CN111539949A - Point cloud data-based lithium battery pole piece surface defect detection method - Google Patents

Point cloud data-based lithium battery pole piece surface defect detection method Download PDF

Info

Publication number
CN111539949A
CN111539949A CN202010401330.7A CN202010401330A CN111539949A CN 111539949 A CN111539949 A CN 111539949A CN 202010401330 A CN202010401330 A CN 202010401330A CN 111539949 A CN111539949 A CN 111539949A
Authority
CN
China
Prior art keywords
point cloud
loss
defect
point
lithium battery
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010401330.7A
Other languages
Chinese (zh)
Other versions
CN111539949B (en
Inventor
陈海永
王涛
刘卫朋
郭婧
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hebei University of Technology
Original Assignee
Hebei University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hebei University of Technology filed Critical Hebei University of Technology
Priority to CN202010401330.7A priority Critical patent/CN111539949B/en
Publication of CN111539949A publication Critical patent/CN111539949A/en
Application granted granted Critical
Publication of CN111539949B publication Critical patent/CN111539949B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G06T7/0006—Industrial image inspection using a design-rule based approach
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
    • G01B11/00—Measuring arrangements characterised by the use of optical techniques
    • G01B11/24—Measuring arrangements characterised by the use of optical techniques for measuring contours or curvatures
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/89—Investigating the presence of flaws or contamination in moving material, e.g. running paper or textiles
    • G01N21/8914—Investigating the presence of flaws or contamination in moving material, e.g. running paper or textiles characterised by the material examined
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/89—Investigating the presence of flaws or contamination in moving material, e.g. running paper or textiles
    • G01N21/892—Investigating the presence of flaws or contamination in moving material, e.g. running paper or textiles characterised by the flaw, defect or object feature examined
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G01N2021/8883—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges involving the calculation of gauges, generating models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10028—Range image; Depth image; 3D point clouds

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Biochemistry (AREA)
  • Chemical & Material Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Textile Engineering (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Signal Processing (AREA)
  • Quality & Reliability (AREA)
  • Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)

Abstract

The invention utilizes a method for detecting the surface defects of the lithium battery pole piece based on point cloud data, and effectively utilizes the geometric characteristic information of the defects. The detection network for detecting the 3D point cloud target has high real-time performance, can adapt to feature extraction of various defects on the surface of a lithium battery pole piece, and simultaneously reduces the sensitivity of small target defects to surrounding background interference information, thereby enhancing the stability of defect detection. The safety of the production link of the lithium battery is guaranteed, and the production quality of the lithium battery is improved.

