Deeply-buried long tunnel defect detection method and computer equipment

Abstract

The invention discloses a deep-buried long tunnel defect detection method and computer equipment, and relates to the field of tunnel defect detection.The method comprises the steps that image data and sonar data collected by an underwater inspection robot in the tunnel operation process are obtained; generating fusion detection data according to the image data and the sonar data; inputting the fusion detection data into a defect detection model obtained by pre-training to obtain a defect detection result; wherein a part of sample data in the training set of the defect detection model is generated based on a sample data construction module, and the sample construction module is used for generating at least one second sample defect data based on the defect example data information and the first sample defect data. Automatic detection of tunnel defects can be achieved based on the artificial intelligence technology, the problems of high cost, low safety and the like existing in manual detection can be effectively avoided, and sample defect data can be expanded to train and obtain a defect detection model with higher detection capacity.

Description

Technical Field

[0001] The present invention relates to the field of tunnel defect detection, and in particular to a method and computer equipment for detecting defects in a deep-buried long tunnel. Background Art

[0002] my country's water resources are unevenly distributed in time and space. In recent years, there are more and more tunnel projects such as water diversion, hydropower diversion, and urban flood control. According to statistics, China has built a total length of more than 500 km of hydraulic tunnels (including pressure tunnels, non-pressure tunnels, pressure steel pipelines, pressure shafts and inclined shafts, etc.), and has built 22 water diversion and power generation tunnels with a length of more than 4 km. Many water diversion and power generation tunnels with a length of more than 20 km and more than 100 km are to be built. Hydraulic tunnels are buried deep underground, and the geological conditions are very complex. In addition to the large cofferdam pressure, they also have to withstand high external and internal water pressures. As these "large, long, and deeply buried" tunnels are gradually put into operation, they are generally not considered during the design and construction stage. The need for daily inspections during operation has been affected by the active fractures, large deformation of soft rocks, high ground stress, high ground temperature, high external water pressure, and construction quality legacy problems, threatening safe and healthy operation. Its operation inspection has always been a major problem at home and abroad.

[0003] Traditional tunnel inspections are all done by draining hydraulic tunnels regularly and then conducting manual inspections. There is no light and water in the tunnel, so the inspection work is time-consuming and labor-intensive, with high inspection costs, large randomness of data, low degree of standardization, serious information silos, poor timeliness, and high safety risks. Missed inspections and blind spots are inevitable. The key is that the rapid changes in stress during the filling and draining process have a great impact on the structure inside the hydraulic tunnel, which may cause stress collapse of the concrete lining and other problems. Summary of the invention

[0004] The object of the present invention is to provide a method and computer equipment for detecting defects in a deep-buried long tunnel.

[0005] In a first aspect, an embodiment of the present invention provides a method for detecting defects in a deep-buried long tunnel, the method comprising:

[0006] Acquire image data and sonar data collected by the underwater inspection robot during operation in the tunnel;

[0007] generating fused detection data according to the image data and the sonar data;

[0008] Inputting the fused detection data into a pre-trained defect detection model to obtain a defect detection result;

[0009] Wherein, during the training process of the defect detection model, part of the sample data in the training set of the defect detection model is generated based on a sample data construction module, and the sample construction module is used to generate at least one second sample defect data based on defect example data information and first sample defect data.

[0010] In a second aspect, an embodiment of the present invention provides a computer device, comprising a processor and a non-volatile memory storing computer instructions, wherein when the computer instructions are executed by the processor, the computer device executes the deep-buried long tunnel defect detection method described in the first aspect.

[0011] Compared with the prior art, the beneficial effects provided by the present invention include: defect detection results can be obtained by acquiring image data and sonar data collected by an underwater inspection robot during tunnel operation, and the image data and sonar data are fused and input into a defect detection model; data can be collected by the inspection robot in a pressurized scenario, and automatic detection of tunnel defects can be achieved based on artificial intelligence technology, which can effectively avoid the high cost and low safety problems of manual detection; and, a sample data construction module is used to construct sample data based on defect example data information and first sample defect data to obtain at least one second sample defect data, thereby expanding the sample defect data on the basis of the first sample defect data to train a defect detection model with stronger detection capabilities, which can effectively reduce the labor time cost while ensuring the reliability of defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0013] Figure 1 A schematic diagram of a process for detecting defects in a deep-buried long tunnel provided by an embodiment of the present invention;

[0014] Figure 2 A structural block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, 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, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0016] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0018] In the description of the present invention, it should be understood that the terms "upper", "lower", "inside", "outside", "left", "right", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, or are directions or positional relationships in which the product of the application is usually placed when in use, or are directions or positional relationships commonly understood by those skilled in the art. These directions or positional relationships are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as a limitation on the present invention.

[0019] Furthermore, the terms “first”, “second”, etc. are merely used for distinguishing descriptions and are not to be understood as indicating or implying relative importance.

[0020] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms such as "setting" and "connection" should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0021] Figure 1 The flowchart of the deep buried long tunnel defect detection method provided by the embodiment of the present invention is as follows. Figure 1 As shown, the method comprises the following steps:

[0022] Step S210, obtaining image data and sonar data collected by the underwater inspection robot during operation in the tunnel.

[0023] In some embodiments, the image data and the sonar data may first be preprocessed, which may include, for example, denoising, contrast enhancement, etc., to improve data quality.

[0024] In some possible implementations, feature extraction may be performed on image data, such as edge detection, texture analysis, etc., and feature extraction may be performed on sonar data, such as echo intensity, signal arrival time, etc., to provide key information for subsequent data analysis and fusion.

[0025] It is worth noting that in a deep and long tunnel environment, due to the large diameter and dark light, it is difficult to achieve the detection accuracy requirements by image acquisition alone. Therefore, an acoustic and optical sensor can be used to collect tunnel inspection data. For example, a sonar device can be used for rough inspection, that is, sonar data can be collected, and then the image acquisition device can be used for corresponding image acquisition.

[0026] Step S220, generating fused detection data according to the image data and the sonar data.

