A Transformer-based power battery weld point cloud defect detection method and system

Abstract

Based on the technical field of image processing, the present invention discloses a method and system for detecting point cloud defects of power battery welds based on Transformer, wherein the detection method comprises the following steps: collecting original point cloud data, and extracting features from the original point cloud data to obtain a point cloud feature matrix; inputting the point cloud feature matrix into a multi-layer attention network, dynamically allocating attention according to the similarity difference between points in the point cloud data, and obtaining multiple feature information; mapping the feature information to different classification directories to obtain type scores. The defect detection method proposed by the present invention has high accuracy and high stability of the detection results; and also improves the detection efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a Transformer-based power battery weld point cloud defect detection method and system. Background Art

[0002] As the energy storage unit of the whole vehicle, the safety and reliability of the battery pack of new energy vehicles are of vital importance. As the positive and negative interfaces of the battery, the poles are currently generally processed by riveting and laser welding. The defects in the laser welding process include burn-through of the weld surface, grooves, cold welding, broken welding, etc. The welding quality of the poles has become a limiting factor in the service life and energy safety of the battery pack, which may lead to problems such as interruption of the battery pack output power, abnormal temperature rise, shortened service life, etc., directly affecting the safety performance of the whole vehicle.

[0003] The industry can currently obtain 3D point cloud information of welds through various structured light sensors, including existing technologies such as laser projection, coded stripes and other structured light. The data structure of 3D point clouds is widely used in deep learning networks; however, current deep neural networks still have shortcomings in the following aspects: 1) Existing networks mainly start from the overall structure of point clouds and lack attention to local detail defects, which affects detection accuracy; 2) Transformer-based neural networks still need to be improved in terms of training convergence speed and stability.

[0004] In summary, it is necessary to design a Transformer-based power battery weld point cloud defect detection method and system to solve the above problems in the prior art. Summary of the invention

[0005] In order to solve the above-mentioned problems in the prior art, the present invention provides a Transformer-based power battery weld point cloud defect detection method and system, which solves the problem that the existing deep neural network lacks attention to local detail defect problems.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] A method for detecting point cloud defects of power battery welds based on Transformer, comprising the following steps:

[0008] Collecting original point cloud data, and performing feature extraction on the original point cloud data to obtain a point cloud feature matrix;

[0009] The point cloud feature matrix is ​​input into a multi-layer attention network, and attention is dynamically allocated according to the similarity difference between points in the point cloud data to obtain multiple feature information;

[0010] Map feature information to different classification directories to obtain type scores.

[0011] In some embodiments of the present invention, the point cloud feature matrix extraction process includes:

[0012] Determine the coordinates of adjacent points of the original point cloud data;

[0013] Subtract the coordinates of the center point from the coordinates of the neighboring points and copy and concatenate them with the K layers of the center point to obtain high-dimensional features;

[0014] After the high-dimensional features are processed by convolution and LBR layers, significant features are extracted from K neighboring points through maximum pooling to obtain the point cloud feature matrix.

[0015] In some embodiments of the present invention, determining the coordinates of neighboring points of the original point cloud data includes:

[0016] The farthest point is sampled for the original point cloud data; and the K neighboring points of each point are found through K-neighbor search in the original point cloud data;

[0017] The coordinates of the K neighboring points are subtracted from the coordinates of the corresponding sampling point to obtain the normalized coordinates of the neighboring points.

[0018] In some embodiments of the present invention, the dynamic allocation of attention according to the similarity difference between points in the point cloud data includes:

[0019] Dynamic similarity is calculated using the Line + BatchNorm algorithm;

[0020] Calculate dynamic similarity attention and output multiple features based on dynamic similarity.

[0021] In some embodiments of the present invention, the calculation formula of the dynamic similarity is:

[0022] DS = LB(Similarity(R(Fi)));

[0023] Among them, Fi represents the input feature; R represents the copy instruction; the similarity in the high-dimensional feature matrix is ​​calculated by matrix dot product, and each element in the obtained similarity matrix corresponds to the similarity between every two elements in the feature matrix; LB is the Line + BatchNorm operation.

[0024] In some embodiments of the present invention, the calculation formula of the dynamic similarity attention is:

[0025] DSA=LB(Similarity(R(Fi)))·SL(Q·K T );

[0026] The output formula of the multiple features is:

[0027] Fout=DSA·V+Fi;

[0028] Among them, SL represents the Softmax+L1Norm algorithm; Fi represents the input feature; Q, K, and V correspond to the query, key, and value in the Transformer network framework respectively, and the calculation formula is as follows:

[0029] (Q,K,V)=Fi·(W q ,W k ,W v );

[0030] Where W q , W k , W v Represents a shared linear layer trainable matrix.

