficient and costly in the large-scale corrosion defect identification process. Therefore, the development of a deep learning-based automatic corrosion defect signal identification method to determine the corrosion status of the copper braid welds in cable accessories is of great significance. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for identifying welding corrosion of copper braided belts of cable accessories based on deep learning, which has a large recognition distance, high recognition efficiency and low cost.

[0007] In order to solve the above technical problems, the present invention provides a method for identifying welding corrosion of copper braids of cable accessories based on deep learning, comprising:

[0008] Collect corrosion defects at the welding points of copper braids of various actual cable accessories;

[0009] By artificially simulating and creating corrosion defects at the welding points of copper braids of various cable accessories;

[0010] Using testing equipment, non-destructive testing was performed on the corrosion defects at the welding points of copper braids of cable accessories, both collected in practice and simulated, to obtain different types of corrosion defect echo samples.

[0011] Obtain various signal features in the time domain, frequency domain, and time-frequency domain from different types of corrosion defect echo samples, label the corrosion defect types, and create a data set;

[0012] Build a corrosion defect recognition model for corrosion defect identification;

[0013] The corrosion defect recognition model is trained by preparing the obtained data set to obtain a trained corrosion defect recognition model;

[0014] The trained corrosion defect recognition model is used to detect the welding points of the copper braids of cable accessories in use and obtain the detection results.

[0015] Furthermore, the actual collection includes not only corrosion defects at the welding points of the copper braids of cable accessories, but also non-corrosion defects at the welding points of the copper braids of cable accessories.

[0016] Furthermore, the detection equipment is an ultrasonic guided wave detection equipment.

[0017] Furthermore, the ratio of the number of corrosion defects actually collected at the welding points of the copper braids of cable accessories to the number of corrosion defects at the welding points of the copper braids of various cable accessories artificially simulated and manufactured is 1:4.

[0018] Furthermore, multiple signal features in the time domain, frequency domain, and time-frequency domain are obtained from different types of corrosion defect echo samples, including form factor, margin index, mean square frequency, frequency variance, mean, cross entropy, and central moment. These multiple signal features form a 7-dimensional feature vector.

[0019] Among them, the shape factor S reflects the ratio of the effective signal segment to the average signal segment:

[0020]

[0021] X rms is the effective signal segment, also known as the mean, and Y is the average signal segment;

[0022] The margin index L is the reserved error allowable range:

[0023]

[0024] X max is the signal peak value;

[0025] The mean square frequency (MSF) calculates random values:

[0026]

[0027] f is the signal frequency, S(f) is the sampling signal;

[0028] Frequency VF variance:

[0029]

[0030] Mean X rms Describe the average energy of the signal:

[0031]

[0032] x is the current signal value, N is the total number of signals;

[0033] Cross entropy is a feature transformation parameter that describes the amount of information in the sample signal, including uncertainty and disorder;

[0034] The central moment B describes the distribution characteristics of the sample values:

[0035]

[0036] is the signal average value;

[0037] The 7-dimensional feature vector is made into a point image, the corrosion type corresponding to the feature image is marked, and a data set is made.

[0038] Furthermore, the corrosion defect recognition model adopts a convolutional neural network model. In the convolutional neural network model, the sample signal image is used as input, and the image input end is enhanced by a data enhancement method. The Mish activation function is used to reduce the computational memory cost and enhance the CNN learning ability. A bottom-up feature pyramid is added after the FPN layer to improve the feature extraction ability. The CIOU_Loss loss function is used for regression training and the DIOU_nms method is used for prediction box screening to complete the output layer.

[0039] Furthermore, the backbone network of the convolutional neural network model consists of three parts: CSPDarknet53, Mish activation function and Dr21opblock; the CSPDarknet53 network is responsible for extracting features from the image and outputting feature maps featuremap1-3 at three scales.

[0040] Furthermore, the convolutional neural network model includes five basic components: ①CBM: the smallest component in the network structure, consisting of Conv+Bn+Mish activation functions; ②CBL: consisting of Conv+Bn+Leaky_relu activation functions; ③Res unit: residual structure, which makes the network construction deeper; ④CSPn: consisting of convolutional layer and n Res unint modules Concat; ⑤SPP: multi-scale fusion is performed using 1×1, 5×5, 9×9, 13×13 maximum pooling methods.