[0027] The fusion of image data and sonar data can be achieved in a variety of ways, such as using a feature-level fusion method to combine the features of the two data sources. For example, different types of features extracted from image data and sonar data can be converted to the same scale or space, such as mapped to a common feature space, such as using principal component analysis technology to reduce the dimension of the features and then fuse them. Feature vectors from different sources can be combined into a single feature vector. This can be achieved by simple data concatenation, or using a more complex weighting method, where the importance of each feature can be determined by its contribution to the final task.

[0028] Step S230, inputting the fused detection data into a pre-trained defect detection model to obtain a defect detection result.

[0029] In which, during the training process of the defect detection model, part of the sample data in the training set of the defect detection model is generated based on the sample data construction module, and the sample construction module is used to generate at least one second sample defect data based on the defect example data information and the first sample defect data.

[0030] In some embodiments, the first sample defect data may be obtained by fusing the image data of a tunnel with defects that are actually collected and the sonar data. For example, by collecting the inspection results of typical tunnels, the tunnel image data and sonar data corresponding to the surface defects of deep-buried long tunnels are collected, such as tunnel cracks, leakage, scouring, pockmarks, falling off, exposed tendons, calcification and other defects.

[0031] Among them, cracks are the most important defects on the surface of the water diversion tunnel. They are often much lower than the background in the image proportion, and have the characteristics of slender shape, diverse structure and bifurcation. Pockmarking is the phenomenon that the surface of the water diversion tunnel concrete is wrapped with small bubbles, showing dense small pits and small pits. When the local area accounts for less than 1% and the depth does not exceed 1cm, it is not harmful. In a dim environment, it does not affect the appearance. It is an early disease of the water diversion tunnel. This phenomenon is mainly related to the early construction conditions. Shedding is when the surface protective layer of the water diversion tunnel is lost and the coarse aggregate with poor stability is directly exposed to the air, forming a defect caused by shedding. The area proportion is often intuitively visible. This is mainly due to the turbidity of the water quality in the tunnel environment, which is corrosive and the displacement of the surrounding rock stress. Exposed reinforcement is the phenomenon of the exposed steel structure in the lining structure of the water diversion tunnel. It may be a legacy problem caused by the constraints of construction conditions during the construction period, or it may be due to the dislocation of steel bars caused by geological changes during the operation of the tunnel. And so on.

[0032] By using sonar data and image data, the characteristics of the above defects can be analyzed, and then it can be detected whether the above defects exist in the tunnel.

[0033] In an embodiment of the present invention, defect detection results can be obtained by acquiring image data and sonar data collected by an underwater inspection robot during tunnel operation, and fusing the image data and sonar data and inputting them into a defect detection model. Data can be collected by the inspection robot in a pressurized scenario, and automatic detection of tunnel defects can be achieved based on artificial intelligence technology, which can effectively avoid the high cost and low safety problems of manual detection. In addition, sample data construction is performed based on defect example data information and first sample defect data using a sample data construction module to obtain at least one second sample defect data. Therefore, on the basis of the first sample defect data, the sample defect data can be expanded to train a defect detection model with stronger detection capabilities, which can effectively reduce labor time costs while ensuring the reliability of defect detection.

[0034] In some embodiments, the training process of the defect detection model may include but is not limited to steps 310 to 340 .

[0035] Step 310 , obtaining first sample defect data and defect data example information corresponding to the first sample defect data, wherein the defect data example information is used to guide the construction of second sample defect data having the same defect type as the first sample defect data.

[0036] In some optional embodiments, the first sample defect data may be obtained from a pre-built sample defect database. In the sample defect database, multiple types of first sample defect data may be stored. The first sample defect data may be obtained by fusing the actually collected tunnel image data with defects and the sonar data.

[0037] In some optional embodiments, the defect data example information is guidance information for instructing the construction of sample data of a unified defect type. By regenerating the first sample defect data based on the defect data example information, at least one second sample defect data that can be used to iterate the parameters of the initial defect detection model can be obtained, which is conducive to realizing the few-sample learning of the initial defect detection model and improving the detection capability and robustness of the trained initial defect detection model (i.e., the updated defect detection model).

[0038] In some optional embodiments, in the process of obtaining defect data example information, at least one candidate defect data information can be first obtained, and a corresponding defect feature representation can be generated for each candidate defect data information; then, for each candidate defect data information, type classification is performed based on the defect feature representation to obtain type distribution information of each candidate defect data information; then, for each candidate defect data information, in response to determining that the type distribution information meets a preset condition, the candidate defect data information is determined as defect data example information.

[0039] In some optional embodiments, the defect feature representation refers to the feature representation that the candidate defect data information can play a guiding role in. For example, if a candidate defect data information can guide the construction of sample data of the same defect type, then the defect feature representation of the candidate defect data information can be "features that can guide the construction of unified defect data".

[0040] In some optional embodiments, in the process of generating corresponding defect feature representations for candidate defect data information, a feature representation of the candidate defect data information can be first established based on image data feature recognition to obtain a first defect representation; and a feature representation of the candidate defect data information can be established based on sonar data feature recognition to obtain a second defect representation; then a first confidence level corresponding to the first defect representation and a second confidence level corresponding to the second defect representation are obtained; and then the feature representation corresponding to the one with a higher level indicator among the first confidence level and the second confidence level is determined as the defect feature representation.

[0041] In some optional embodiments, in the process of establishing a first defect representation based on a feature representation of candidate defect data information according to image data feature recognition, the candidate defect data information can be preprocessed first, and image data feature recognition can be performed on the preprocessed candidate defect data information to capture image features of the candidate defect data information; then, a corresponding first defect representation is generated for the candidate defect data information based on the acquired image features.

[0042] In some optional embodiments, establishing a feature representation of the candidate defect data information based on sonar data feature recognition means obtaining a feature representation of the candidate defect data information by identifying the sonar data feature recognition in the candidate defect data information. In the process of establishing a feature representation of the candidate defect data information based on sonar data feature recognition to obtain a second defect representation, the candidate defect data information may first be subjected to sonar data feature recognition extraction; then, a set of feature representation sets may be generated based on at least one extracted sonar data feature recognition, the feature representation set including at least one candidate feature representation, and the candidate feature representations correspond one-to-one to the sonar data features; then, the candidate defect data information is distance matched with each candidate feature representation, and the candidate feature representation with the largest distance is determined as the second defect representation of the candidate defect data information.