Description

Point cloud data-based lithium battery pole piece surface defect detection method
Technical Field
The invention relates to the technical field of lithium battery pole piece surface defect detection, in particular to a point cloud data-based lithium battery pole piece surface defect detection method.
Background
Lithium batteries are a new type of high-energy batteries using lithium metal or lithium alloy as a negative electrode material and a nonaqueous electrolyte solution. With the development of scientific technology, various mobile electronic devices are widely applied in human life, and the comprehensive performances of lithium batteries in specific capacity, no memory effect, long service life, environmental protection and the like far exceed those of other secondary batteries, so that the lithium batteries become a mainstream power supply scheme of the mobile electronic devices. Meanwhile, a large-capacity lithium battery becomes a power supply, and has started to be applied to electric vehicles and aerospace.
The coating of the lithium battery electrode is that an electrode coating is smeared on a metal foil such as aluminum or copper, and then drying and dedusting are carried out, so that the coating can be used as a material of the lithium battery electrode. A common method for detecting the surface defects of the lithium battery pole piece is generally realized by using an industrial camera and an LED light source to acquire image data for processing. The surface defects of the lithium battery, such as scratches, bubbles, foreign matters, pinholes and the like, have the characteristics of weak texture, low chromatic aberration and the like, and have certain difficulty in image processing. The spatial geometric characteristics of the high-precision three-dimensional scanning point cloud data can be fully utilized, the information quantity is larger, the acquired data is closer to the real form of the defect, and more characteristic information can be brought theoretically.
Wuqinghua (Wuqinghua, theory and application research of online detection of three-dimensional surface defects based on line structured light scanning [ D ]. Huazhong university of science and technology, 2013.) uses an ICP (inductively coupled plasma) registration method, and after the surface to be detected is registered with a defect-free template in each detection, inconsistent regions are extracted for judgment, the method is established on the premise that the defect-free part accounts for most of point clouds, meanwhile, the method has larger calculation amount depending on the registration precision, and can not classify and position defect pairs with various sizes in effect.
Disclosure of Invention
In order to solve the problems that detection in the prior art is difficult to adapt to various size defects, and defects of small targets are strongly interfered by surrounding backgrounds, the invention provides a method for detecting defects on the surface of a lithium battery pole piece based on point cloud data.
In order to achieve the purpose, the invention adopts the following technical scheme:
a lithium battery pole piece surface defect detection method based on point cloud data is used for detection by a learning-based method and is characterized by comprising the following steps:
(1) point cloud acquisition:
firstly, obtaining point clouds on the surface of a lithium battery pole piece through a line scanning type laser sensor, then removing invalid points, averagely dividing the point clouds on the surface into B point cloud blocks, and respectively carrying out random down-sampling and unit processing on each point cloud block through parallel processing to form a batch input detection network;
(2) making a data set sample and a label:
sorting the sizes, the directions and the positions of various manually marked defect bounding box labels:
counting the average bounding box size of each type of defect as a reference for learning the defect bounding box size, wherein a certain defect bounding box size consists of the average bounding box size of the type to which the defect belongs and a regression factor of the residual size of the bounding box of the defect;
on the plane of the lithium battery pole piece, the angle values of all the defect bounding boxes are normalized to a clockwise rotation angle within the range of 0-360 degrees, then 0-360 degrees are equally divided into a plurality of intervals, the angles in the intervals are regarded as a class, and the rotation angle of each defect bounding box consists of the class of the interval and residual regression;
sampling point clouds in the defect bounding box by adopting a farthest point to obtain a plurality of surface key points, taking the surface key points and the center point of the defect bounding box as voting key points of the defect, recording the offset of the surface key points relative to the center point of the defect bounding box, and determining the position of the defect in the current point cloud;
(3) constructing a multi-scale feature extraction backbone network with graph attention:
(3-1) design the feature layer FL containing the graph attention convolution:
expressing Point cloud by using a graph structure, introducing an attention mechanism, adding Point-wise attention and Channel-wise attention to the graph structure in a serial mode, and giving two attention weights to the graph structure; finally, inputting the graph structure containing two attention weights into a third MLP network, wherein the third MLP network is connected with a symmetric function, the implementation mode of the symmetric function adopts maximum pooling, the high-dimensional characteristics of the graph structure with attention are automatically learned, and the design of the graph attention convolution layer is completed, so that a characteristic layer FL containing graph attention convolution is formed;
