Defect Detection Method and Device for Insulating Tools in Live Working

Abstract

Embodiments of the present invention provide a method and apparatus for defect detection of live working insulation tools. The method includes obtaining A-scan signal display waveform data of the insulation tool by an ultrasonic detection system as sampling data; preprocessing the sampling data to obtain candidate samples; performing recall processing on high echo samples and head and tail echo samples in the candidate samples, and using the samples after recall processing as a training set; constructing a defect judgment model, training the defect judgment model using the training set, and using the defect judgment model that meets the ranking loss as a trained defect judgment model; inputting the sampling data of the target insulation tool into the trained defect judgment model and outputting a defect result. Through the insulation tool defect detection algorithm of the present application, small defects in the insulation tool can be stably detected, the flaw detection quality of the insulation tool is guaranteed, and the maximization of economic benefits is ensured.

Description

Technical Field

[0001] The present invention generally relates to the field of non-destructive testing, and more specifically, to a method and device for detecting defects in live working insulation tools. Background Art

[0002] Live working insulation tools are generally made of hollow tubes and solid rods processed from glass fiber-reinforced epoxy resin composites, and have good mechanical and insulation properties, and are widely used in the power and grid industries.

[0003] Since the insulation tools are subjected to the combined action of high-voltage electricity and mechanical stress during long-term use, microscopic defects such as delamination, air gaps, and impurities are likely to form inside them, resulting in a decline in their mechanical and insulation properties, and ultimately endangering the safety of power operation personnel and maintenance equipment. The use of ultrasonic testing technology can effectively diagnose internal defects of materials and ensure the safety of operators. Traditional non-destructive testing methods use manual hand-held detectors and ordinary piezoelectric ultrasonic probes to detect insulation rods.

[0004] There are technical problems in the prior art that rely on the experience, operation methods, and professionalism of flaw detection personnel when detecting insulation rods, the data cannot be recorded and traced in real time, and the reliability and credibility of flaw detection are not high. Summary of the Invention

[0005] According to an embodiment of the present invention, a defect detection solution for live working insulation tools is provided. In this solution, the ultrasonic probe is manually attached to the surface of the insulation rod, and then the ultrasonic probe is pushed on the surface of the insulation rod by manual or mechanical power to collect a large amount of ultrasonic data from different directions, and a defect detection algorithm based on machine learning is used to identify abnormal waveforms, so as to screen out the damage defects existing in the insulation rods to be detected, and solve the technical problems of relying on the experience of flaw detection personnel and low reliability and credibility in the prior art.

[0006] In the first aspect of the present invention, a method for detecting defects in live working insulation tools is provided. The method includes:

[0007] Obtain the A-scan signal display waveform data of the insulation tool by the ultrasonic detection system as sampling data.

[0008] Preprocess the sampling data to obtain candidate samples.

[0009] Perform a recall process on the high echo samples and the head and tail echo samples in the candidate samples, and use the samples after the recall process as the training set.

[0010] Construct a defect judgment model, train the defect judgment model using the training set, and use the defect judgment model that meets the ranking loss as the trained defect judgment model.

[0011] Input the sampling data of the target insulation tool into the trained defect judgment model to output the defect result.

[0012] The preprocessing of the sampling data includes:

[0013] Filling the missing values in the sampling data using linear interpolation. And / or

[0014] Identifying the abnormal sampling points in the sampling data, and replacing the data of the abnormal sampling points with the data linearly interpolated from the sampling points before and after the abnormal sampling points. And / or

[0015] Expanding the statistical features and morphological features at the sample level and window level in the sampling data. And / or

[0016] Normalizing the sampling data.

[0017] The filling of the missing values in the sampling data using linear interpolation includes:

[0018]

[0019] where X k is the missing value to be filled; X i is the missing value; X i-1 is the non-missing value before the missing value X i ; X j is the non-missing value after the last missing value; p is the set of missing values; i is the subscript corresponding to the first missing value; j is the subscript corresponding to the next value after the last missing value.

