Search

CN-121982379-A - Basalt fiber 3D printing-oriented defect detection method

CN121982379ACN 121982379 ACN121982379 ACN 121982379ACN-121982379-A

Abstract

The invention relates to a basalt fiber 3D printing-oriented defect detection method, and belongs to the technical field of defect detection. The method comprises the steps of constructing a defect data set containing fiber path deviation, accurately marking, designing and constructing MABD-Net defect detection models, training MABD-Net models by using training sets, optimizing training parameters of the models, improving detection accuracy, and applying the optimized MABD-Net models to a basalt fiber 3D printer to detect defects in real time. The MABD-Net model provided by the invention can efficiently identify and accurately position the micro defects through an innovative feature extraction and fusion mechanism, particularly when the detail problems such as complex fiber path deviation and the like are processed, the invention has excellent adaptability and precision, and has remarkable application prospect and industrial value.

Inventors

  • YUE ZHIJIAN
  • YIN HONGYU
  • YANG XIAOXIA
  • ZHOU XIAOHUI
  • HUANG JIANGPENG
  • WANG JIAMING
  • Gao Xuanbing
  • YANG YUZE
  • DENG YUHANG
  • HE YINDA
  • GAO ZIQIAN

Assignees

  • 重庆航天火箭电子技术有限公司

Dates

Publication Date
20260505
Application Date
20260108

Claims (10)