(3-2) integrating multi-scale features by using a pyramid structure consisting of a plurality of feature layers FL containing graph attention convolution and a feature propagation layer with jump connection to complete construction of a multi-scale feature extraction backbone network with graph attention;
(4) a depth Hough voting mechanism is introduced, and the key points of the defect target are robustly searched:
inputting the output of the multi-scale feature extraction backbone network with the drawing attention into a fourth MLP network, automatically generating votes by the fourth MLP network through learning, and aggregating the votes to obtain voting key points VkeyInformation and features V after voting clusteringfea；
(5) According to the aggregated voting key point information obtained in the step (4) and the feature V after voting clusteringfeaPredicting a series of candidate defect target proposals, including voting key points VkeyRemoving repeated bounding boxes in the candidate defect target and screening out the most possible defect positions.
In the point cloud obtaining mode in the step (1), a motor drives a roller to enable a lithium battery pole piece to move horizontally, a line scanning type laser sensor is fixed above the lithium battery pole piece, the line scanning type laser sensor is triggered by an encoder to acquire data frame by frame, and each 1024 frame of data is set to be combined into one point cloud; the surface point cloud is evenly divided into 16 point cloud blocks, and 8192 points are obtained after random down-sampling.
The adding mode of the Point-wise attention and the Channel-wise attention in the step (3) is as follows:
for a graph structure G, on the dimension of a feature Channel, firstly, a pooling layer is used for gathering Point features to the Channel-wise dimension to obtain a Point-wise response RpThen accessing the first MLP network to learn the connection weight W of each neighborhoodp(ii) a G and WpMultiplying corresponding elements to obtain a graph G with Point-wise attentionp；
For GpIn the K dimension, the Channel characteristics are firstly gathered to the Point-wise dimension by using the pooling layer to obtain Channel-wise response RcThen accessing a second MLP network to learn the weight W of each channelc；GpAnd WcMultiplying the corresponding elements to obtain a graph G with two attentionspcAnd K is the neighborhood number.
The concrete process of the pyramid structure is as follows: the method comprises the steps that a plurality of feature layers FL containing graph attention convolution are included, the feature layers FL are sequentially stacked to form a plurality of feature layers with different scales, the feature layer with the first scale is firstly subjected to point cloud coarsening processing by adopting farthest point sampling, then point cloud subjected to coarsening processing is sent into the graph attention convolution with the scale, the output of the feature layer with the first scale is subjected to point cloud coarsening processing by adopting the farthest point sampling, high-dimensional features output by the feature layer with the first scale are sent into the feature layer with the second scale together, and the like, the stacking of the feature layers with the different scales is completed, and the fusion of multi-scale features is completed through interpolation, connection and feature generation among the feature layers with the different scales.
Compared with the prior art, the invention has the beneficial effects that:
the invention utilizes a method for detecting the surface defects of the lithium battery pole piece based on point cloud data, and effectively utilizes the geometric characteristic information of the defects. The detection network for detecting the 3D point cloud target for detecting the defects on the surface of the lithium battery pole piece is provided, has high real-time performance, can adapt to the feature extraction of various defects on the surface of the lithium battery pole piece, and simultaneously reduces the sensitivity of small target defects to surrounding background interference information, thereby enhancing the stability of defect detection. The method is beneficial to guaranteeing the safety of the production link of the lithium battery, improves the production quality of the lithium battery, meets the precision requirement of the lithium battery industry on defect detection and positioning, and can realize classification and positioning of various size defects such as scratches, cracks, bubbles and particles.
The method organically combines the graph structure, two attention mechanisms and multi-scale fusion, simultaneously refers to a depth Hough voting mechanism, obtains surface key points and center points for prediction, fully utilizes point pair information, is more stable in model, and improves prediction accuracy. The introduction of the attention of the graph strengthens the correlation between local point pairs, defines the key degree between different channel characteristics, is favorable for learning the characteristic difference between the background and the defect and between the defect and the defect, utilizes a pyramid structure to fuse multi-scale characteristics, strengthens the detection capability of the small target defect, obtains the defect key points through a voting mechanism, and further improves the robustness of background noise.
Drawings
FIG. 1 illustrates a point cloud acquisition method used in the present invention.
FIG. 2 is a flow chart of the work of the method for detecting the surface defects of the lithium battery pole piece based on the point cloud data.
FIG. 3 is a block diagram of the overall network model in the method of the present invention.
FIG. 4 is a schematic diagram of the structure of the attention convolution layer in the feature pyramid of FIG. 3.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Example 1
The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data comprises the following steps:
(1) and point cloud obtaining.