[0020] The identification of the abnormal sampling points in the sampling data includes:

[0021] If two consecutive sampling points in the sampling data meet the abnormal sampling point judgment condition, then the latter sampling point of the two sampling points is the abnormal sampling point.

[0022] The abnormal sampling point judgment condition is.

[0023] |X a -X a-1 |>kσ

[0024] where σ is the standard deviation of the sampling data; k is a constant; X a and X a-1 are two consecutive sampling points.

[0025] The recall processing of the high echo samples in the candidate samples includes:

[0026] Obtaining the set of intervals where the sample values are greater than the first threshold from the candidate samples.

[0027] Merge the continuous points in the interval set into intervals to obtain several sub-intervals.

[0028] If the length of at least one sub-interval is greater than the second threshold, the sample corresponding to this sub-interval is a high echo sample, and the high echo sample is recalled.

[0029] The head and tail echo samples in the candidate samples are recalled, including:

[0030] If there are high echo samples in the candidate samples that meet the first condition or the second condition of the high echo interval, the high echo samples that meet the conditions are recalled.

[0031] The first condition is:

[0032] max(t ai ,t bi )<M left

[0033] The second condition is:

[0034] max(t ai ,t bi )>M right

[0035] Among them, max(t ai ,t bi ) is the larger value of t ai and t bi ; t ai , t bi are the left and right boundaries of the i-th high echo interval respectively; M left is the left endpoint of the recall interval, and M right is the right endpoint of the recall interval.

[0036] The defect judgment model includes a typing layer and a first fully connected neural network layer connected in sequence, and a convolutional layer, a multi-head self-attention network layer, a second fully connected neural network layer, and an average pooling layer connected in sequence; it also includes a splicing layer and a third fully connected neural network layer.

[0037] The typing layer, as an input layer of the defect judgment model, is used to obtain sample-level features, perform look-up table embedding, and the output of the typing layer is input to the first fully connected neural network layer. The first fully connected neural network layer interacts and assigns weights to the input sliding window features and the original sequence, and outputs the first hidden vector.

[0038] The convolutional layer, as another input layer of the defect judgment model, is used to obtain window-level features and the sampling data, map the data, output the mapping result to the multi-head self-attention network layer for feature cross, and the cross result passes through the second fully connected neural network layer and the average pooling layer in sequence and then outputs the second hidden vector.

[0039] A splicing layer for splicing the first hidden vector and the second hidden vector.

[0040] A third fully-connected neural network layer takes the output result of the splicing layer as input and outputs a defect score result, and takes the defect score result as the output result of the defect judgment model.

[0041] Training the defect judgment model using the training set, and taking the defect judgment model that satisfies the ranking loss as the trained defect judgment model, including:

[0042] Select a defective waveform from the training set as a label sample, and calculate the DTW distance between all samples in the training set and the label sample.

[0043] Sort all samples in the training set in ascending order of the DTW distance from the label sample to obtain a reference sequence.

[0044] Sort the defect score results of each sample output by the defect judgment model in descending order according to the reference sequence to obtain a prediction sequence.

[0045] Calculate the pairwise ranking loss function according to the reference sequence and the prediction sequence.

[0046] The pairwise ranking loss is calculated as follows:

[0047]

[0048] Where s i and s j are the scores of the model for the i-th and j-th samples, and there is

[0049]

[0050] Inputting the sampling data of the target insulating tool into the trained defect judgment model and outputting a defect result, including:

[0051] If the scores output by the trained defect judgment model are all less than a preset third threshold or there are no samples for recall processing, it is considered that the target insulating tool has no defect.

[0052] If there are samples with scores greater than the preset third threshold in the scores output by the trained defect judgment model, the waveform with the highest score is output as the defect result.

[0053] In a second aspect of the present invention, a defect detection device for live working insulating tools is provided. The device includes:

[0054] A data acquisition module for acquiring the A-scan signal display waveform data of an insulation tool by an ultrasonic detection system as sampling data.

[0055] A preprocessing module for preprocessing the sampling data to obtain candidate samples.

[0056] A recall processing module for performing recall processing on the high echo samples and the first and last echo samples in the candidate samples, and using the samples after recall processing as the training set.