[0043] In some optional embodiments, in order to improve the accuracy of the feature representation of the candidate defect data information, after obtaining the first defect representation and the second defect representation, the first trust level corresponding to the first defect representation and the second trust level corresponding to the second defect representation can also be obtained, and then the first trust level and the second trust level are compared, and the feature representation corresponding to the trust level with a higher level index is determined as the defect feature representation of the candidate defect data information. Wherein, when obtaining the first trust level corresponding to the first defect representation and the second trust level corresponding to the second defect representation, the pre-trained first feature representation establishment module can be used to establish the feature representation of the candidate defect data information based on the image data feature recognition, and the pre-trained second feature representation establishment module can be used to establish the feature representation of the candidate defect data information based on the sonar data feature recognition extraction. At this time, the first feature representation establishment module will output the first defect representation of the candidate defect data information and the first trust level corresponding to the first defect representation, and the second feature representation establishment module will output the second defect representation of the candidate defect data information and the second trust level corresponding to the second defect representation, so the trust levels of the first defect representation and the second defect representation can be obtained in the output results of the first feature representation establishment module and the second feature representation establishment module. Among them, the pre-trained first feature representation establishment module and the second feature representation establishment module can both be composed of a conventional deep neural network model or a convolutional neural network model.

[0044] In some optional embodiments, when the candidate defect data information is classified into types based on the defect feature representation, each defect feature representation can be first tensorized to obtain a feature representation tensor corresponding to each defect feature representation; then, based on the distance between different feature representation tensors, all defect feature representations are clustered to obtain at least one feature representation cluster; then, based on the feature representation cluster in which the defect feature representation is located, the candidate defect data information corresponding to the defect feature representation is classified into types to obtain type distribution information. This processing can effectively improve the accuracy of the classification of candidate defect data information, thereby obtaining more accurate defect data example information that meets preset conditions.

[0045] In some optional embodiments, when tensorizing the defect feature representation to obtain the feature representation tensor corresponding to the defect feature representation, the defect feature representation can be firstly data mapped, for example, by mapping the defect feature representation to a unique identifier or index through integer encoding or one-hot encoding, to obtain the feature representation code corresponding to the defect feature representation; then, for the feature representation code, the corresponding feature representation tensor is obtained through a word embedding model query, that is, the feature representation tensor corresponding to the defect feature representation is obtained. This processing can conveniently obtain the feature representation tensor of the defect feature representation, thereby improving the efficiency of classifying the candidate defect data information according to the feature representation tensor.

[0046] In some optional embodiments, after obtaining the feature representation tensor of the defect feature representation, cluster analysis can be performed on all defect feature representations based on the distances between different feature representation tensors. Specifically, first, for each feature representation tensor, the cosine distance, Euclidean distance or other distance measurement methods can be used to determine the distance between it and other feature representation tensors; then the distances of all feature representation tensors are combined into a distance matrix, and each element in this distance matrix represents the similarity between two feature representation tensors; then, a cluster analysis algorithm is used to determine the distance matrix as input data for cluster analysis to obtain different feature representation clusters, and each defect feature representation is classified into a corresponding feature representation category, where each feature representation cluster represents a group of similar defect feature representations, and these similar defect feature representations are closer to each other in the vector space. This processing process can accurately perform cluster analysis on defect feature representations, which is conducive to obtaining more accurate defect data example information that meets preset requirements.

[0047] In some optional embodiments, after obtaining at least one candidate defect data information, these candidate defect data information can be first labeled with defect feature representations, and then a defect data example information library can be constructed based on the candidate defect data information corresponding to the defect feature representations that meet preset conditions.

[0048] In some optional embodiments, the defect data example information stored in the defect data example information library will all carry defect feature representations, so that when the sample data of the first sample defect data is subsequently constructed based on the defect data example information, the defect data example information can be directly randomly selected from the defect data example information library to construct the sample data of the first sample defect data, which can not only improve the convenience of obtaining the defect data example information, but also improve the richness of the second sample defect data obtained, thereby facilitating the use of the second sample defect data to improve the generalization ability and robustness of the model. In addition, the defect data example information library can be iteratively expanded to increase the diversity of the defect data example information.

[0049] Step 320 : Using a sample data construction module to construct sample data for the first sample defect data based on the defect data example information, to obtain at least one second sample defect data.

[0050] In some optional embodiments, the sample data construction module may be a pre-trained neural network model, which may extract features or information useful for defect detection from defect data example information.

[0051] In some optional embodiments, after obtaining the defect data example information and the first sample defect data, the defect data example information and the first sample defect data may be preprocessed to construct input data of the sample data construction module, and then the input data is input into the sample data construction module, so that the sample data construction module can perform sample data construction based on the input data to obtain at least one second sample defect data. Since the defect data example information can be used to guide the sample data construction module to generate sample defect data of the same defect type as the first sample defect data, the at least one second sample defect data output by the sample data construction module is of the same defect type as the first sample defect data.

[0052] In some optional embodiments, the process in which the sample data construction module constructs the sample data for the first sample defect data based on the defect data example information to obtain the second sample defect data may include but is not limited to steps 410 to 450 .

[0053] Step 410 , performing data mapping on the first sample defect data based on the defect data example information to obtain defect coding data and intermediate layer parameters.

[0054] In some optional embodiments, after receiving input data consisting of defect data example information and first sample defect data, the sample data construction module may perform data mapping on the input data to obtain defect coding data and intermediate layer parameters. The intermediate layer parameters may refer to the weights and bias items of the intermediate layer in the sample data construction module. The sample data construction module includes an input layer, an output layer and at least one intermediate layer, wherein the input layer is used to receive input data, the output layer is used to generate prediction results, and the intermediate layer is located between the input layer and the output layer, responsible for processing and learning data features. Therefore, after the sample data construction module performs data mapping on the input data to obtain defect coding data, it may obtain the intermediate layer parameters in the sample data construction module when the defect coding data is obtained. For example, the sample data construction module may include an encoding unit and a decoding unit, wherein the encoding unit is used to perform data mapping on the first sample defect data based on the defect data example information, and the decoding unit is used to generate the second sample defect data based on the data mapping result of the first sample defect data. Then the sample data construction module performs data mapping on the first sample defect data based on the defect data example information, which means that the encoding unit in the sample data construction module performs data mapping on the first sample defect data based on the defect data example information. After the encoding unit performs data mapping on the first sample defect data based on the defect data example information to obtain the defect coding data, the intermediate layer parameters can be obtained by obtaining the weights and bias items of the intermediate layers in the encoding unit when the encoding unit outputs the defect coding data. In one example, both the encoding unit and the decoding unit can have multiple layers. The output of the previous layer of encoding units will be determined as the input of the next layer of encoding units, and the output of the last layer of encoding units is the output of the entire encoding unit. The output of the previous layer of decoding units will be determined as the input of the next layer of decoding units, and the output of the last layer of decoding units is the output of the entire decoding unit. Among them, the number of layers of the encoding unit and the decoding unit can be appropriately selected based on the actual application situation, and is not specifically limited here. For example, the encoding unit and the decoding unit can both have 6 layers.