  1. 1. A basalt fiber 3D printing-oriented defect detection method is characterized by comprising the steps of constructing a defect data set containing fiber path deviation, marking and dividing the data set, establishing MABD-Net defect detection models, training the models by using divided training sets, optimizing model parameters in the training process, applying the optimized MABD-Net defect detection models to a basalt fiber 3D printer for real-time defect detection, setting at least a feature extraction network, a feature fusion network and a detection head network in the MABD-Net defect detection models, The feature extraction network is combined with the enhanced lightweight convolutional neural network MobileNetV and the SPPCSPC-ATT module to perform multi-scale feature extraction, wherein two C2f modules are introduced into the MobileNetV network, each C2f module extracts features with different scales and outputs the features to the feature fusion network and the SPPCSPC-ATT module, the SPPCSPC-ATT module enhances target edge information by introducing SimAM attention mechanisms, long-distance dependence modeling capability among BiFormer attention mechanism enhanced features is introduced, and the output features are also input into the feature fusion network; The feature fusion network comprises a SE-AdaptiveConv module, a BiFPN-Concat module, an up-sampling module and a C2f-AdaptiveConv module, wherein the BiFPN-Concat module receives features processed by the SE-AdaptiveConv module and the up-sampling module and features output by the feature extraction network to perform bidirectional feature fusion, and the features are processed by the C2f-AdaptiveConv module, wherein the SE-AdaptiveConv module introduces an adaptive adjustment mechanism to optimize convolution kernel weights in real time according to path data features; the detection head network combines the decoupling detection head and the SE-AdaptiveConv module to independently process the positioning and classifying tasks on different feature maps so as to finish defect detection.
  2. 2. The method for detecting defects in basalt fiber-oriented 3D printing of claim 1, wherein the input image of the feature extraction network is represented as The feature map is output through a 2D convolution layer, batch normalization and HARDSWISH activation functions Feature map After being processed by two bottleneck layers Bneck and a C2f module, a first scale characteristic diagram is obtained First scale feature map The feature fusion network is directly input, and a second scale feature map is obtained through the three bottleneck layers Bneck and the C2f module And, likewise, a second scale feature map The third scale feature map is obtained by directly inputting the feature fusion network and processing the feature fusion network through a bottleneck layer Bneck and a SPPCSPC-ATT module Third scale feature map A direct input feature fusion network, wherein, The input features of SPPCSPC-ATT modules are sequentially processed by two CBS blocks, then processed by a SimAM mechanism, the input features are processed by three branches to be subjected to maximum average pooling, the characteristics output by Concat blocks and SimAM blocks are spliced, the spliced features are spliced with the input features processed by one CBS block after being processed by two CBS blocks, and the spliced features are processed by one CBS block to obtain final output by a BiFormer mechanism.
  3. 3. The method for detecting the defects of the basalt fiber-oriented 3D printing of claim 2, wherein the SimAM mechanism of the SPPCSPC-ATT module is processed as follows: the attention mechanism is first implemented using a nonlinear function, wherein the importance of each neuron is estimated by computing an energy function, which is defined as: In the formula, And Is that And In the form of a linear transformation of (a), The weights and offsets of the linear transformations respectively, Representing the values to which the target neuron and other neurons are expected to be mapped, Representing the number of neurons on a channel, Is an index in the spatial dimension; The minimum energy is then calculated: In the formula, And Representing the mean and variance of the channel features in the spatial dimension respectively, Is a super parameter; the zoom operation refines and enhances features to adjust the attention mechanism: In the formula, The input characteristic diagram is represented by a graph of the input characteristics, Representing all channels and spatial dimensions In the summary of (a), Is a normalization function.
  4. 4. The method for detecting the defects of the basalt fiber-oriented 3D printing of claim 2, wherein the BiFormer attention mechanism of the SPPCSPC-ATT module comprises the following processing steps: First, a feature map is input Is divided into Non-overlapping blocks of size, each region comprising Individual feature vectors and map the feature images Mapping to queries, keys, and value matrices: Is a learned weight matrix, and the shape of the projected query, key and value matrix is ; Then, for the query And key Averaging the matrix to obtain an aggregate representation of the region level And And calculate the correlation between the regions and obtain the adjacency matrix : Next, for the adjacent matrix Performing row-by-row Top-k operation, extracting the most relevant neighboring region of each query, and obtaining an index matrix : Wherein, the Is an index matrix representing the most relevant of each query A plurality of regions; then the related keys are aggregated according to the index matrix Sum value Matrix, get the gathered matrix And : Finally, based on the aggregated matrix And Standard attention operations are performed to calculate the final output: Wherein the attention is operated Is to 、 And Weighted summation is performed and a loss function parameterized by a depth convolution is introduced Optimizing Is characterized by the following.
  5. 5. The method for detecting the defects of basalt fiber-oriented 3D printing of claim 2, wherein the third scale features received by the feature fusion network The second scale characteristics are processed by the SE-AdaptiveConv module and the up-sampling module and then input into one BiFPN-Concat module, and the BiFPN-Concat module and the second scale characteristics processed by the SE-AdaptiveConv module Performing bidirectional fusion, and further processing by C2f-AdaptiveConv to obtain primary fusion characteristics ; Preliminary fusion features Then after being processed by the SE-AdaptiveConv module and the up-sampling module, the sample is processed by the BiFPN-Concat module and the SE-AdaptiveConv module to obtain the first scale characteristic Performing bidirectional fusion, and further processing by C2f-AdaptiveConv to obtain a first fusion characteristic ; First fusion feature Inputting into a detection head network for defect detection, processing again by a SE-AdaptiveConv module, and fusing with a primary feature by a BiFPN-Concat module Fusing and further processing by C2f-AdaptiveConv to obtain a second fused feature ; Second fusion feature Directly inputting the processed data into a detection head network for defect detection, processing the processed data again by a SE-AdaptiveConv module, and extracting the processed data and the third scale characteristics output by a characteristic extraction network by a BiFPN-Concat module Fusing and further processing by C2f-AdaptiveConv to obtain a third fused feature Third fusion feature Directly inputting the defect detection data into a detection head network for defect detection.
  6. 6. The method for detecting defects in basalt fiber-oriented 3D printing of claim 5, wherein the SE-AdaptiveConv module comprises the following steps: fiber path profile for a given input Its spatial dimensions are compressed into a global feature vector using adaptive averaging pooling: routing networks through linear transformations Global feature vector Mapping into a set of routing weights w: Wherein, the Is of the size of A matrix of weights for the linear transformation, In order to input the number of channels of the feature map, For the number of expert convolution kernels, As a result of the bias term, Is a Sigmoid activation function for limiting the weight to be within the (0, 1) range; Routing weights Determining each expert convolution kernel Contribution to the final convolution operation in a conditional convolution layer, multiple expert convolution kernels Is pre-trained and the final adaptive convolution kernel And dynamically combining according to the routing weight w to obtain: using the resulting adaptive convolution kernel For input fiber path feature map Performing convolution operation to generate an output feature map : The generated feature map Y is input into the SE attention mechanism module, which dynamically adjusts the attention feature.
  7. 7. The basalt fiber 3D printing-oriented defect detection method of claim 5, wherein the C2f-AdaptiveConv module fuses the C2f module and the AdaptiveConv module, the C2f module extracts features by using multi-scale convolution operation and fuses the features to improve the processing capacity of a network on multi-scale information, and the AdaptiveConv module is embedded in a bottleneck position of the network to dynamically adjust convolution kernel weights so that the network is more flexible when facing different inputs.
  8. 8. The basalt fiber 3D printing-oriented defect detection method of claim 5, wherein the BiFPN-Concat module performs bidirectional feature fusion, makes full use of feature information of different scales, the up-sampling module is used for recovering details of low-resolution feature images, ensuring close combination of low-level and high-level features to capture global semantic information and transmit the global semantic information to lower-level feature images, the down-sampling operation reduces spatial resolution of the low-level feature images and fuses the low-level feature images with the high-level feature images to capture detailed information and transmit the detailed information to higher-level feature images, the feature fusion is achieved through a weighted summation operation in BiFPN, each fusion node distributes a leachable weight for the input feature images and then is used for weighted summation to promote the feature fusion, and the weighted fusion operation is as follows: Wherein, the Is a weight that can be learned and is, Is an input feature map which is used to input a feature map, Is a minimum number for avoiding zero divide errors.
  9. 9. The method for detecting basalt fiber-oriented 3D printing defects according to claim 5, wherein the head network is combined with the decoupling detection head and the SE-AdaptiveConv module, and the SE-AdaptiveConv is used for independently processing positioning and classifying tasks on different feature graphs by dynamically adjusting input parameters, and the features are fused with first fused features output by the network Second fusion feature And a third fusion feature And respectively processing by the corresponding SE-AdaptiveConv modules to finish defect detection.
  10. 10. The basalt fiber 3D printing-oriented defect detection method of claim 9, wherein training is performed on MABD-Net defect detection models, and a loss function is used for jointly optimizing a boundary box loss, a classification loss and a confidence loss, wherein the boundary box loss adopts a normalized weighted distance loss function, and the normalized weighted distance loss function is expressed as: In the formula, Representing the normalized wasperstein distance, And Respectively represented by bounding boxes And A modeled gaussian distribution; The classification loss and the confidence loss use a binary cross entropy loss function to measure the consistency of the prediction probability and the real label so as to distinguish different defect categories, and the mathematical expression is as follows: Wherein, the Representing the total number of bounding boxes in the batch, Represent the first True labels of the bounding boxes, 1 indicating the presence of a defect, 0 indicating the absence; representing the model's predicted probability of the defect type in the bounding box.