[0057] A model training module for constructing a defect judgment model, training the defect judgment model using the training set, and using the defect judgment model that meets the sorting loss as the trained defect judgment model.

[0058] A defect judgment module for using the trained defect judgment model to judge the defect result of the sampling data of the target insulation tool.

[0059] It should be understood that the content described in the invention content part is not intended to limit the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. Description of the Drawings

[0060] In combination with the drawings and with reference to the following detailed description, the above and other features, advantages and aspects of the embodiments of the present invention will become more obvious. In the drawings, the same or similar reference numerals represent the same or similar elements, where:

[0061] Figure 1 Shows a flowchart of a method for detecting defects of live working insulation tools according to an embodiment of the present invention.

[0062] Figure 2 Shows a flowchart of a defect judgment model according to an embodiment of the present invention.

[0063] Figure 3 Shows a block diagram of a device for detecting defects of live working insulation tools according to an embodiment of the present invention. Detailed Embodiments

[0064] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0065] In addition, the term "and / or" in this text is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this text generally represents an "or" relationship between the preceding and following associated objects.

[0066] As shown in this embodiment Figure 1 、 Figure 2 as follows. The method includes:

[0067] S101. Obtain the A-scan signal display waveform data of the insulation tool by the ultrasonic detection system as sampling data.

[0068] In this embodiment, the A-scan signal display waveform data is a way of ultrasonic detection display, which uses the form of rectangular coordinates to display the amplitude and propagation time of ultrasonic echoes.

[0069] S102. Preprocess the sampling data to obtain candidate samples.

[0070] In this embodiment, since the A-scan signal display waveform data may vary greatly due to different sampling methods, it is necessary to preprocess the sampling data into the same distribution.

[0071] In this embodiment, after obtaining the sampling data, when it is found that there are missing segments in the sampling data, linear interpolation is used to fill the missing values in the sampling data.

[0072] The filling of the missing values in the sampling data by using linear interpolation includes:

[0073]

[0074] where X k is the missing value to be filled; X i is the missing value; X i-1 is the previous non-missing value of the missing value X i ; X j is the next non-missing value after the last missing value; p is the set of missing values; i is the subscript corresponding to the first missing value; j is the subscript corresponding to the next value after the last missing value.

[0075] In this embodiment, when there is noise or machine failure during the process of obtaining the sampling data, the data of the sampling points will deviate greatly from the normal values. Further, the sampling points that deviate greatly from the normal values are screened and set as abnormal sampling points.

[0076] Further, identify the abnormal sampling points in the sampling data, and replace the data of the abnormal sampling points with the linearly interpolated data of the previous and next sampling points of the abnormal sampling points.

[0077] Identify abnormal sampling points in the sampled data, including:

[0078] If two consecutive sampling points in the sampled data meet the abnormal sampling point judgment condition, the latter sampling point of the two sampling points is an abnormal sampling point.

[0079] The abnormal sampling point judgment condition is:

[0080] |X a -X a-1 |>kσ

[0081] where σ is the standard deviation of the sampled data; k is a constant; X a and X a-1 are two consecutive sampling points.

[0082] In this embodiment, in order to facilitate the machine learning model to identify waveform features and patterns, it is necessary to extract and expand the sample-level data and window-level data in the collected data, including: expansion of sample-level statistical features and morphological features and expansion of window-level statistical features and morphological features.

[0083] Among them, expanding the statistical features and morphological features of the sample level and window level in the sampled data includes:

[0084] For the statistical features and morphological features of the sample level.

[0085] · Mean, standard deviation.

[0086] · Minimum value, maximum value.

[0087] · Median, quartile, interquartile range.

[0088] · Deflection, torsion, central moment, second moment.

[0089] For the statistical features and morphological features of the window level.

[0090] · Mean, standard deviation of the 10-step sliding window.

[0091] · Minimum and maximum values of the 10-step sliding window.

[0092] · Deflection, torsion of the 10-step sliding window.

[0093] · Mean, standard deviation of the 30-step sliding window.