The motor drives the roller to enable the lithium battery pole piece to move horizontally, the line scanning type laser sensor is fixed above the lithium battery pole piece, the encoder triggers the line scanning type laser sensor to acquire data frame by frame, every 1024 frame of data is combined into a point cloud, and the lithium battery pole piece moves horizontally by about 80 mm.
The preprocessing firstly removes invalid points, then divides dense surface point cloud containing a large number of points, averagely divides the dense surface point cloud into 16 point cloud blocks, randomly samples each point cloud block to 8192 points through parallel processing, and then conducts unitization processing. And forming a batch _ size ═ B input detection network. The characteristic that the neural network can be processed in batch is fully utilized, and more geometric information is kept as far as possible.
(2) A data set sample and a label are made.
The manually marked labels (type, position, size and direction) of the bounding boxes with the scratches, cracks, bubbles and particle defects are sorted, so that the learning of the detection network is facilitated. Mainly the loss of the defect envelope size and its rotation angle. The size and rotation angle regression task of the defect bounding box is converted into a classification task and a regression task respectively, and the detection precision is improved by utilizing the property of the strong classification capability of the CNN.
(2-1) counting the average bounding box size of each type of defect as a reference for learning the bounding box size of the defect.
Average bounding box size Bbox for each type of defectmean_k＝(lmean_k,wmean_k,hmean_k) In which lmean_kMean bounding box length, w, for the k-th defectmean_kMean width of bounding box for k-th defect, hmean_kThe mean height of the bounding box for the k-th defect.
A certain defect bounding box size is defined by the average bounding box size Bbox of the class to which the defect belongsmean_kAnd the regression factor Res of the residual size of the bounding box of the defectbbox_d＝(Resl_d,Resw_d,Resh_d) Composition, i.e. certain defect bounding box size Bboxd＝Bboxmean_k+Resbbox_d⊙Bboxmean_k，
Where ⊙ denotes the Hadamard product, Resl_d,Resw_d,Resh_dThe values of the regression factors of the residual sizes in the length direction, the width direction and the height direction are (-1, 1)]。
(2-2) on the plane of the lithium battery pole piece, the defect surrounding box only comprises rotation of one degree of freedom, and in order to avoid the influence of the periodicity of the rotation angle on network learning, the angle value α is normalized to the clockwise rotation angle α within the range of 0-360 degreesf。
Figure BDA0002487793790000041
Where α' ═ α mod (2 π), mod denotes the remainder.
Will make an angle αfAnd converting into classification and regression tasks. Specifically, 0-360 ° needs to be equally divided into 12 sections, and the defect enclosure rotation angle falling within a certain section is regarded as a class. As shown in the formula, the rotation angle of each defect bounding box is composed of the type of the located interval and residual regression.
Figure BDA0002487793790000042
Wherein classrota0,1,2, …, s denotes the classification of s intervals, residual regression factor ResrotaIn the range of (-1, 1)]。
And (2-3) making a voting label. Sampling 16 points from the point cloud in the defect bounding box by adopting the farthest point as a surface key point, taking the surface key point and the center point of the defect bounding box as a voting key point of the defect, recording the offset of the surface key point relative to the center point of the defect bounding box, and determining the position of the defect in the current point cloud.
At this point, the dimensions and rotation angles of the defect bounding box are represented by category and residual regression, and data set samples and labels are prepared. The Loss function used for classification of defect bounding box size and rotation during training is Cross Engine Loss, and the Loss function used for regression of bounding box size and rotation is Huber Loss, which is used to minimize classification score error and residual error.
(3) And constructing a multi-scale feature extraction backbone network with graph attention.
(3-1) design feature layer FL with graph attention convolution
(3-1-1) expressing the point cloud using a graph structure:
from the given point cloud, a map structure G ═ (v, e) is created. v includes all points x of the point cloudiE is an edge feature function eijSet of (2), x found in Euclidean space by K-NN methodiK neighborhood points xijWhere i is 0,1,2, …, m, j is 0,1,2, …, K.
Let edge feature function eij＝h(xi,xij)＝h(xi,xij-xi) The network learns the geometric relations of all points and the neighborhood points, thus simultaneously considering the global characteristics and the local characteristics and obtaining the graph structure G with the shape of [ m, C, K ]]Wherein m is the number of all point clouds, and C is the number of characteristic channels. According to the point cloud density, the invention takes the neighborhood number K as 16.
(3-1-2) an attention mechanism is introduced to add Point-wise and Channel-wise attention in series to the graph structure, giving the graph structure two attention weights.
Specifically, on the feature Channel dimension of the graph structure G, firstly, a maximum pooling layer is used for gathering Point features to the Channel-wise dimension to obtain a Point-wise response Rp. Then accessing a first MLP network (multilayer perceptron network) to learn attention W of each neighborhood pointpI.e. by
Wp＝w2(w1Rp)，
Wherein w1、w2The learning parameters of the two fully-connected layers are respectively the activation function Relu.
Then the graph G with Point-wise attentionp＝Wc⊙Rp。
Similarly, for GpIn K dimension, firstly, using maximum pooling layer to gather channel characteristics to Point-wise dimension to obtain CHANDEL-WISE RESPONSE Rc. Then, a second MLP network (consisting of two fully-connected layers and a Relu function) is accessed, and the weight W of each channel is learnedcI.e. by