[0055] In some optional embodiments, in the process of performing data mapping on the first sample defect data based on the defect data example information to obtain defect coding data and intermediate layer parameters, the first sample defect data can be first embedded with information based on the defect data example information to obtain a first defect vector; and feature association measurement and data regeneration are performed on the first defect vector to obtain third process data; feature engineering is then performed on the third process data to obtain first feature data; and then data regeneration is performed on the first feature data based on the third process data to obtain defect coding data; and then the intermediate layer parameters of the sample data construction module when the defect coding data is obtained are obtained. Wherein, in the case where the sample data construction module includes an encoding unit and a decoding unit, obtaining the intermediate layer parameters of the sample data construction module when the defect coding data is obtained refers to obtaining the intermediate layer parameters of the intermediate layer in the encoding unit when the encoding unit outputs the defect coding data.

[0056] In some optional embodiments, information embedding is performed on the first sample defect data based on defect data example information to obtain a first defect vector, which can not only extract key features of the first sample defect data, but also reduce the dimension of the first sample defect data, thereby helping to reduce the complexity of data processing.

[0057] In some optional embodiments, after obtaining the first defect vector, a self-attention mechanism may be used to perform feature association measurement on the first defect vector to obtain an association measurement representation. For example, the similarity distance between the probe parameter and the alignment parameter may be determined first, and then a weight distribution may be obtained through a softmax function, and then the weight distribution may be applied to the data carrier to obtain an association measurement representation.

[0058] In some optional embodiments, after performing feature association measurement on the first defect vector to obtain an association measurement representation, the association measurement representation can be used to perform data regeneration on the first defect vector to obtain third process data. Wherein, when performing data regeneration on the first defect vector using the association measurement representation, the association measurement representation and the first defect vector can be first summed to obtain a sum result, and then the sum result can be normalized to obtain the third process data. The summation operation refers to the operation of adding the association measurement representation and the first defect vector. By performing a summation operation and a normalization process on the association measurement representation and the first defect vector, the purpose of performing data regeneration on the first defect vector using the association measurement representation can be achieved, so as to improve the accuracy of obtaining defect coding data.

[0059] In some optional embodiments, after obtaining the third process data, feature engineering can be performed on the third process data to obtain first feature data, and then the first feature data can be regenerated based on the third process data to obtain defect coding data. The first feature data obtained by performing feature engineering on the third process data can better capture various features and abstract representations of the first sample defect data, so that the sample data construction module can better understand the semantics of the first sample defect data and achieve better performance in subsequent processing. In addition, when regenerating the first feature data based on the third process data, the third process data and the first feature data can be summed to obtain the sum result, and then the sum result can be normalized to obtain defect coding data. The defect coding data is an intermediate representation obtained after a series of data processing of the first sample defect data, and the intermediate representation contains the conversion and encoding information of the first sample defect data, which can be used to capture the semantic information and feature information of the first sample defect data, so as to help the sample data construction module better understand and process the first sample defect data.

[0060] Step 420, perform feature association measurement and data regeneration on the defect coding data to obtain first process data.

[0061] In some optional embodiments, when performing feature association measurement and data regeneration on defect coded data, the self-attention mechanism described above can be used to perform feature association measurement on the defect coded data to obtain an association measurement representation of the defect coded data, and then the association measurement representation of the defect coded data can be used to perform data regeneration on the defect coded data to obtain first process data. Wherein, when performing data regeneration on the defect coded data using the association measurement representation of the defect coded data, the association measurement representation of the defect coded data and the defect coded data can be first summed to obtain a sum result, and then the sum result can be normalized to obtain the first process data. By summing and normalizing the association measurement representation of the defect coded data and the defect coded data, the purpose of performing data regeneration on the defect coded data using the association measurement representation of the defect coded data can be achieved, so as to improve the accuracy of obtaining the second sample defect data.

[0062] Step 430: perform data integration on the intermediate layer parameters and the first process data to obtain integrated data.

[0063] In some optional embodiments, when data integration is performed on the intermediate layer parameters and the first process data, a self-attention mechanism can be used to achieve data integration between the intermediate layer parameters and the first process data. Specifically, the self-attention mechanism can be used to perform feature association measurement on the intermediate layer parameters and the first process data to achieve data integration of the intermediate layer parameters and the first process data. Wherein, when performing feature association measurement, the probe parameters and the alignment parameters are both intermediate layer parameters, and the data carrier is the first process data, that is, the intermediate layer parameters are determined as probe parameters and alignment parameters, and the first process data is determined as the data carrier, and feature association measurement is performed to obtain integrated data to achieve data integration of the intermediate layer parameters and the first process data.

[0064] Step 440: regenerate the integrated data based on the first process data to obtain second process data.

[0065] In some optional embodiments, when the integrated data is regenerated based on the first process data, the first process data and the integrated data may be summed to obtain a summed result, and then the summed result may be normalized to obtain the second process data. By summing and normalizing the first process data and the integrated data, the purpose of regenerating the integrated data using the first process data may be achieved, so as to improve the accuracy of obtaining the second sample defect data.

[0066] Step 450 : obtaining second sample defect data based on the second process data.

[0067] In some optional embodiments, in order to obtain the second sample defect data that is more in line with the defect data example information, feature engineering can be performed on the second process data to obtain the second feature data, and then the second feature data can be predicted by type classification to obtain the second sample defect data. The second feature data can better capture the various features and abstract representations of the defect coding data, so that the sample data construction module can better understand the semantics of the defect coding data, and then can more accurately decode the defect coding data to obtain the second sample defect data.