[0031] In some embodiments of the present invention, a power battery weld point cloud defect detection system based on Transformer is provided, comprising:

[0032] Feature embedding module, used to extract and aggregate features after high-dimensional mapping of raw point cloud data;

[0033] The dynamic similarity attention module is used to import the high-dimensional features output by the feature embedding module as input into the attention network based on dynamic similarity to redistribute the attention weights and extract multiple feature information of the point cloud;

[0034] The classification module is used to map the obtained high-dimensional feature information into different classification directories and obtain and output type scores.

[0035] In some embodiments of the present invention, the classification module includes three linear layers, which are used to gradually map the features extracted and learned by the dynamic similarity attention module to the sample annotation space.

[0036] In some embodiments of the present invention, there is provided an electronic device, comprising:

[0037] A processor, and a memory and a transceiver communicatively connected to the processor;

[0038] The memory stores computer-executable instructions; the transceiver is used to send and receive data;

[0039] The processor executes the computer-executable instructions stored in the memory to implement the above-mentioned calculation method.

[0040] In some embodiments of the present invention, a computer-readable storage medium is provided, characterized in that:

[0041] The computer-readable storage medium stores computer-executable instructions, which are used to implement the above-mentioned calculation method when executed by a processor.

[0042] The technical solution of the present invention has the following technical effects compared with the prior art:

[0043] The defect detection method proposed by the present invention has high accuracy and high stability of the detection result; at the same time, the detection efficiency is also improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0045] Figure 1 Schematic diagram of the structure of a Transformer-based power battery weld point cloud defect detection system shown in the embodiment.

[0046] Figure 2 Schematic diagram of the processing flow of the feature embedding module shown in the embodiment.

[0047] Figure 3 Schematic diagram of the processing flow of the dynamic similarity attention module shown in the embodiment.

[0048] Figure 4 Schematic diagram of the accuracy of the training process of each point cloud module.

[0049] Figure 5 Schematic diagram of the accuracy and loss value of the test set of the model shown in the embodiment.

[0050] Figure 6 FIG. 1 is a schematic diagram of the structure of an electronic device shown in the embodiment.

[0051] Figure numerals: 110, feature embedding module; 120, dynamic similarity attention module; 130, classification module; 200, electronic device; 210, processor; 220, memory; 230, transceiver. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] In the description of the present application, it should be understood that the terms "center", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0054] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0055] In the description of this application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0056] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0057] The disclosure below provides many different embodiments or examples to implement different structures of the present invention. In order to simplify the disclosure of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the present invention. In addition, the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed.

[0058] Embodiment 1: A method for detecting point cloud defects of a power battery weld based on Transformer, comprising the following steps:

[0059] Step S1: collecting original point cloud data, and performing feature extraction on the original point cloud data to obtain a point cloud feature matrix;

[0060] Specifically, refer to Figure 2 As shown, we first use a 1×1 convolution layer to transform the original point cloud from N in ×3 dimensions increased to N in ×d dimension, the convolution layer can interact the features between different channels and improve the feature extraction capability. Then the LBR layer is used to further extract high-dimensional features and normalize them to prevent overfitting.

[0061] In order to take into account the local features of the point cloud, the input feature data is first sampled at the farthest point (FPS), and the K nearest neighbor search (KNN) is used to find the K nearest neighbor points of each point in the original point cloud. The coordinates of these neighboring points are subtracted from the coordinates of the corresponding sampling point to obtain the normalized coordinates of the neighboring points. Then, these coordinates are subtracted from the coordinates of the center point and connected with the K layers of the center point to obtain N out ×2K×d dimensional features. After being processed by the convolution and LBR layers, in order to ensure the permutation invariance of the point cloud, the most significant features are extracted from the K neighboring points through the maximum pooling operation, and finally a new point cloud feature matrix is ​​obtained.