Wc＝w′2(w1′Rc)，
Wherein w1′、w′2Respectively, the learning parameters of the two fully connected layers.
Diagram G with two kinds of attentionpc＝Wc⊙Rc。
(3-1-3) adding GPCInputting the data into a third MLP network, wherein the third MLP network is connected with a symmetric function in a rear part, and the implementation mode of the symmetric function adopts maximum pooling to automatically learn the high-dimensional characteristics of the graph structure with attention, so as to complete the design of the graph attention convolution layer, thereby forming a characteristic layer FL containing graph attention convolution;
the third MLP network described above consists of two-dimensional convolution layers (size of convolution kernel 1 × 1), Relu function.
And (3-2) integrating the multi-scale features by using a pyramid structure consisting of a plurality of feature layers FL and a feature propagation layer with jump connection (interpolation, connection and feature).
And each feature layer FL containing the graph attention convolution firstly adopts the farthest point sampling to carry out point cloud coarsening treatment, and then sends the point cloud subjected to the coarsening treatment into the graph attention convolution. The output of the first layer FL is subjected to point cloud coarsening through farthest point sampling, then the coarsened point cloud and the high-dimensional features output by the first layer FL are input into the second layer, and by parity of reasoning, 4 feature layers with different scales are respectively formed by sequentially stacking the 4 feature layers FL and are respectively marked as FL1、FL2、FL3、FL4。
In particular FL1、FL2、FL3、FL4Comprises xyz coordinates of point cloud points 4096, 2048, 1024 and 512 obtained by point cloud coarsening treatment and high-dimensional characteristics FL corresponding to the point cloud points after each layer of graph attention convolution1f、FL2f、FL3f、FL4f。
FL (general-purpose lamp)4fAnd the FL3fThrough interpolation, connection, feature generation and FL3fPoint-consistent features FL3f', will FL3f'AND' AND2fThrough interpolation, connection, feature generation and FL2fPoint-consistent features FL2f', will FL2f'AND' AND1fThrough interpolation, connection, feature generation and FL1fPoint-consistent features FL1f', completing the fusion of the multi-scale features.
(4) And a depth Hough voting mechanism is introduced to robustly search key points of the defect target.
Different from the traditional method that voting needs to be designed manually, the voting is automatically generated by a fourth MLP network (comprising three layers of convolution and two Relu functions) through learning, and voting is aggregated to obtain voting key point information VkeyAnd features V after voting clusteringfeat(number of points and number of channels).
In particular, the output FL of the multi-scale feature extraction backbone network with drawing attention1f' inputting the information into a fourth MLP network to generate voting information, wherein the MLP network used for generating the voting consists of one-dimensional convolution (the size of a convolution kernel is 1) and ReLU; the aggregate voting operation still uses the feature layer FL with the graph attention convolution.
(5) And (4) predicting a series of candidate defect target proposals according to the aggregation voting key point information obtained in the step (4), and finally screening out the most possible defect positions.
Clustering features V of votesfeatExtracting high-dimensional features of votes through a fifth MLP network (one-dimensional convolution (size 1 of convolution kernel) and ReLU composition) to generate candidate defect target proposal results including voting key points VkeyDeviation of information, defect bounding box size, angle, defect type probability. And removing repeated bounding boxes in the candidate target by using three-dimensional non-maximum suppression (3D-NMS) to screen out the target which is most likely to be a defect.
Specifically, the number of output candidate proposals is set to 256, and the IoU threshold of the 3D-NMS is set to 0.25.
(6) The total Loss is composed of a plurality of Loss tasks, namely voting Loss, bounding box Loss and semantic Loss.
Loss＝Lossvote+λ1LossBbox+λ2Losssem
Wherein λ1And λ2To balance the weight lost by the various parts.
The bounding box loss specifically includes: loss of center point LosscenterBounding box size classification Losssize_classAnd Loss of residue Losssize_resLost Loss of bounding box rotation classificationrota_classAnd Loss of residue Lossrota_res。
LossBbox＝Losscenter+Losssize_class+Losssize_res+Lossrota_class+Lossrota_resAnd saving the optimal primary network model training parameters and determining the network model parameters.
The model of the present invention (constituting the detection network via steps (3) - (6)) uses a random gradient descent (SGD) to update the training parameters, momentum 0.9, and initial learning rate 0.001. In order to improve the noise resistance of the network, data enhancement processing is carried out along with iterative loading, wherein the data enhancement processing comprises that point cloud coordinate values xyz are respectively randomly zoomed within the range of 0.9-1.1 times, and the rotation angle of the defect bounding box label is increased and randomly rotated within the range of-5 degrees to 5 degrees.
The model trains 200 epochs, with a learning rate that decays to 0.1 times every 40 epochs. And evaluating the Loss of the network model on the verification set every 10 epochs, and saving the optimal network model training parameters. The training set and the verification set are each composed of a certain number (in this embodiment, the ratio is 4:1) of samples randomly decimated by the proportion of the data set samples obtained in the above step (2).
Nothing in this specification is said to apply to the prior art.