[0068] In some optional embodiments, the sample data construction module may include an encoding unit and a decoding unit. In the case where the sample data construction module includes an encoding unit and a decoding unit, the step 410 described above may be performed in the encoding unit, and the steps 420 to 450 described above may be performed in the decoding unit. Among them, the decoding unit may include a feedforward network layer and a softmax type divider. The feedforward network layer can be used to perform feature engineering on the second process data to obtain second feature data, and the softmax type divider can be used to perform type division prediction on the second feature data to obtain second sample defect data.

[0069] In some optional embodiments, after information embedding is performed on the first sample defect data based on the defect data example information to obtain the first defect vector, the first defect vector may be subjected to feature association measurement to obtain an association measurement representation, and then the association measurement representation and the first defect vector are summed and normalized to obtain third process data, and then feature engineering is performed on the third process data to obtain first feature data, and then the first feature data and the third process data are summed and normalized to obtain defect coded data. In addition, by obtaining the weights and bias items of the intermediate layer in the encoding unit when the encoding unit outputs the defect coded data, the intermediate layer parameters may be obtained.

[0070] After obtaining the defect coding data, the defect coding data can be subjected to feature association measurement to obtain an association measurement representation of the defect coding data, and then the defect coding data and the association measurement of the defect coding data can be summed and normalized to obtain the first process data, and then the first process data and the intermediate layer parameters can be subjected to feature association measurement to obtain integrated data, and then the first process data and the integrated data can be summed and normalized to obtain the second process data, and then the second process data can be input into the feedforward network layer of the decoding unit to perform feature engineering to obtain the second feature data, and then the second feature data can be input into the softmax type divider of the decoding unit to perform type classification prediction on the second feature data to obtain the second sample defect data.

[0071] In some optional embodiments, by inputting the first sample defect data into the sample data construction module for sample data construction, at least one second sample defect data having the same defect type as the first sample defect data can be obtained, so that on the basis of the first sample defect data, the sample defect data having the same defect type can be expanded to achieve data regeneration of the first sample defect data.

[0072] Step 330 , performing parameter iteration on the initial defect detection model based on at least one second sample defect data to obtain an updated defect detection model.

[0073] In some optional embodiments, before performing parameter iteration on the initial defect detection model based on at least one second sample defect data, the initial defect detection model may be pre-trained using the first sample defect data so that the initial defect detection model has basic defect detection capabilities.

[0074] In some optional embodiments, after the initial defect detection model is pre-trained so that the initial defect detection model has basic task processing capabilities, since the second sample defect data is obtained by regenerating the first sample defect data based on the defect data example information, and the defect type of the second sample defect data is consistent with that of the first sample defect data, after the initial defect detection model is iterated with parameters (i.e., fine-tuned training) using at least one second sample defect data, an updated defect detection model with better model effect and robustness can be obtained. In other words, the initial defect detection model will first go through a training phase, and then fine-tune the initial defect detection model based on at least one second sample defect data in an adjustment phase to obtain an updated defect detection model with better model effect and robustness.

[0075] In some optional embodiments, the first sample defect data can be first determined as the first training set to pre-train the initial defect detection model to obtain an initial defect detection model with basic defect detection capabilities; then the sample data construction module is used to regenerate the first training set to obtain a second training set (i.e., a training set composed of at least one second sample defect data), wherein the defect types of the first training set and the second training set are unified; then the initial defect detection model is fine-tuned based on the second training set to improve the model effect and robustness of the initial defect detection model.

[0076] In some optional embodiments, when performing parameter iteration on the initial defect detection model based on at least one second sample defect data, the at least one second sample defect data and the first sample defect data can be first merged together to determine a sample defect data set, and then the initial defect detection model can be fine-tuned and trained based on the sample defect data set to obtain an updated defect detection model with better generalization ability and robustness.

[0077] Step 340 , verifying the detection capability of the updated defect detection model to obtain detection capability indication information, and updating the initial defect detection model according to the detection capability indication information.

[0078] In some optional embodiments, after performing parameter iteration on the initial defect detection model based on at least one second sample defect data to obtain an updated defect detection model, the detection capability of the updated defect detection model can be verified, and the initial defect detection model can be updated according to the detection capability indication information.

[0079] In some optional embodiments, when the initial defect detection model is updated according to the detection capability indication information, in response to determining that the detection capability indication information indicates that the defect detection capability of the updated defect detection model is enhanced, the initial defect detection model can be updated to the updated defect detection model; in response to determining that the detection capability indication information indicates that the defect detection capability of the updated defect detection model is not enhanced, the initial defect detection model is not updated to the updated defect detection model. In other words, it is possible to determine whether the defect detection capability of the updated defect detection model is enhanced according to the detection capability indication information, and only when it is determined that the defect detection capability of the updated defect detection model is enhanced, the initial defect detection model is updated to the updated defect detection model, thereby ensuring that the update iterations of the initial defect detection model are positively effective, thereby avoiding the problem of unstable model training effectiveness in the process of iterating parameters of the initial defect detection model. In addition, since the model effect of the updated defect detection model has been optimized, it means that the current at least one second sample defect data can have a positive training effectiveness on the initial defect detection model, that is, the current at least one second sample defect data can be determined as a high-quality model training sample. Therefore, the current at least one second sample defect data can be saved in the sample defect database so that these second sample defect data can be used for other training samples of the same type of defect detection.

[0080] In some optional embodiments, after the detection capability of the updated defect detection model is verified, in response to determining that the detection capability indication information indicates that the defect detection capability of the updated defect detection model has not been enhanced, it means that the current at least one second sample defect data does not have a positive training effectiveness on the initial defect detection model, that is, the current at least one second sample defect data is not a high-quality model training sample. Therefore, the updating of the defect detection model can be abandoned, the initial defect detection model is not updated to the updated defect detection model, and the current at least one second sample defect data is not added to the sample defect database.

[0081] The process of verifying the detection capability of the updated defect detection model to obtain the detection capability indication information may include the following steps 510 to 530 .

[0082] Step 510: Acquire sample verification defect data and verification reference information corresponding to the sample verification defect data.