[0062] Step S2, inputting the point cloud feature matrix into a multi-layer attention network, dynamically allocating attention according to the similarity difference between points in the point cloud data, and obtaining multiple feature information;

[0063] The input features of the multi-layer attention network are recorded as Fi, and W is used q , W k , W v Represents the shared linear layer trainable matrix, the input features Fi and W q , W k , W vThe matrices are multiplied to get Q, K, and V, which correspond to the query, key, and value in the Transformer network framework. The formula is as follows:

[0064] (Q,K,V)=Fi·(W q ,W k ,W v );

[0065] The dynamic similarity-attention (DSA) mechanism is adopted. DSA implements a non-uniform attention mechanism. In the point cloud feature space after feature extraction, the attention scores are weighted according to the similarity differences between high-dimensional points. Therefore, DSA can accurately distinguish different parts in the point cloud. For the battery welding defect judgment task, different weights need to be assigned to point clouds in different areas. The dynamic similarity calculation formula is as follows:

[0066] DS = LB(Similarity(R(Fi)));

[0067] Among them, LB is the Line + BatchNorm operation, Similarity is to calculate the similarity in the high-dimensional feature matrix through the matrix dot product. Each element in the obtained similarity matrix corresponds to the similarity between every two elements in the feature matrix. Finally, the similarity matrix is ​​normalized through the LB layer. Figure 3 The SIMI in is Similarity. Analogous to the traditional attention model, the dynamic similarity attention is as follows:

[0068] DSA=LB(Similarity(R(Fi)))·SL(Q·K T )

[0069] SL represents the Softmax+L1Norm algorithm;

[0070] And the final output is F out , the formula is as follows:

[0071] Fout=DSA·V+Fi

[0072] The improved attention model dynamically allocates attention weights for each round of training, which can show more detailed modeling of different parts of the point cloud, thereby realizing the relationship learning between different points and improving classification accuracy.

[0073] Step S3: Map the feature information to different classification directories to obtain type scores.

[0074] Specifically, three linear layers are used to gradually map the features extracted and learned by the previous modules to the sample annotation space. The last linear layer reduces the features to categories of size 1×Nc, i.e., one-dimensional sequence values.

[0075] In this embodiment, the classification catalog includes qualified and unqualified, that is, corresponding to two types of scores "OK" and "NG".

[0076] In some other embodiments, the classification catalog may include qualified, defect type 1, defect type 2, etc., and each classification corresponds to a score. The operator may determine the current defect type based on the type score obtained by the detection method of the present invention.

[0077] The technical solution of the present invention has the following technical effects compared with the prior art:

[0078] The defect detection method proposed by the present invention has high accuracy and high stability of the detection result; at the same time, the detection efficiency is also improved.

[0079] Example 2: This example will be based on Figure 1 and Figure 6 A Transformer-based power battery weld point cloud defect detection system and electronic device 200 are described.

[0080] For the detection system, refer to Figure 6 As shown, including:

[0081] A feature embedding module 110 is used to extract and aggregate features after high-dimensional mapping of the original point cloud data;

[0082] A dynamic similarity attention module 120 is used to import the high-dimensional features output by the feature embedding module as input into the attention network based on dynamic similarity to redistribute attention weights and extract multiple feature information of the point cloud;

[0083] The classification module 130 is used to map the obtained high-dimensional feature information into different classification directories and obtain and output type scores.

[0084] In some embodiments of the present application, the main function of the feature embedding module 110 is to map the input points to a higher-dimensional feature space to facilitate subsequent learning, and the main implementation method is a multi-layer perceptron (MLP).

[0085] In some embodiments of the present application, the dynamic similarity attention module 120 refers to Figure 3As shown in Figure 1, it is improved based on the self-attention module. Each point cloud has its own characteristics. The traditional attention mechanism is globally uniform and insensitive to the relative positions between points. That is, it considers the interactions between all points in the point cloud, but does not consider the local relationships in the point cloud.

[0086] The dynamic similarity attention module 120 in this embodiment introduces a dynamic similarity attention (DSA) mechanism. DSA implements a non-uniform attention mechanism. In the point cloud feature space after feature extraction, the attention scores are weighted according to the similarity differences between high-dimensional points. Therefore, DSA can accurately distinguish different parts in the point cloud, especially for battery welding defect judgment tasks, which require different weights to be assigned to point clouds in different areas. Increasing the similarity calculation of input features can improve the accuracy of defect detection.

[0087] Experimental verification shows that the improved attention model dynamically allocates attention weights for each round of training, which can show more detailed modeling of different parts in the point cloud, thereby realizing the relationship learning between different points and improving classification accuracy.

[0088] In some embodiments of the present application, the classification module 130 includes three linear layers, which are used to gradually map the features extracted and learned by the dynamic similarity attention module to the sample annotation space. The last linear layer reduces the features to a category of size 1×Nc, i.e., a one-dimensional sequence value. In addition, in order to avoid overfitting, a dropout layer is added to the linear layer.