Claims (10)

1. A lithium battery pole piece surface defect detection method based on point cloud data is used for detection by a learning-based method and is characterized by comprising the following steps:
(1) point cloud acquisition:
firstly, obtaining point clouds on the surface of a lithium battery pole piece through a line scanning type laser sensor, then removing invalid points, averagely dividing the point clouds on the surface into B point cloud blocks, and respectively carrying out random down-sampling and unit processing on each point cloud block through parallel processing to form a batch input detection network;
(2) making a data set sample and a label:
sorting the sizes, the directions and the positions of various manually marked defect bounding box labels:
counting the average bounding box size of each type of defect as a reference for learning the defect bounding box size, wherein a certain defect bounding box size consists of the average bounding box size of the type to which the defect belongs and a regression factor of the residual size of the bounding box of the defect;
on the plane of the lithium battery pole piece, the angle values of all the defect bounding boxes are normalized to a clockwise rotation angle within the range of 0-360 degrees, then 0-360 degrees are equally divided into a plurality of intervals, the angles in the intervals are regarded as a class, and the rotation angle of each defect bounding box consists of the class of the interval and residual regression;
sampling point clouds in the defect bounding box by adopting a farthest point to obtain a plurality of surface key points, taking the surface key points and the center point of the defect bounding box as voting key points of the defect, recording the offset of the surface key points relative to the center point of the defect bounding box, and determining the position of the defect in the current point cloud;
(3) constructing a multi-scale feature extraction backbone network with graph attention:
(3-1) design the feature layer FL containing the graph attention convolution:
expressing Point cloud by using a graph structure, introducing an attention mechanism, adding Point-wise attention and Channel-wise attention to the graph structure in a serial mode, and giving two attention weights to the graph structure; finally, inputting the graph structure containing two attention weights into a third MLP network, wherein the third MLP network is connected with a symmetric function, the implementation mode of the symmetric function adopts maximum pooling, the high-dimensional characteristics of the graph structure with attention are automatically learned, and the design of the graph attention convolution layer is completed, so that a characteristic layer FL containing graph attention convolution is formed;
(3-2) integrating multi-scale features by using a pyramid structure consisting of a plurality of feature layers FL containing graph attention convolution and a feature propagation layer with jump connection to complete construction of a multi-scale feature extraction backbone network with graph attention;
(4) a depth Hough voting mechanism is introduced, and the key points of the defect target are robustly searched:
inputting the output of the multi-scale feature extraction backbone network with the drawing attention into a fourth MLP network, automatically generating votes by the fourth MLP network through learning, and aggregating the votes to obtain voting key points VkeyInformation and features V after voting clusteringfea；
(5) According to the aggregated voting key point information obtained in the step (4) and the feature V after voting clusteringfeaPredicting a series of candidate defect target proposals, including voting key points VkeyRemoving repeated bounding boxes in the candidate defect target and screening out the most possible defect positions.
2. The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 1, wherein the method comprises the following steps: in the point cloud obtaining mode in the step (1), a motor drives a roller to enable a lithium battery pole piece to move horizontally, a line scanning type laser sensor is fixed above the lithium battery pole piece, the line scanning type laser sensor is triggered by an encoder to acquire data frame by frame, and each 1024 frame of data is set to be combined into one point cloud; the surface point cloud is evenly divided into 16 point cloud blocks, and 8192 points are obtained after random down-sampling.
3. The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 1, wherein the method comprises the following steps: a certain defect bounding box size Bbox in (2)dThe expression of (a) is: bboxd＝Bboxmean_k+Resbbox_d⊙Bboxmean_kWhere ⊙ denotes the Hadamard product, the average bounding box size Bbox for each type of defectmean_kBounding box residual size regression factor Resbbox_d；
Angle of rotation α for each defect boxfThe expression is as follows:
Figure FDA0002487793770000021
wherein classrota0,1,2, …, s denotes the classification of s intervals, residual regression factor ResrotaIn the range of (-1, 1)]。
4. The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 1, wherein the method comprises the following steps: the adding mode of the Point-wise attention and the Channel-wise attention in the step (3) is as follows:
for a graph structure G, on the dimension of a feature Channel, firstly, a pooling layer is used for gathering Point features to the Channel-wise dimension to obtain a Point-wise response RpThen accessing the first MLP network to learn the connection weight W of each neighborhoodp(ii) a G and WpMultiplying corresponding elements to obtain a graph G with Point-wise attentionp；
For GpIn the K dimension, the Channel characteristics are firstly gathered to the Point-wise dimension by using the pooling layer to obtain Channel-wise response RcThen accessing a second MLP network to learn the weight W of each channelc；GpAnd WcMultiplying the corresponding elements to obtain a graph G with two attentionspcAnd K is the neighborhood number.
5. The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 4, wherein the method comprises the following steps: the first MLP network and the second MLP network are both composed of two fully connected layers and a Relu function sandwiched between the two fully connected layers.
6. The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 1, wherein the method comprises the following steps: the concrete process of the pyramid structure is as follows: the method comprises the steps that a plurality of feature layers FL containing graph attention convolution are included, the feature layers FL are sequentially stacked to form a plurality of feature layers with different scales, the feature layer with the first scale is firstly subjected to point cloud coarsening processing by adopting farthest point sampling, then point cloud subjected to coarsening processing is sent into the graph attention convolution with the scale, the output of the feature layer with the first scale is subjected to point cloud coarsening processing by adopting the farthest point sampling, high-dimensional features output by the feature layer with the first scale are sent into the feature layer with the second scale together, and the like, the stacking of the feature layers with the different scales is completed, and the fusion of multi-scale features is completed through interpolation, connection and feature generation among the feature layers with the different scales.
7. The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 1, wherein the method comprises the following steps: for the depth Hough voting mechanism in the step (4), the aggregation voting operation adopts a feature layer FL containing graph attention convolution; the fourth MLP network used to generate votes consists of three convolution kernels, one-dimensional convolution layers of size 1, and ReLU functions placed between the convolution layers.
8. The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 1, wherein the method comprises the following steps: the total Loss function Loss of the detection method consists of a plurality of Loss tasks which are voting Loss respectivelyvoteLoss of bounding box1LossBboxLoss of semanticssemThe expression is:
Loss＝Lossvote+λ1LossBbox+λ2Losssem，
wherein λ1And λ2To balance the weight lost by each section;
the bounding box loss specifically includes: loss of center point LosscenterBounding box size classification Losssize_classAnd Loss of residue Losssize_resLost Loss of bounding box rotation classificationrota_classAnd Loss of residue Lossrota_resThe expression is:
LossBbox＝Losscenter+Losssize_class+Losssize_res+Lossrota_class+Lossrota_res。
9. the method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 1, wherein the method comprises the following steps: generating a candidate defect target proposal in the step (5) and obtaining the candidate defect target proposal by adopting a fifth MLP network, wherein the fifth MLP network consists of three one-dimensional convolution layers with convolution kernel size of 1 and ReLU functions arranged among the convolution layers; the 3D-NMS is inhibited from removing the repeated bounding boxes in the candidate target by using the three-dimensional non-maximum value, and the IoU threshold of the 3D-NMS is 0.25.
10. The method for detecting the surface defects of the lithium battery pole piece based on the point cloud data as claimed in claim 1, wherein the method comprises the following steps: when the multi-scale features are integrated, four scales are provided, namely, an xyz coordinate containing point cloud points of 4096, 2048, 1024 and 512 and a high-dimensional feature FL corresponding to the point cloud points after the point cloud points are subjected to attention convolution of each layer of graph1f、FL2f、FL3f、FL4f。
CN202010401330.7A 2020-05-12 2020-05-12 Point cloud data-based lithium battery pole piece surface defect detection method Active CN111539949B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010401330.7A CN111539949B (en) 2020-05-12 2020-05-12 Point cloud data-based lithium battery pole piece surface defect detection method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010401330.7A CN111539949B (en) 2020-05-12 2020-05-12 Point cloud data-based lithium battery pole piece surface defect detection method