[0083] In some optional embodiments, the sample verification defect data refers to the sample data used when testing the model performance of the updated defect detection model. These sample data are data that are not used for parameter iteration of the initial defect detection model, and thus can be used to evaluate the performance of the updated defect detection model on new data.

[0084] In some optional embodiments, the verification reference information corresponding to the sample verification defect data may be a defect type label corresponding to the sample verification defect data, or may be a second verification defect data having the same defect type as the sample verification defect data. It may be appropriately selected based on different detection capability verification methods for the updated defect detection model, and is not limited here.

[0085] Step 520 , using the updated defect detection model to perform defect detection on the sample verification defect data to obtain a first detection result.

[0086] Step 530 , based on the first detection result and the verification reference information, obtain detection capability indication information for verifying the detection capability of the updated defect detection model.

[0087] In some optional embodiments, when the verification reference information is a defect type label corresponding to the sample verification defect data, in the process of obtaining the detection capability indication information for verifying the detection capability of the updated defect detection model based on the first detection result and the verification reference information, the iterative loss parameter between the first detection result and the defect type label can be first determined, and then the iterative loss parameter can be compared with the loss parameter threshold to obtain the first indication information; in response to determining that the first indication information is used to indicate that the iterative loss parameter is not higher than the loss parameter threshold, the detection capability indication information of the updated defect detection model with enhanced defect detection capability can be obtained based on the first indication information; in response to determining that the first indication information is used to indicate that the iterative loss parameter is higher than the loss parameter threshold, the detection capability indication information of the updated defect detection model with non-enhanced defect detection capability can be obtained based on the first indication information. It should be noted that the value of the loss parameter threshold can be appropriately selected based on the actual application situation, and is not specifically limited here.

[0088] In some optional embodiments, when the verification reference information is the second verification defect data of the same defect type as the sample verification defect data, in the process of obtaining the detection capability indication information for verifying the detection capability of the updated defect detection model based on the first detection result and the verification reference information, the distance between the first detection result and the second verification defect data can be first determined, and the distance can be compared with the distance threshold parameter to obtain the second indication information. In response to determining that the second indication information is used to indicate that the distance is not higher than the distance threshold parameter, the detection capability indication information of the updated defect detection model that the defect detection capability is not enhanced can be obtained based on the second indication information; in response to determining that the second indication information is used to indicate that the distance is higher than the distance threshold parameter, the detection capability indication information of the updated defect detection model that the defect detection capability is enhanced can be obtained based on the second indication information. It should be noted that the value of the distance threshold parameter can be determined based on the specific application and requirements, and is not specifically limited here.

[0089] Through the model training method of steps 310 to 340, after obtaining the first sample defect data and the defect data example information corresponding to the first sample defect data, the sample data construction module is used to guide the construction of the second sample defect data with the same defect type as the first sample defect data based on the defect data example information, and the first sample defect data is sampled to construct the defect data, and at least one second sample defect data is obtained, thereby expanding the sample defect data with the same defect type on the basis of the first sample defect data, and realizing the data regeneration of the first sample defect data. In addition, since the second sample defect data is obtained through the sample data construction module, it does not need to be manually collected and marked, thereby effectively reducing the time-consuming labor cost. After obtaining at least one second sample defect data, the initial defect detection model is firstly iterated based on the at least one second sample defect data to obtain an updated defect detection model. Since the second sample defect data is obtained based on the data regeneration of the first sample defect data, the few-sample learning of the initial defect detection model can be realized; then, the detection capability of the updated defect detection model is verified to obtain the detection capability indication information. In response to the determination that the detection capability indication information indicates that the acquired defect detection capability of the updated defect detection model is enhanced, it is indicated that the second sample defect data has played a positive training effectiveness on the initial defect detection model, and can improve the model effect and robustness of the initial defect detection model. Therefore, the initial defect detection model can be updated to the updated defect detection model, thereby obtaining a model with better defect detection effect and stronger robustness. In addition, since the initial defect detection model is updated to the updated defect detection model only when the acquired defect detection capability of the updated defect detection model is enhanced, each model update can be positively effective, thereby avoiding the problem of unstable model training effectiveness.

[0090] For example, after executing the above process, at least one second sample defect data can be obtained. At this time, when the user uses at least one second sample defect data to iterate the parameters of the initial defect detection model, the at least one second sample defect data can be formed into a sample set, and then the sample set and the first sample defect data pre-stored in the sample defect database are determined as training samples of the initial defect detection model, and the training samples are input into the initial defect detection model, so as to fine-tune the initial defect detection model and obtain an updated defect detection model. After obtaining the updated defect detection model, the detection capability of the updated defect detection model can be verified using a test data set (i.e., sample verification defect data) and a verification data set (i.e., verification reference information corresponding to the sample verification defect data). In the process of verifying the detection capability of the updated defect detection model using the sample verification defect data and the verification reference information, the sample verification defect data can be first input into the updated defect detection model for defect detection to obtain a first detection result, and then the detection capability indication information for verifying the detection capability of the updated defect detection model can be obtained based on the first detection result and the verification reference information. In one example, the defect type label corresponding to the sample verification defect data with the verification reference information is used as an example for explanation. In the process of obtaining the detection capability indication information for verifying the detection capability of the updated defect detection model based on the first detection result and the verification reference information, the iterative loss parameter between the first detection result and the verification reference information can be first determined, and the iterative loss parameter can be compared with the loss parameter threshold. When the iterative loss parameter is not higher than the loss parameter threshold, it means that the model effect of the updated defect detection model has been optimized, indicating that the second sample defect data can play a positive role in the training effectiveness of the initial defect detection model, and can enhance the model effect and robustness of the initial defect detection model. At this time, this batch of second sample defect data can be saved in the sample defect database, and the initial defect detection model can be updated to the updated defect detection model. When the iterative loss parameter is higher than the loss parameter threshold, it means that the defect detection capability of the updated defect detection model has not been enhanced, indicating that the second sample defect data has failed to have a positive training effectiveness on the initial defect detection model, and has failed to enhance the model effect and robustness of the initial defect detection model. At this time, this batch of second sample defect data can be abandoned, that is, this batch of second sample defect data will not be saved in the sample defect database, and the initial defect detection model will remain unchanged.