[0041] The beneficial effects of implementing the present invention are:

[0042] 1) Ultrasonic guided waves are used to detect the welding points of the copper braids of cable accessories. The ultrasonic echo information is received by the ultrasonic probe, with a long propagation distance, a large detection distance and high detection efficiency.

[0043] 2) The convolutional neural network model has good corrosion damage feature extraction capabilities and can perform large-scale rapid and automatic identification of the welding points of copper braids of cable accessories with low cost and high recognition efficiency.

[0044] 3) Multiple signal features are acquired from different types of corrosion defect echo samples in the time, frequency, and time-frequency domains, including form factor, margin index, mean square frequency, frequency variance, mean, cross entropy, and central moment. These features form a 7-dimensional feature vector. This 7-dimensional feature vector ensures the accuracy of model recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a principle flow chart of a method for identifying welding corrosion of copper braided belts of cable accessories based on deep learning in the present invention;

[0046] Figure 2 This is a schematic diagram of the principle of a corrosion defect recognition model used in a method for identifying welding corrosion of copper braided tapes of cable accessories based on deep learning in the present invention;

[0047] Figure 3 This is the main framework diagram of the convolutional neural network algorithm;

[0048] Figure 4 Schematic diagram for detecting and identifying welding corrosion defects on copper braids of cable accessories. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It is hereby stated that any directional terms such as "up," "down," "left," "right," "front," "back," "inside," and "outside" that appear or will appear herein are based solely on the accompanying drawings and are not intended to limit the present invention.

[0050] like Figure 1 As shown, the present invention provides a method for identifying welding corrosion of copper braids of cable accessories based on deep learning, comprising the following steps:

[0051] S1. Collect corrosion defects at the welding points of copper braids of various actual cable accessories;

[0052] In this step, the corrosion defects collected include pitting corrosion, uniform corrosion, stress corrosion cracking, crevice corrosion, filiform corrosion, etc.; in addition, it also includes corrosion-free defects at the welding points of the copper braids of cable accessories, that is, no corrosion defects occur.

[0053] S2. Artificially simulate and create corrosion defects at the welding points of copper braids of various cable accessories;

[0054] S3. Using ultrasonic guided wave testing equipment, non-destructive testing was performed on the corrosion defects at the welds of the copper braids of cable accessories collected in practice and simulated, obtaining different types of corrosion defect echo samples.

[0055] In this step, the ratio of the number of corrosion defects actually collected at the welding points of the copper braided belts of cable accessories to the number of corrosion defects at the welding points of the copper braided belts of various cable accessories artificially simulated and manufactured is 1:4. The purpose of mixing the two is to reduce the signal error of the artificially produced corrosion defects.

[0056] S4. Obtain various signal features in the time domain, frequency domain, and time-frequency domain from different types of corrosion defect echo samples, label the corrosion defect types, and create a data set;

[0057] In this step, multiple signal features in the time domain, frequency domain, and time-frequency domain are obtained from different types of corrosion defect echo samples, including form factor, margin index, mean square frequency, frequency variance, mean, cross entropy, and central moment. These multiple signal features form a 7-dimensional feature vector.

[0058] Among them, the shape factor S reflects the ratio of the effective signal segment to the average signal segment:

[0059]

[0060] X rms is the effective signal segment, also known as the mean, and Y is the average signal segment;

[0061] The margin index L is the reserved error allowable range:

[0062]

[0063] X max is the signal peak value;

[0064] The mean square frequency (MSF) calculates random values:

[0065]

[0066] f is the signal frequency, S(f) is the sampling signal;

[0067] Frequency VF variance:

[0068]

[0069] Mean X rms Describe the average energy of the signal:

[0070]

[0071] x is the current signal value, N is the total number of signals;

[0072] Cross entropy is a feature transformation parameter that describes the amount of information in the sample signal, including uncertainty and disorder;

[0073] The central moment B describes the distribution characteristics of the sample values:

[0074]

[0075] is the signal average value;

[0076] The 7-dimensional feature vector is made into a point image, the corrosion type corresponding to the feature image is marked, and a data set is made.

[0077] S5. Build a corrosion defect recognition model for corrosion defect recognition;

[0078] In this step, if Figure 2 As shown in the figure, the corrosion defect recognition model adopts a convolutional neural network model. In the convolutional neural network model, the sample signal image is used as input, and the data enhancement method is used to enhance the image input. The Mish activation function is used to reduce the computational memory cost and enhance the CNN learning ability. A bottom-up feature pyramid is added after the FPN layer to improve the feature extraction ability. The CIOU_Loss loss function is used for regression training and the DIOU_nms method is used for prediction box screening to complete the output layer.

[0079] The convolutional neural network model's backbone consists of three components: CSPDarknet53, Mish activation function, and Dr21 opblock. The CSPDarknet53 network extracts image features and outputs feature maps (featuremap1-3) at three scales. CSPDarkNet53 is used to extract model features, significantly reducing network parameters while maintaining accuracy and minimizing computational bottlenecks and memory costs.

[0080] like Figure 3 As shown in the figure, the convolutional neural network model includes five basic components: ①CBM: the smallest component in the network structure, consisting of Conv+Bn+Mish activation functions; ②CBL: consisting of Conv+Bn+Leaky_relu activation functions; ③Res unit: residual structure, which makes the network deeper; ④CSPn: consists of a convolutional layer and n Res unint modules Concat; ⑤SPP: uses 1×1, 5×5, 9×9, 13×13 maximum pooling methods for multi-scale fusion.

[0081] The first step of the process is to obtain the recognition prediction results from the three feature layers; the second step is to decode the recognition prediction results. The coordinates of the center point of the final bounding box are obtained using the following calculation formula:

[0082]

[0083] Among them, c x and c y is the number of grids between the upper left corner of the grid where the point is located and the upper left corner of the entire image, t x and t y are the x and y distances from the target center to the upper left corner, p w and p h is the side length of the prior box, t w and t h are the width and height of the prediction box respectively, σ is the activation function, generally the sigmoid function is used, which is between [0,1].

[0084] S6. Training the corrosion defect recognition model using the prepared data set to obtain a trained corrosion defect recognition model;

[0085] Set the training parameters. To reduce training time and obtain a good training model, set the training parameters before training. For example, you can input multiple sample images per training session, set the image input quantity to 16, and the maximum number of iterations should not exceed 6000. The learning rate can be set to 0.001, the learning rate step size can be set to 4800 and 5400, and the control parameter can be set to 1000. Training ends when the loss value is less than the set threshold or the maximum number of iterations is reached, resulting in a convolutional neural network model for identifying weld corrosion defects in copper braids of cable accessories.

[0086] S7. Use the trained corrosion defect recognition model to detect the welding parts of the copper braids of the cable accessories in use and obtain the detection results.

[0087] like Figure 4 The figure below shows a schematic diagram of corrosion defect detection and identification for copper braid welds of cable accessories. At the inspection site, ultrasonic guided wave testing equipment is used to inspect multiple sets of copper braid welds in use. The detected signals are coded and sent to a backend server for processing. A convolutional neural network model trained to automatically identify corrosion defect types is used on the multiple sets of signal images of copper braid welds transmitted to the server to identify the corrosion type. The defective results and their corresponding types are then transmitted back to the inspection site. Once the results are obtained on-site, the corresponding copper braid welds are repaired or replaced to ensure safe and reliable operation of the grounding system.

[0088] Although the description of the present disclosure has been quite detailed and particularly describes several embodiments, it is not intended to be limited to any of these details or embodiments or any particular embodiment, but should be regarded as providing a broad possible interpretation of these claims in view of the prior art by reference to the appended claims, thereby effectively covering the intended scope of the present disclosure. In addition, the above description of the present disclosure is based on the embodiments foreseen by the inventors, which is intended to provide a useful description, and those non-substantial changes to the present disclosure that have not yet been foreseen may still represent equivalent changes to the present disclosure.