Publications (2)

Publication Number Publication Date
CN111539949A true CN111539949A (en) 2020-08-14
CN111539949B CN111539949B (en) 2022-05-13

Family

ID=71977663

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010401330.7A Active CN111539949B (en) 2020-05-12 2020-05-12 Point cloud data-based lithium battery pole piece surface defect detection method

Country Status (1)

Country Link
CN (1) CN111539949B (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112085123A (en) * 2020-09-25 2020-12-15 北方民族大学 Point cloud data classification and segmentation method based on salient point sampling
CN112525106A (en) * 2020-10-23 2021-03-19 清华大学 Three-phase machine cooperative laser-based 3D detection method and device
CN112801059A (en) * 2021-04-07 2021-05-14 广东众聚人工智能科技有限公司 Graph convolution network system and 3D object detection method based on graph convolution network system
CN113095205A (en) * 2021-04-07 2021-07-09 北京航空航天大学 Point cloud target detection method based on improved Hough voting
CN113435461A (en) * 2021-06-11 2021-09-24 中国地质大学（武汉） Point cloud local feature extraction method, device, equipment and storage medium
CN114882024A (en) * 2022-07-07 2022-08-09 深圳市信润富联数字科技有限公司 Target object defect detection method and device, electronic equipment and storage medium

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102022114910A1 (en) 2022-06-14 2023-12-14 Deutsches Zentrum für Luft- und Raumfahrt e.V. Method, device and computer program for detecting defects and method and device for depositing fiber material

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015002114A1 (en) * 2013-07-03 2015-01-08 Mitsubishi Electric Corporation Method and apparatus for determining pose of object in scene
CN104463871A (en) * 2014-12-10 2015-03-25 武汉大学 Streetscape facet extraction and optimization method based on vehicle-mounted LiDAR point cloud data
CN105809118A (en) * 2016-03-03 2016-07-27 重庆中科云丛科技有限公司 Three-dimensional object identifying method and apparatus
US20180322623A1 (en) * 2017-05-08 2018-11-08 Aquifi, Inc. Systems and methods for inspection and defect detection using 3-d scanning
CN110348310A (en) * 2019-06-12 2019-10-18 西安工程大学 A kind of Hough ballot 3D colour point clouds recognition methods
CN110390302A (en) * 2019-07-24 2019-10-29 厦门大学 A kind of objective detection method
CN110674829A (en) * 2019-09-26 2020-01-10 哈尔滨工程大学 Three-dimensional target detection method based on graph convolution attention network
CN110827302A (en) * 2019-11-14 2020-02-21 中南大学 Point cloud target extraction method and device based on depth map convolutional network
CN110991534A (en) * 2019-12-03 2020-04-10 上海眼控科技股份有限公司 Point cloud data processing method, device, equipment and computer readable storage medium

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015002114A1 (en) * 2013-07-03 2015-01-08 Mitsubishi Electric Corporation Method and apparatus for determining pose of object in scene
CN104463871A (en) * 2014-12-10 2015-03-25 武汉大学 Streetscape facet extraction and optimization method based on vehicle-mounted LiDAR point cloud data
CN105809118A (en) * 2016-03-03 2016-07-27 重庆中科云丛科技有限公司 Three-dimensional object identifying method and apparatus
US20180322623A1 (en) * 2017-05-08 2018-11-08 Aquifi, Inc. Systems and methods for inspection and defect detection using 3-d scanning
CN110348310A (en) * 2019-06-12 2019-10-18 西安工程大学 A kind of Hough ballot 3D colour point clouds recognition methods
CN110390302A (en) * 2019-07-24 2019-10-29 厦门大学 A kind of objective detection method
CN110674829A (en) * 2019-09-26 2020-01-10 哈尔滨工程大学 Three-dimensional target detection method based on graph convolution attention network
CN110827302A (en) * 2019-11-14 2020-02-21 中南大学 Point cloud target extraction method and device based on depth map convolutional network
CN110991534A (en) * 2019-12-03 2020-04-10 上海眼控科技股份有限公司 Point cloud data processing method, device, equipment and computer readable storage medium