[0091] In one embodiment, the method may specifically include but is not limited to steps 601 to 606.

[0092] Step 601, obtain first sample defect data and at least one candidate defect data information, establish feature representations for all candidate defect data information based on image data feature recognition to obtain a first defect representation, and establish feature representations for all candidate defect data information based on sonar data feature recognition to obtain a second defect representation.

[0093] Step 602, obtaining a first confidence level corresponding to the first defect representation and a second confidence level corresponding to the second defect representation, and determining a feature representation corresponding to the one with a higher level index of the first confidence level and the second confidence level as the defect feature representation of the candidate defect data information.

[0094] Step 603 , quantize each defect feature representation tensor to obtain a feature representation tensor corresponding to each defect feature representation, and perform cluster analysis on all defect feature representations based on the distances between different feature representation tensors to obtain at least one feature representation cluster.

[0095] Step 604, based on the feature representation cluster in which the defect feature representation of the candidate defect data information is located, obtain type distribution information of the candidate defect data information, and for each candidate defect data information, in response to determining that the type distribution information satisfies a preset condition, determine the candidate defect data information as defect data example information.

[0096] Step 605 , embedding information into the first sample defect data based on the defect data example information to obtain a first defect vector.

[0097] Step 606 , perform feature association measurement and data regeneration on the first defect vector to obtain third process data, and perform feature engineering on the third process data to obtain first feature data.

[0098] Step 607 , regenerating the first characteristic data based on the third process data to obtain defect coding data, and acquiring the intermediate layer parameters of the sample data construction module when obtaining the defect coding data.

[0099] Step 608, perform feature association measurement and data regeneration on the defect coding data to obtain first process data.

[0100] Step 609, determining the intermediate layer parameters as the probe parameters and alignment parameters in the feature association measurement, and determining the first process data as the data carrier in the feature association measurement, performing the feature association measurement, and obtaining integrated data.

[0101] Step 610: regenerate the integrated data based on the first process data to obtain second process data, and perform feature engineering on the second process data to obtain second feature data.

[0102] Step 611, perform type classification prediction on the second feature data to obtain second sample defect data.

[0103] Step 612: perform parameter iteration on the initial defect detection model based on at least one second sample defect data to obtain an updated defect detection model.

[0104] Step 613: Acquire sample verification defect data and verification reference information corresponding to the sample verification defect data.

[0105] Step 614 , using the updated defect detection model to perform defect detection on the sample verification defect data to obtain a first detection result.

[0106] Step 615, when the verification reference information is the defect type label corresponding to the sample verification defect data, determine the iterative loss parameter between the first detection result and the defect type label, and compare the iterative loss parameter with the loss parameter threshold to obtain first indication information; in response to determining that the first indication information is used to indicate that the iterative loss parameter is not higher than the loss parameter threshold, based on the first indication information, obtain detection capability indication information of the updated defect detection model with enhanced defect detection capability; in response to determining that the first indication information is used to indicate that the iterative loss parameter is higher than the loss parameter threshold, based on the first indication information, obtain detection capability indication information of the updated defect detection model with non-enhanced defect detection capability.

[0107] Step 616, updating the initial defect detection model according to the detection capability indication information.

[0108] Through the above method of step 601 to step 616, after obtaining the first sample defect data and the defect data example information corresponding to the first sample defect data, the sample data construction module is used to construct the first sample defect data based on the defect data example information to guide the construction of the second sample defect data with the same defect type as the first sample defect data, and at least one second sample defect data is obtained. Therefore, on the basis of the first sample defect data, the sample defect data with the same defect type can be expanded to achieve data regeneration for the first sample defect data. In addition, since the second sample defect data is obtained through the sample data construction module, it is not necessary to collect and mark it manually, so that the time-consuming cost of manual work can be effectively reduced. After obtaining at least one second sample defect data, the initial defect detection model is firstly iterated based on the at least one second sample defect data to obtain an updated defect detection model. Since the second sample defect data is obtained based on the data regeneration of the first sample defect data, the few-sample learning of the initial defect detection model can be achieved; then, the detection capability of the updated defect detection model is verified to obtain the detection capability indication information, and the initial defect detection model is updated according to the detection capability indication information. Since the update of the initial defect detection model needs to be performed based on the detection capability indication information, there will not be a situation where the task model is automatically iterated and updated regardless of whether it is optimized, thereby avoiding the problem of unstable model training effectiveness. For example, when the detection capability indication information indicates that the defect detection capability of the updated defect detection model is enhanced, it means that the second sample defect data has a positive training effectiveness on the initial defect detection model, which can improve the model effect and robustness of the initial defect detection model. At this time, the initial defect detection model can be updated to an updated defect detection model, thereby obtaining a model with better model effect and stronger robustness. Since the initial defect detection model can be updated to an updated defect detection model only when the defect detection capability of the updated defect detection model is enhanced, each model update can be made positive and effective, thereby effectively avoiding the problem of unstable model training effectiveness.

[0109] It is to be understood that the steps involved in the above embodiments are not necessarily performed in sequence. Unless clearly stated in the present embodiment, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the above steps may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0110] The embodiment of the present invention further provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned deep buried long tunnel defect detection method. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111 , a processor 112 , and a communication unit 113 .

[0111] To achieve data transmission or interaction, the memory 111, the processor 112, and the communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines. The processor 112 is used to execute computer instructions stored in the memory 111 to implement the deep buried long tunnel defect detection method involved in the above embodiment.

[0112] In one embodiment, the memory 111 is a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor 112, the computer device executes the steps of the deep buried long tunnel defect detection method involved in the above embodiment.

[0113] An embodiment of the present invention provides a readable storage medium, which includes a computer program. When the computer program is running, it controls the computer device where the readable storage medium is located to execute the aforementioned deep-buried long tunnel defect detection method.

[0114] For the purpose of illustration, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. In accordance with the above teachings, numerous modifications and variations are possible. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application. For the purpose of illustration, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. In accordance with the above teachings, numerous modifications and variations are possible. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A method for detecting defects in a deep buried long tunnel, characterized in that: The method comprises: Obtaining image data and sonar data collected by the underwater inspection robot during operation in the tunnel; generating fused detection data according to the image data and the sonar data; Inputting the fused detection data into a pre-trained defect detection model to obtain a defect detection result; Wherein, during the training process of the defect detection model, part of the sample data in the training set of the defect detection model is generated based on a sample data construction module, and the sample construction module is used to generate at least one second sample defect data based on defect example data information and first sample defect data.