[0089] In terms of loss function, this embodiment adopts the classic cross entropy loss function.

[0090] In some embodiments of the present application, the accuracy and stability of other point cloud network models in weld defect detection tasks are compared, such as PCT, PointNet, and DGCNN. In order to verify the effectiveness of the dynamic similarity attention (DSA) module proposed in this embodiment in the attention weight allocation, we also added the TDSPnet network without the DSA module, namely TDSPnet-naive, to the comparative experiment, where the TDSPnet network is the network used in this application. The accuracy of each model training process is shown in Figure 2. Figure 4 As shown, the accuracy and loss values ​​of the test set of the TDSPnet network are as follows Figure 5 The following are the specific analysis results:

[0091] In terms of convergence speed, PointNet, DGCNN, and TDSPnet showed faster convergence speeds within the first 50 cycles of training and quickly reached higher accuracy. Specifically, PointNet converged to an accuracy of 0.75 in 27 cycles; DGCNN and TDSPnet reached an accuracy of 0.85 in 9 and 12 cycles, and an accuracy of 0.90 in 44 and 25 cycles, respectively. In contrast, the initial convergence speeds of TDSPnet-naive and PCT were relatively slow, reaching convergence in about 60 cycles and 180 cycles, respectively.

[0092] In terms of model stability, it can be observed from the figure that TDSPnet and DGCNN show good stability and small fluctuations after reaching the highest accuracy. In contrast, TDSPnet-naive and PCT still have certain fluctuations in the later stages of training, which shows that their stability is slightly lower than that of TDSPnet. Although the accuracy curve of PointNet is relatively stable, its overall accuracy is significantly lower than that of other models.

[0093] In terms of the final performance of the model, at the end of the training, TDSPnet and PCT had the highest accuracy, reaching 0.9594 and 0.9527 respectively, showing their strong ability in processing point cloud data. The final accuracy of TDSPnet-naive and DGCNN was slightly lower, at 0.9401 and 0.9289 respectively. PointNet had the lowest final accuracy of 0.8073, indicating that its performance on this task was inferior to other models. In the test set, TDSPnet also showed the highest accuracy of 0.9555.

[0094] In terms of actual engineering applications, TDSPnet showed the lowest missed detection rate in the test set, only 4.52%, which strengthened its reliability. The missed detection rates of PCT and TDSPnet-naive were 7.04% and 8.04% respectively. In comparison, the missed detection rates of PointNet and DGCNN were as high as 40.21% and 13.07% respectively. In addition, TDSPnet also performed better in detection efficiency.

[0095] Based on the above analysis, the TDSPnet used in this application performs well in many aspects such as convergence speed, stability and final accuracy, and is superior to other reference network models. In the task of processing weld defect point cloud data, TDSPnet demonstrates strong capabilities and reliability, and is suitable for application in practical engineering.

[0096] It should be understood that the detection system here is embodied in the form of a functional module. The term "module" here may refer to an application specific integrated circuit (ASIC), an electronic circuit, a processor (e.g., a shared processor, a proprietary processor, or a group processor, etc.) and a memory for executing one or more software or firmware programs, a combined logic circuit, and / or other suitable components that support the described functions. In an optional example, it will be appreciated by those skilled in the art that the detection system may be specifically the electronic device 200 in the above-described embodiment, or the function of the electronic device 200 in the above-described embodiment may be integrated in the detection system, and the detection system may be used to execute the various processes and / or steps corresponding to the electronic device 200 in the above-described method embodiment, and to avoid repetition, it will not be repeated here.

[0097] The detection system has the function of implementing the corresponding steps performed by the electronic device 200 of the calculation method in Example 1; the above functions can be implemented by hardware, or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above functions. For example, the acquisition module can be a communication interface, such as a transceiver interface.

[0098] In the embodiments of the present application, Figure 1 The detection system may also be a chip or a chip system, such as a system on chip (SoC).

[0099] Reference Figure 6 As shown, in this embodiment, an electronic device 200 is provided, including:

[0100] A processor 210, and a memory 220 and a transceiver 230 communicatively connected to the processor;

[0101] The memory 220 stores computer-executable instructions; the transceiver 230 is used to send and receive data;

[0102] The processor 210 executes the computer-executable instructions stored in the memory 220 to implement the calculation method in Example 1.