[0094] · Minimum and maximum values of the 30-step sliding window.

[0095] · Deflection, torsion of the 30-step sliding window.

[0096] For the statistical features and morphological features at the window level, zeros are filled at the tails of the statistical features and morphological features, and they are stacked together with the sampled data in the form of a vector matrix to obtain the augmented sample data.

[0097] In this embodiment, it is necessary to perform further normalization processing on the sampled data to ensure the unity of the feature space where each sampled data is located.

[0098] In this embodiment, the normalization processing of the sampled data includes:

[0099] Perform MinMax normalization operation. On the dimension of the original data, the samples are normalized using the maximum and minimum values of all known data.

[0100] The MinMax normalization operation is:

[0101]

[0102] where x is the dimension of the original data, x i,norm is the sample value after the normalization operation, x i is the current value of the normalization operation, x min is the minimum value in the sample data, x max is the maximum value in the sample data.

[0103] In this embodiment, when preprocessing the sampled data obtained in step S101, it is possible to analyze according to the real-time situation of the sampled data and use one or more data preprocessing methods in step S102 for processing to obtain candidate samples.

[0104] S103. Perform recall processing on the high echo samples and the head and tail echo samples in the candidate samples, and use the samples after the recall processing as the training set.

[0105] In this embodiment, since the number of candidate samples returned by the defect detection system is huge, directly using the machine learning model is inefficient, and the sampled data with noise will interfere with the machine learning model. Therefore, it is necessary to perform recall processing on the candidate samples to perform the initial screening of the candidate samples to obtain a smaller-scale candidate sample.

[0106] In this embodiment, there are two ways to perform recall processing on the candidate samples, including performing recall processing on the high echo samples in the candidate samples and performing recall processing on the head and tail echo samples in the candidate samples.

[0107] Furthermore, because the waveform of the defect candidate sample has the high echo characteristic, it is necessary to perform recall processing on the high echo samples in the candidate samples.

[0108] Performing recall processing on the high echo samples in the candidate samples includes:

[0109] Obtain a set of intervals from the candidate samples where the sample values are greater than the first threshold.

[0110] Merge the continuous points in the interval set into intervals to obtain several sub - intervals.

[0111] If the length of at least one sub - interval is greater than the second threshold, the sample corresponding to this sub - interval is a high - echo sample, and recall the high - echo samples.

[0112] In this embodiment, the second threshold is generally set to 5.

[0113] Furthermore, since the high - echo of the defect candidate samples only appears in the middle of the waveform, it is necessary to use the head and tail echo samples in the candidate samples for recall processing to filter out the high - echoes generated by noise interference at the head and tail of the candidate samples.

[0114] The recall processing of the head and tail echo samples in the candidate samples includes:

[0115] If there are high - echo samples in the candidate samples that meet the first condition or the second condition of the high - echo interval, recall the high - echo samples that meet the conditions.

[0116] The first condition is:

[0117] max(t ai ,t bi )<M left

[0118] The second condition is:

[0119] max(t ai ,t bi )>M right

[0120] Where max(t ai ,t bi ) is the larger value of t ai and t bi ; t ai , t bi are the left and right boundaries of the i - th high - echo interval respectively; M left is the left endpoint of the recall interval, and M right is the right endpoint of the recall interval.

[0121] Where M left is generally set to 0.15 times the length of a single - waveform sample, and M right is generally set to 0.85 times the length of a single - waveform sample.

[0122] S104. Construct a defect judgment model, train the defect judgment model using the training set, and use the defect judgment model that meets the ranking loss as the trained defect judgment model.

[0123] In this embodiment, after obtaining the training set obtained in step S103, construct a defect judgment model and train the defect judgment model using the training set, as Figure 2 shown:

[0124] In this embodiment, the defect judgment model includes a hashing layer and a first fully connected neural network layer connected in sequence, and a convolutional layer, a multi-head self-attention network layer, a second fully connected neural network layer, and an average pooling layer connected in sequence; it also includes a splicing layer and a third fully connected neural network layer.