Description

Basalt fiber 3D printing-oriented defect detection method Technical Field The invention belongs to the technical field of defect detection, and relates to a basalt fiber 3D printing-oriented defect detection method. Background 3D printing, a leading edge additive manufacturing technology, is becoming the focus of many field applications. The core principle is that the three-dimensional digital model is directly converted into a solid product by stacking materials layer by layer. Compared with the traditional material reduction manufacturing method, the 3D printing technology has obvious advantages in the aspects of manufacturing complex structures, realizing personalized customization, rapid forming and the like. The advantage enables the 3D printing technology to develop rapidly in a short time, and to be widely applied to various fields such as aerospace, national defense, medical treatment, industrial manufacturing and the like, and has great market potential and technical prospect. To meet the different application requirements and complex usage scenarios, the variety of 3D printing materials is gradually expanding, and various material systems including plastics, metals, ceramics and biological materials have been developed from the original thermoplastics. In recent years, along with the further development of 3D printing technology, the introduction of CFRCs into a 3D printing process has become a research direction of great interest. CFRCs are unique in that they have excellent mechanical properties and environmental suitability, and can meet specific engineering requirements of high strength, light weight, durability, etc. The CFRCs is applied to 3D printing, so that the manufacture of high-strength fiber reinforced parts can be realized, the printed product has excellent structural strength and toughness, and the method is widely applied to industrial and engineering scenes requiring wear resistance, compression resistance and light weight. In the CFRCs 3D printing process, the matrix material is typically heated to a liquid state in a hot melt nozzle, impregnating the continuous fibers and extruding them together. Then, the matrix material and the reinforcing fibers are printed layer by layer on a printing platform, and the required fiber reinforced structural component is formed after cooling and solidification. However, this process is extremely sensitive to the setting of printing parameters, and improper parameter setting may result in significant deviations between the fiber lay path and the intended path. For example, too high a printing speed may result in the fibers not being fully impregnated in the matrix material, and improper temperature setting may affect the flowability of the matrix material and the adhesion of the fibers. In addition, printing accuracy problems, equipment calibration deficiencies, and the like may affect the accuracy of the printing path, thereby causing problems such as delamination between material layers, fiber misalignment, and the like, and ultimately resulting in serious impact on the structural integrity and functionality of the printed product. Disclosure of Invention In view of the above, the invention aims to provide a basalt fiber 3D printing-oriented defect detection method which can accurately identify and position tiny defects, and particularly in practical application, the method has adaptability and high efficiency for detail problems such as complex fiber path deviation. In order to achieve the above purpose, the present invention provides the following technical solutions: a basalt fiber 3D printing-oriented defect detection method comprises the steps of constructing a defect data set containing fiber path deviation, marking and dividing the data set, establishing MABD-Net defect detection models, training the models by using divided training sets, optimizing model parameters in the training process, applying the optimized MABD-Net defect detection models to a basalt fiber 3D printer for real-time defect detection, setting at least a feature extraction network, a feature fusion network and a detection head network in a MABD-Net defect detection model, The feature extraction network is combined with the enhanced lightweight convolutional neural network MobileNetV and the SPPCSPC-ATT module to perform multi-scale feature extraction, wherein two C2f modules are introduced into the MobileNetV network, each C2f module extracts features with different scales and outputs the features to the feature fusion network and the SPPCSPC-ATT module, the SPPCSPC-ATT module enhances target edge information by introducing SimAM attention mechanisms, long-distance dependence modeling capability among BiFormer attention mechanism enhanced features is introduced, and the output features are also input into the feature fusion network; The feature fusion network comprises an SE-AdaptiveConv module, a BiFPN-Concat module, an up-sampling module and a C2f-AdaptiveConv modu