[0103] It should be understood that the electronic device 200 can be used to execute the corresponding steps and / or processes in the above method embodiments. Optionally, the memory 220 may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory 220 may also include a non-volatile random access memory. For example, the memory 220 may also store information about the device type. The processor 210 may be used to execute the instructions stored in the memory 220, and when the processor 210 executes the instructions, the processor 210 may execute the corresponding steps and / or processes in the above method embodiments.

[0104] It should be understood that in the embodiment of the present application, the processor 210 may be a central processing unit (CPU), and the processor 210 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0105] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor 210 or an instruction in the form of software. The steps of the method disclosed in conjunction with the embodiment of the present application can be directly embodied as a hardware processor for execution, or a combination of hardware and software modules in the processor 210 for execution. The software module can be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor executes the instructions in the memory, and completes the steps of the above method in conjunction with its hardware. To avoid repetition, it is not described in detail here.

[0106] Embodiment 3: In this embodiment, a computer-readable storage medium is provided, characterized in that:

[0107] The computer-readable storage medium stores computer-executable instructions, which are used to implement the calculation method in Example 1 when executed by a processor.

[0108] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0109] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0110] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0111] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks or optical disks.

[0112] In the description of the above embodiments, specific features, structures, materials or characteristics may be combined in a suitable manner in any one or more embodiments or examples.

[0113] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A power battery weld point cloud defect detection method based on Transformer, characterized in that: The following steps are involved: Collecting original point cloud data, and performing feature extraction on the original point cloud data to obtain a point cloud feature matrix; The point cloud feature matrix is ​​input into a multi-layer attention network, and attention is dynamically allocated according to the similarity difference between points in the point cloud data to obtain multiple feature information; The dynamic allocation of attention according to the similarity difference between points in the point cloud data includes: The dynamic similarity is calculated by Line + BatchNorm algorithm; the calculation formula of the dynamic similarity is: ； Among them, Fi represents the input feature; R represents the copy instruction; Similarity is to calculate the similarity in the high-dimensional feature matrix through the matrix dot product, and each element in the obtained similarity matrix corresponds to the similarity between every two elements in the feature matrix; LB is the Line + BatchNorm operation; dynamic similarity attention is calculated based on dynamic similarity and multiple features are output; The calculation formula of the dynamic similarity attention is: ； The output formula of the multiple features is: ； Among them, SL represents the Softmax+L1Norm algorithm; Fi represents the input feature; Q, K, and V correspond to the query, key, and value in the Transformer network framework respectively, and the calculation formula is as follows: ； Where W q , W k , W v Represents a shared linear layer trainable matrix; Map feature information to different classification directories to obtain type scores.

2. According to the Transformer-based power battery weld point cloud defect detection method of claim 1, it is characterized in that: The extraction process of the point cloud feature matrix includes: Determine the coordinates of adjacent points of the original point cloud data; Subtract the coordinates of the center point from the coordinates of the neighboring points and copy and concatenate them with the K layers of the center point to obtain high-dimensional features; After the high-dimensional features are processed by convolution and LBR layers, significant features are extracted from K neighboring points through maximum pooling to obtain the point cloud feature matrix.

3. According to the Transformer-based power battery weld point cloud defect detection method of claim 2, it is characterized in that: Determining the coordinates of the neighboring points of the original point cloud data includes: The farthest point is sampled for the original point cloud data; and the K neighboring points of each point are found through K-neighbor search in the original point cloud data; The coordinates of the K neighboring points are subtracted from the coordinates of the corresponding sampling point to obtain the normalized coordinates of the neighboring points.

4. A power battery weld point cloud defect detection system based on Transformer, used to execute a power battery weld point cloud defect detection method based on Transformer according to any one of claims 1 to 3, characterized in that: include: Feature embedding module, used to extract and aggregate features after high-dimensional mapping of raw point cloud data; The dynamic similarity attention module is used to import the high-dimensional features output by the feature embedding module as input into the attention network based on dynamic similarity to redistribute the attention weights and extract multiple feature information of the point cloud; The classification module is used to map the obtained high-dimensional feature information into different classification directories and obtain and output type scores.

5. According to the Transformer-based power battery weld point cloud defect detection system of claim 4, it is characterized in that: The classification module includes three linear layers, which are used to gradually map the features extracted and learned by the dynamic similarity attention module to the sample annotation space.

6. An electronic device, characterized in that: include: A processor, and a memory and a transceiver communicatively connected to the processor; The memory stores computer-executable instructions; the transceiver is used to send and receive data; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 3.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 3 when executed by a processor.

Images