[0125] The hashing layer, as an input layer of the defect judgment model, is used to obtain sample-level features, perform look-up table embedding, and the output of the hashing layer is input to the first fully connected neural network layer. The first fully connected neural network layer interacts with and assigns weights to the input sliding window features and the original sequence, and outputs a first hidden vector.

[0126] Among them, after obtaining the sample-level features, the defect judgment model equally bins the sample-level features into discrete features, then performs look-up table embedding, and inputs them into the first fully connected neural network layer.

[0127] The convolutional layer, as another input layer of the defect judgment model, is used to obtain window-level features and the sampled data, map the data, output the mapping result to the multi-head self-attention network layer for feature crossing, and the crossing result passes through the second fully connected neural network layer and the average pooling layer in sequence and then outputs a second hidden vector.

[0128] Among them, for the sliding window features and the original sequence, first map the data through a 1x1 convolutional network, then perform window-level feature crossing through a multi-head attention network, and finally output through the second fully connected network layer and the average pooling layer.

[0129] The splicing layer is used to splice the first hidden vector and the second hidden vector.

[0130] The third fully connected neural network layer takes the output result of the splicing layer as input and outputs a defect score result, and uses the defect score result as the output result of the defect judgment model.

[0131] Among them, the third fully connected neural network layer can scale the feature dimension to 1 and output a defect score.

[0132] Furthermore, train the defect judgment model using the training set, and use the defect judgment model that meets the ranking loss as the trained defect judgment model.

[0133] In this embodiment, since there are deviations in the distribution of different insulation tool candidate samples, in order to avoid the influence of the deviations, the DTW distance is not directly used as the label to train the model, and the ranking loss is used instead.

[0134] Select a defective waveform from the training set as the label sample, and calculate the DTW distance between all samples in the training set and the label sample.

[0135] Sort all samples in the training set in ascending order of the DTW distance from the label sample to obtain a reference sequence.

[0136] Sort the defect score results of each sample output by the defect judgment model in descending order according to the reference sequence to obtain a prediction sequence.

[0137] Calculate the pairwise ranking loss function according to the reference sequence and the prediction sequence.

[0138] The pairwise ranking loss is calculated as follows:

[0139]

[0140] where s i and s j are the scores of the model for the i-th and j-th samples, and there is

[0141]

[0142] In this embodiment, the loss function is used to calculate the gradient backpropagation in deep learning training to update the parameters.

[0143] S105. Input the sampling data of the target insulation tool into the trained defect judgment model to output a defect result.

[0144] In this embodiment, the step of inputting the sampling data of the target insulation tool into the trained defect judgment model to output a defect result includes:

[0145] If all the scores output by the trained defect judgment model are less than a preset third threshold or there are no samples for recall processing, it is considered that the target insulation tool has no defect.

[0146] Among them, the preset third threshold is the score obtained by the defect judgment model for the ultrasonic sample at the defect critical point.

[0147] If there are samples with scores greater than the preset third threshold among the scores output by the trained defect judgment model, output the waveform with the highest score as the defect result.

[0148] Using the insulation tool defect detection algorithm, it is determined whether there is a defect in the detected target from the large-scale A-scan waveform data of the detected target scanned in different directions collected by the ultrasonic detection system. If there is a defect, the corresponding abnormal waveform is returned at the same time. Through the insulation tool defect detection algorithm of the present application, small defects in the insulation tool can be stably detected, ensuring the inspection quality of the insulation tool and maximizing the economic benefits.

[0149] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0150] The above is the introduction of the method embodiments. The following is a further description of the solution of the present invention through the device embodiments having the same inventive concept as the methods in the foregoing embodiments.

[0151] As Figure 3 shown, the device 300 includes:

[0152] A data acquisition module 310, configured to acquire the A-scan signal display waveform data of the insulation tool by the ultrasonic detection system as sampling data.

[0153] A preprocessing module 320, configured to preprocess the sampling data to obtain candidate samples.

[0154] A recall processing module 330, configured to perform recall processing on the high echo samples and the head and tail echo samples in the candidate samples, and use the samples after the recall processing as the training set.

[0155] A model training module 340, configured to construct a defect judgment model, train the defect judgment model using the training set, and use the defect judgment model that meets the sorting loss as the trained defect judgment model.

[0156] A defect judgment module 350, configured to use the trained defect judgment model to judge the defect result of the sampling data of the target insulation tool.

[0157] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the described modules can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0158] In the technical solution of the present invention, the acquisition, storage, and application of the user's personal information involved all comply with the provisions of relevant laws and regulations and do not violate public order and good customs.

[0159] The various embodiments of the systems and techniques described above in this specification can be implemented in digital electronic circuitry, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on a chip (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include: being implemented in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be a special-purpose or general-purpose programmable processor that receives data and instructions from, and transmits data and instructions to, a storage system, at least one input device, and at least one output device.

[0160] The program code for implementing the methods of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing device, such that the program codes, when executed by the processor or controller, cause the functions / operations specified in the flowchart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.

[0161] In the context of this invention, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0162] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, speech input, or tactile input).

[0163] The systems and techniques described herein can be implemented in a computing system including backend components (e.g., as a data server), or a computing system including middleware components (e.g., an application server), or a computing system including frontend components (e.g., a user computer having a graphical user interface or a web browser through which the user can interact with an implementation of the systems and techniques described herein), or a computing system including any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected to each other by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local area network (LAN), wide area network (WAN), and the Internet.

[0164] A computer system can include a client and a server. The client and the server are generally remote from each other and typically interact through a communication network. The relationship between the client and the server is generated by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, or a server of a distributed system, or a server incorporating a blockchain.

[0165] It should be understood that various forms of the processes shown above can be used, with steps reordered, added, or deleted. For example, the steps recited in the present invention can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of the present invention can be achieved, and no limitation is imposed herein.

[0166] The above specific embodiments do not constitute a limitation on the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A defect detection method for insulating tools for live working, characterized in that: include: Acquire the A-scan signal display waveform data of the ultrasonic detection system for the insulating tool as sampling data; Preprocess the sampled data to obtain candidate samples; Recall the high echo samples and the first and last echo samples in the candidate samples, and use the recalled samples as training sets; Construct a defect judgment model, use the training set to train the defect judgment model, and use the defect judgment model that satisfies the sorting loss as the trained defect judgment model; Input the sampled data of the target insulation tool into the trained defect judgment model and output the defect result; Among them, the high echo samples in the candidate samples are recalled, including: Obtaining, from the candidate samples, a set of intervals whose sample values ​​are greater than a first threshold; Merge consecutive points in the interval set into intervals to obtain several sub-intervals; If there is at least one subinterval whose length is greater than the second threshold, the sample corresponding to the subinterval is a high-echo sample, and the high-echo sample is recalled; The first and last echo samples in the candidate samples are recalled, including: If there are candidate samples that meet the first condition or the second condition of the high echo interval, the high echo samples that meet the conditions are recalled; The first condition is: max(t ai ,t bi )<M left The second condition is: max(t ai ,tb i )>M right Among them, max(t ai ,t bi ) is t ai and t bi The larger value; t ai ,t bi are the left and right boundaries of the ith high echo interval respectively; M left is the left endpoint of the recall interval, M right is the right endpoint of the recall interval.

2. The method according to claim 1, characterized in that: The preprocessing of the sampled data comprises: Fill missing values ​​in the sampled data using linear interpolation; and / or Identify abnormal sampling points in the sampled data, and replace the data of the abnormal sampling points with data of linear interpolation of sampling points before and after the abnormal sampling points; and / or Expanding sample-level and window-level statistical and morphological features in the sampled data; and / or Normalize the sampled data.

3. The method according to claim 2, characterized in that The method of filling missing values ​​in the sampled data by using linear interpolation includes: Among them, X k is the missing value to be filled; X i is a missing value; X i-1 For missing values ​​X i The previous non-missing value of X j is the non-missing value after the last missing value; p is the set of missing values; i is the subscript corresponding to the first missing value; j is the subscript corresponding to the next value of the last missing value.