Claims

1. A method for identifying welding corrosion of copper braided belts of cable accessories based on deep learning, characterized in that: include: Collect corrosion defects at the welding points of copper braids of various actual cable accessories; By artificially simulating and creating corrosion defects at the welding points of copper braids of various cable accessories; Using testing equipment, non-destructive testing was performed on the corrosion defects at the welding points of copper braids of cable accessories, both collected in practice and simulated, to obtain different types of corrosion defect echo samples. Obtain various signal features in the time domain, frequency domain, and time-frequency domain from different types of corrosion defect echo samples, label the corrosion defect types, and create a data set; Build a corrosion defect recognition model for corrosion defect identification; The corrosion defect recognition model is trained by preparing the obtained data set to obtain a trained corrosion defect recognition model; The trained corrosion defect recognition model is used to detect the welding points of the copper braids of cable accessories in use and obtain the test results. The detection equipment is an ultrasonic guided wave detection equipment; Acquire multiple signal features in the time domain, frequency domain, and time-frequency domain from different types of corrosion defect echo samples, including form factor, margin index, mean square frequency, frequency variance, mean, cross entropy, and central moment. The mean is the average energy of the signal. These multiple signal features form a 7-dimensional feature vector. Among them, the shape factor S reflects the ratio of the effective signal segment to the average signal segment: X rms is the effective signal segment, also known as the mean, and Y is the average signal segment; The margin index L is the reserved error allowable range: X max is the signal peak value; The mean square frequency (MSF) calculates random values: f is the signal frequency, S(f) is the sampling signal; Frequency VF variance: Mean X rms Describe the average energy of the signal: x is the current signal value, N is the total number of signals; Cross entropy is a feature transformation parameter that describes the amount of information in the sample signal, including uncertainty and disorder; The central moment B describes the distribution characteristics of the sample values: is the signal average value; The 7-dimensional feature vector is made into a point image, the corrosion type corresponding to the feature image is marked, and a data set is made.

2. A cable accessory copper braid welding corrosion identification method based on deep learning according to claim 1, characterized in that: The actual data collected include not only the corrosion defects at the welding points of the copper braids of cable accessories, but also the non-corrosion defects at the welding points of the copper braids of cable accessories.

3. A method for identifying welding corrosion of copper braided tape of cable accessories based on deep learning according to any one of claims 1-2, characterized in that: The ratio of the number of corrosion defects at the welding points of copper braids of cable accessories actually collected to the number of corrosion defects at the welding points of copper braids of various cable accessories artificially simulated and manufactured is 1:

4.

4. The method for identifying welding corrosion of copper braided tape of cable accessories based on deep learning according to claim 1 is characterized in that: The corrosion defect recognition model adopts a convolutional neural network model. In the convolutional neural network model, the sample signal image is used as input, and the image input end is enhanced by the data enhancement method. The Mish activation function is used to reduce the computational memory cost and enhance the CNN learning ability. A bottom-up feature pyramid is added after the FPN layer to improve the feature extraction ability. The CIOU_Loss loss function is used for regression training and the DIOU_nms method is used for prediction box screening to complete the output layer.

5. A cable accessory copper braid welding corrosion identification method based on deep learning according to claim 4, characterized in that: The backbone network of the convolutional neural network model consists of three parts: CSPDarknet53, Mish activation function and Dropblock; the CSPDarknet53 network is responsible for extracting features from the image and outputting feature maps featuremap1-3 at three scales.

6. A cable accessory copper braid welding corrosion identification method based on deep learning according to claim 5, characterized in that: The convolutional neural network model includes five basic components: ①CBM: the smallest component in the network structure, consisting of Conv+Bn+Mish activation functions; ②CBL: consisting of Conv+Bn+Leaky_relu activation functions; ③Resunit: residual structure, which makes the network deeper; ④CSPn: consisting of convolutional layer and n Res unit modules Concat; ⑤SPP: multi-scale fusion is performed using 1×1, 5×5, 9×9, 13×13 maximum pooling methods.

Images