2. The method according to claim 1, characterized in that The training of the defect detection model includes: Acquire first sample defect data and defect data example information corresponding to the first sample defect data, wherein the defect data example information is used to guide the construction of second sample defect data with the same defect type as the first sample defect data; Using a sample data construction module to construct sample data for the first sample defect data based on the defect data example information, to obtain at least one second sample defect data; Iterate the parameters of the initial defect detection model based on at least one of the second sample defect data to obtain an updated defect detection model; Verifying the detection capability of the updated defect detection model to obtain detection capability indication information; The initial defect detection model is updated according to the detection capability indication information.

3. The method according to claim 2, characterized in that The verifying the detection capability of the updated defect detection model to obtain the detection capability indication information includes: Acquire sample verification defect data and verification reference information corresponding to the sample verification defect data; Using the updated defect detection model to perform defect detection on the sample verification defect data to obtain a first detection result; Based on the first detection result and the verification reference information, detection capability indication information for verifying the detection capability of the updated defect detection model is obtained.

4. The method according to claim 3, characterized in that The verification reference information is a defect type label corresponding to the sample verification defect data; The obtaining, based on the first detection result and the verification reference information, detection capability indication information for verifying the detection capability of the updated defect detection model includes: determining an iterative loss parameter between the first detection result and the defect type label, and comparing the iterative loss parameter with a loss parameter threshold to obtain first indication information; In response to determining that the first indication information is used to indicate that the iterative loss parameter is not higher than the loss parameter threshold, obtaining detection capability indication information of the updated defect detection model with enhanced defect detection capability based on the first indication information; In response to determining that the first indication information is used to indicate that the iterative loss parameter is higher than the loss parameter threshold, detection capability indication information indicating that defect detection capability of the updated defect detection model is not enhanced is obtained based on the first indication information.

5. The method according to claim 3, characterized in that: The verification reference information is second verification defect data with the same defect type as the sample verification defect data; The obtaining, based on the first detection result and the verification reference information, detection capability indication information for verifying the detection capability of the updated defect detection model includes: Determine a distance between the first detection result and the second verification defect data, and compare the distance with a distance threshold parameter to obtain second indication information; In response to determining that the second indication information is used to indicate that the distance is not greater than the distance threshold parameter, obtaining, based on the second indication information, detection capability indication information indicating that the defect detection capability of the updated defect detection model is not enhanced; In response to determining that the second indication information is used to indicate that the distance is higher than the distance threshold parameter, detection capability indication information of the updated defect detection model with enhanced defect detection capability is obtained based on the second indication information.

6. The method according to claim 4 or 5, characterized in that: The updating of the initial defect detection model according to the detection capability indication information includes: In response to determining that the detection capability indication information is an obtained defect detection capability enhancement of the updated defect detection model, updating the initial defect detection model to the updated defect detection model; Alternatively, in response to determining that the detection capability indication information indicates that the acquired defect detection capability of the updated defect detection model is not enhanced, the initial defect detection model is not updated to the updated defect detection model.

7. The method according to claim 2, characterized in that The defect data example information is obtained based on the following steps: Acquire at least one candidate defect data information, and generate a corresponding defect feature representation for each candidate defect data information; For each of the candidate defect data information, classify the types based on the defect feature representation to obtain type distribution information of each of the candidate defect data information; For each of the candidate defect data information, in response to determining that the type distribution information satisfies a preset condition, the candidate defect data information is determined as the defect data example information.

8. The method according to claim 7, characterized in that The process of generating the defect feature representation for the candidate defect data information comprises the following steps: Establishing a feature representation of the candidate defect data information based on image data feature recognition to obtain a first defect representation; Establishing a feature representation of the candidate defect data information based on sonar data feature recognition to obtain a second defect representation; Obtaining a first confidence level corresponding to the first defect representation and a second confidence level corresponding to the second defect representation; Determine the feature representation corresponding to the one with a higher level index between the first confidence level and the second confidence level as the defect feature representation; The classifying the candidate defect data information into types based on the defect feature representation to obtain type distribution information of the candidate defect data information includes: Quantize each of the defect feature representations to obtain a feature representation tensor corresponding to each of the defect feature representations; Based on the distances between different feature representation tensors, cluster analysis is performed on all the defect feature representations to obtain at least one feature representation cluster; Based on the feature representation cluster in which the defect feature representation corresponding to the candidate defect data information is located, type distribution information of the candidate defect data information is obtained.

9. The method according to claim 2, characterized in that: The process of the sample data construction module constructing the sample data for the first sample defect data based on the defect data example information to obtain the second sample defect data includes the following steps: Performing data mapping on the first sample defect data based on the defect data example information to obtain defect coding data and intermediate layer parameters; Performing feature association measurement and data regeneration on the defect coding data to obtain first process data; Performing data integration on the intermediate layer parameters and the first process data to obtain integrated data; regenerating the integrated data based on the first process data to obtain second process data; obtaining the second sample defect data based on the second process data; The performing data mapping on the first sample defect data based on the defect data example information to obtain defect coding data and intermediate layer parameters includes: Based on the defect data example information, information is embedded into the first sample defect data to obtain a first defect vector; Performing feature association measurement and data regeneration on the first defect vector to obtain third process data; Performing feature engineering on the third process data to obtain first feature data; Regenerating the first characteristic data based on the third process data to obtain defect code data; Acquire the intermediate layer parameters of the sample data construction module when obtaining the defect coding data; The step of integrating the intermediate layer parameters and the first process data to obtain integrated data includes: Determine the intermediate layer parameters as probe parameters and alignment parameters in feature association measurement, determine the first process data as data carriers in feature association measurement, perform feature association measurement, and obtain integrated data; The obtaining the second sample defect data based on the second process data comprises: performing feature engineering on the second process data to obtain second feature data; Perform type classification prediction on the second feature data to obtain the second sample defect data.

10. A computer device, characterized in that: The computer device includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device executes the deep-buried long tunnel defect detection method described in any one of claims 1-9.

Images