4. The method according to claim 2, characterized in that: The identifying of abnormal sampling points in the sampled data includes: If two consecutive sampling points in the sampling data meet the abnormal sampling point judgment condition, the latter of the two sampling points is the abnormal sampling point; The abnormal sampling point judgment condition is: |X a -X a-1 |>kσ Among them, σ is the standard deviation of the sampled data; k is a constant; X a With X a-1 are two consecutive sampling points.

5. The method according to claim 1, characterized in that: The defect judgment model includes a scoring layer and a first fully connected neural network layer connected in sequence, and a convolutional layer, a multi-head self-attention network layer, a second fully connected neural network layer and an average pooling layer connected in sequence; and also includes a splicing layer and a third fully connected neural network layer; The scoring layer, as an input layer of the defect judgment model, is used to obtain sample-level features and perform table lookup embedding. The output of the scoring layer is input to the first fully connected neural network layer. The first fully connected neural network layer interacts and weights the input sliding window features and the original sequence, and outputs a first hidden vector. The convolution layer, as another input layer of the defect judgment model, is used to obtain the window level features and the sampled data, map the data, and output the mapping results to the multi-head self-attention network layer for feature crossover. The crossover results are sequentially passed through the second fully connected neural network layer and the average pooling layer to output the second hidden vector; A concatenation layer, used for concatenating the first hidden vector and the second hidden vector; The third fully connected neural network layer takes the output result of the concatenation layer as input, and outputs a defect score result, and uses the defect score result as the output result of the defect judgment model.

6. The method according to claim 1, characterized in that The defect judgment model is trained by using the training set, and the defect judgment model that satisfies the sorting loss is used as the trained defect judgment model, including: Select a defect waveform from the training set as a label sample, and calculate the DTW distance between all samples in the training set and the label sample; Sort all samples in the training set by their DTW distance to the label samples from small to large to obtain a reference sequence; The defect score results of each sample output by the defect judgment model are sorted from large to small according to the reference sequence to obtain a prediction sequence; Calculating a pairwise ranking loss function based on the reference sequence and the predicted sequence; The pairwise ranking loss is calculated as follows: where s i and j is the model's score for the i-th and j-th samples, and there is 7. The method according to claim 1, characterized in that The step of inputting the sampled data of the target insulation tool into the trained defect judgment model and outputting the defect result comprises: If the output scores of the trained defect judgment model are all less than the preset third threshold or there is no sample for recall processing, the target insulating tool is considered to be free of defects; If there is a sample whose output score of the trained defect judgment model is greater than the preset third threshold, the waveform with the highest score is output as the defect result.

8. A defect detection device for insulating tools for live working, characterized in that: include: A data acquisition module is used to acquire the A-scan signal display waveform data of the ultrasonic detection system for the insulation tool as sampling data; A preprocessing module is used to preprocess the sampled data to obtain candidate samples; A recall processing module is used to perform recall processing on high echo samples and first and last echo samples in candidate samples, and use the recalled samples as training sets; The model training module is used to build a defect judgment model, train the defect judgment model using the training set, and use the defect judgment model that meets the sorting loss as the trained defect judgment model; A defect judgment module, used to judge defect results of sampling data of a target insulation tool using a trained defect judgment model; Among them, the high echo samples in the candidate samples are recalled, including: Obtaining, from the candidate samples, a set of intervals whose sample values ​​are greater than a first threshold; Merge consecutive points in the interval set into intervals to obtain several sub-intervals; If there is at least one subinterval whose length is greater than the second threshold, the sample corresponding to the subinterval is a high-echo sample, and the high-echo sample is recalled; The first and last echo samples in the candidate samples are recalled, including: If there are candidate samples that meet the first condition or the second condition of the high echo interval, the high echo samples that meet the conditions are recalled; The first condition is: max(t ai ,t bi )<M left The second condition is: max(t ai ,t bi )>M right Among them, max(t ai ,t bi ) is t ai and t bi The larger value; t ai ,t bi are the left and right boundaries of the ith high echo interval respectively; M left is the left endpoint of the recall interval, M right is the right endpoint of the recall interval.

Images