Non-Patent Citations (7)

* Cited by examiner, † Cited by third party
Title
CAN CHEN等: ""GAPNet: Graph Attention based Point Neural Network for Exploiting Local Feature of Point Cloud"", 《ARXIV》 *
CHARLES R. QI等: ""Deep Hough Voting for 3D Object Detection in Point Clouds"", 《ARXIV》 *
ZHE LIU等: ""TANet: Robust 3D Object Detection from Point Clouds with Triple Attention"", 《ARXIV》 *
ZHENGLI ZHA等: ""Multi-Scale Dynamic Graph Convolution Network for Point Clouds Classification"", 《IEEE ACCESS》 *
张凯霖等: "复杂场景下基于C-SHOT特征的3D物体识别与位姿估计", 《计算机辅助设计与图形学学报》 *
汪汉云: ""高分辨率三维点云目标识别技术研究"", 《博士电子期刊》 *
罗欢: ""高分辨率激光扫描点云语义标注研究"", 《博士电子期刊》 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112085123A (en) * 2020-09-25 2020-12-15 北方民族大学 Point cloud data classification and segmentation method based on salient point sampling
CN112525106A (en) * 2020-10-23 2021-03-19 清华大学 Three-phase machine cooperative laser-based 3D detection method and device
CN112525106B (en) * 2020-10-23 2022-08-26 清华大学 Three-phase machine cooperative laser-based 3D detection method and device
CN112801059A (en) * 2021-04-07 2021-05-14 广东众聚人工智能科技有限公司 Graph convolution network system and 3D object detection method based on graph convolution network system
CN113095205A (en) * 2021-04-07 2021-07-09 北京航空航天大学 Point cloud target detection method based on improved Hough voting
CN113095205B (en) * 2021-04-07 2022-07-12 北京航空航天大学 Point cloud target detection method based on improved Hough voting
CN113435461A (en) * 2021-06-11 2021-09-24 中国地质大学（武汉） Point cloud local feature extraction method, device, equipment and storage medium
CN113435461B (en) * 2021-06-11 2023-07-14 深圳市规划和自然资源数据管理中心(深圳市空间地理信息中心) Point cloud local feature extraction method, device, equipment and storage medium
CN114882024A (en) * 2022-07-07 2022-08-09 深圳市信润富联数字科技有限公司 Target object defect detection method and device, electronic equipment and storage medium

Also Published As

Publication number Publication date
CN111539949B (en) 2022-05-13

Similar Documents

Publication Publication Date Title
CN111539949B (en) Point cloud data-based lithium battery pole piece surface defect detection method
CN111598861B (en) Improved Faster R-CNN model-based non-uniform texture small defect detection method
Saberironaghi et al. Defect detection methods for industrial products using deep learning techniques: A review
CN110457515B (en) Three-dimensional model retrieval method of multi-view neural network based on global feature capture aggregation
CN111339935B (en) Optical remote sensing picture classification method based on interpretable CNN image classification model
Wu et al. Using YOLOv5 for garbage classification
CN113705631B (en) 3D point cloud target detection method based on graph convolution
CN110569926B (en) Point cloud classification method based on local edge feature enhancement
CN110532946A (en) A method of the green vehicle spindle-type that is open to traffic is identified based on convolutional neural networks
CN111028238B (en) Robot vision-based three-dimensional segmentation method and system for complex special-shaped curved surface
CN115439694A (en) High-precision point cloud completion method and device based on deep learning
Wu et al. Underwater trash detection algorithm based on improved YOLOv5s
CN115170816A (en) Multi-scale feature extraction system and method and fan blade defect detection method
CN114373099A (en) Three-dimensional point cloud classification method based on sparse graph convolution
CN113989291A (en) Building roof plane segmentation method based on PointNet and RANSAC algorithm
CN115719445A (en) Seafood identification method based on deep learning and raspberry type 4B module
CN115032648A (en) Three-dimensional target identification and positioning method based on laser radar dense point cloud
CN115526863A (en) Cylindrical lithium battery surface defect detection method and device
CN117495891B (en) Point cloud edge detection method and device and electronic equipment
Yanmin et al. Research on ear recognition based on SSD_MobileNet_v1 network
CN112668662B (en) Outdoor mountain forest environment target detection method based on improved YOLOv3 network
CN114120066A (en) Small sample steel surface defect classification method based on lightweight network
Yang et al. A lightweight defect detection algorithm of insulators for power inspection
Yang et al. Detection of defects in voltage-dependent resistors using stacked-block-based convolutional neural networks
CN111967579A (en) Method and apparatus for performing convolution calculation on image using convolution neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant