Wen 丨 Xiaofei is a little annoying
Editor丨Xiaofei is a little annoying
Virtual manufacturing is a simulation-based technique used to define, simulate, and visualize manufacturing processes during the design phase, where product defect detection is closely related to quality assurance.
In this context, the defect detection of mechanical gears is particularly important as a power transmission element widely used in industrial machinery such as turbines, motor vehicles and aircraft.
Although the inspection of 3D objects has been extensively studied, the detection of gear defects is still of critical importance in the field of virtual manufacturing because it enables timely capture of defect failures during manufacturing simulations.
However, in the actual industrial environment, gear defects are difficult to avoid, almost 80% of mechanical transmission system failures are caused by gear defects, which not only involves personal safety issues, but also leads to manufacturing and financial losses, so defect detection in mechanical systems is particularly necessary.
Mechanical rotation processes and their technical challenges
Researchers can monitor the operating state of rotating machinery by manually collecting the characteristics of vibration and acoustic emission signals, although the signal-based method has shown some effectiveness in detecting gear defects.
However, accurate physical models and signal processing experience are often required, which may not be sufficient to meet the requirements of modern industrial intelligence.
In fact, sensor data is the basis for fault detection, and researchers have traditionally analyzed defect characteristics by collecting information from different sensors.
However, since the defect vibration signal is obtained by the operation of the gear, the defect signal may be mixed in the strong meshing harmonics of various rotating parts.
Therefore, we want to dig deep into its excellent feature extraction and nonlinear approximation capabilities, which can show great advantages in the field of image classification and object detection.
In addition to the above methods, intelligent data-driven fault diagnosis technology has also gradually attracted attention, and researchers have proposed a deep learning method based on separation and fusion to analyze the multi-modal characteristics of gearbox vibration measurement and generate diagnostic results.
However, extracting and detecting the characteristics of gear modulated signals may not be practical for traditional methods.
But image-based computer vision methods can also be applied to defect detection, and in addition to converting vibration signals into grayscale images, researchers have also tried to use two-dimensional images of gears to identify their defects.
Due to the complex structure of the gear surface, especially the defects of the tooth surface, this method may be affected by texture interference and stains, which can affect the accurate identification of defects.
Compared with image-based methods, using 3D point cloud models with depth data can effectively avoid the influence of image texture or stains on gear defect recognition.
Some researchers proposed a point cloud network - PointNet, and then the deep learning network based on point cloud was successfully applied to 3D shape classification, object detection, tracking and 3D segmentation.
However, obtaining large-scale labeled point cloud data with defect information is the key to ensure the good detection performance of neural networks, but in reality, it is difficult to collect enough machine data, which becomes a constraint for intelligent fault diagnosis.
To address the lack of labeled data, the researchers employed transfer learning, semi-supervised, and unsupervised learning methods to obtain noise-free point clouds from computer-aided design (CAD) models of gears through virtual fabrication.
Therefore, it is necessary to use point cloud data for defect detection, and defective gear models have complex local structures, which can be completely represented by point clouds.
In the study, we discovered a new artificial neural network called Gear-PCNet++, which is based on point cloud data extracted from CAD models.
In this network, a novel combinatorial convolutional block (CCB) is introduced to replace the convolutional layer in the multilayer perception (MLP) network, so as to better extract the detailed features of gear defects.
Key contributions to this research include:
(1) A 3D gear model dataset containing four typical gear defects (fracture, pitting corrosion, gluing and wear) was established;
(2) By introducing CCB structure, the accuracy of defect detection is improved;
(3) A new network model, Gear-PCNet++, was developed to identify various types of gear defects with high accuracy.
These improvements and discoveries are expected to drive the application of gear defect detection technology in the field of virtual manufacturing, further improving the reliability and performance of mechanical systems.
The point cloud data can be obtained by 3D scanning, but the scanned original point cloud is difficult to accurately label the category, and a 3D gear data generation method is proposed based on the geometric characteristics of the gear.
Construction of a 3D gear sample set
In this study, gear defects are divided into four typical types: wear, pitting, gluing, and fracture, and gear defects can be represented as a combination of four typical defects, as shown in the figure below.
Point cloud data can be obtained by 3D scanning, but the scanned original point cloud is difficult to accurately label categories, and CAD models can be converted to point cloud models.
Although the CAD model of the gear has a large number of surface elements, the effective surface can still be used to build a point cloud dataset of 10,000 gear samples, some of which are shown in the figure below.
Gear defects often occur in similar locations on the tooth surface, and it is difficult to identify gear defects from the local features of the point cloud, for which boundary information is more important than other details.
In addition to the above methods, convolutional layer operation is also an efficient way to share parameters in artificial neural networks, which is widely used in various deep learning tasks.
Researchers have proposed a method called PointConv, which uses Monte Carlo approximation for convolution operations, suitable for unstructured point cloud data.
In addition, researchers have demonstrated that extended convolution and downsampling are effective methods for expanding data.
In this context, the adoption of K· D nearest neighbors replace the original k nearest neighbors to obtain a wider range of receptive fields, which, although somewhat similar to dilated convolution, may result in a loss of local feature detail.
PointNet++ adopts a neighborhood-based feature extraction method to replace the independent learning channel of each point, and multi-scale analysis is also another strategy to improve the effect of image semantic segmentation, which can enrich the feature information.
Not only that, the feature pyramid network is also one of the most common frameworks, through multi-scale or multi-level information interaction, this multi-scale synthesis strategy is applied to convolution operations, using relatively small convolution kernels to obtain rich feature information.
Specifically, convolution kernels of different sizes are used to extract features under different receptive fields before attaching them to the results of this module.
This strategy has been widely used in ResNet, GoogLeNet, and other architectures, where 1×1 convolution operations are often used to implement dimensional transformations to reduce the number of parameters.
In research, the basic geometric elements of gears include points, lines, and surfaces, with two points determining a straight line and three points determining a plane.
However, point cloud data is usually sparser compared to the original 3D model, and we can assume that the surface contains at least three points, two of which form the boundary line of the gear point cloud.
We can then use convolution operations with kernel sizes of 1, 2, and 3 to extract relevant geometric element information, which helps to resolve features into projected points, dashed lines, and imaginary faces to some extent.
It is important to note that point clouds entered into the network are often unordered, and in the example, the gears exhibit pitting and wear defects, which are represented by green and blue boxes, respectively.
Ppit-j, Pwear-i, and Pwear-k represent different points of pitting and wear, and for point Pwear-i, using larger convolution kernels makes it easier to extract features related to that point, especially when the point contributes less to the final result.
To ensure the effectiveness of extracting features, neighborhood based on distance definition of points is a general strategy that has been applied in many networks, such as PointNet++, SpiderCNN, and EdgeConv.
Due to the differences between the research in this paper and the above methods, we propose a distance-based optimization network architecture strategy to assign corresponding weights to the features extracted by convolution kernels of different sizes.
Network architecture
We propose a gear defect recognition network based on one-dimensional convolution operation: Gear-PCNet, which consists of feature extraction (CCB-MLP) and final classification module.
x, while CCB can output features containing single points and inter-point information, making point clouds rotational and translational invariance, projecting point cloud data into a 2D image or representing it as voxels may result in information loss.
In PointNet, the above two problems are dealt with using the maximum value, while in Gear-PCNet, the maximum and average functions are used to extract the features of the point cloud and connect them.
As mentioned above, the hierarchical feature learning framework is further applied to Gear-PCNet, and Gear-PCNet++ is constructed based on 2D convolution operations. The structure of Gear-PCNet++ is shown in the figure below.
Based on the data that builds the local area set, the dataset is relatively more concentrated, allowing the radius of the local region to be set smaller, and based on the above data, we can already make a simple discussion of the study.
Discussion of the study
The method was evaluated on a set of 10,000 samples (defective gears) whose characteristics can be divided into 5 types, basic gears, fracture, pitting corrosion, gluing and wear.
The 10,000 samples were divided into training set, validation set, and test set according to the ratio of 8:1:1, and the experiment ran the CPU on a PC equipped with the "NVIDIAGeForceRTX3070" GPU and "Intel [email protected]".
It can be analyzed that PointNet is a classic point cloud classification and segmentation network, and the number of parameters in Gear-PCNet is less than the number of parameters in PointNet (vanilla).
The effectiveness of Gear-PCNet was evaluated based on the classification performance of the three networks on the gear dataset, and the combined CCB in Gear-PCNet was replaced with the structure shown in the figure below to verify the superiority of the comprehensive feature information over the single feature information.
Through research, it can be found that CCB achieves good results by extracting richer features, and Gear-PCNet++ and several classical networks are tested on gear datasets.
Observing the prediction accuracy of the training and validation sets during training, it can be seen that Gear-PCNet++ and PointNet++ converge faster.
Although we got the results we wanted in this study, there are still some flaws in this study.
Defect identification discussion
The classification and prediction of Gear-PCNet and Gear-PCNet++ are given in the experimental results section, but the types and numbers of defects of different gear models are different.
Based on the performance of Gear-PCNet++ on the test sample (1000 gear models), analyzing the identification of defects and points in different models, we found that 97.90% of the model recognition accuracy was above 95.00%.
If the identification is successful, the defect type is considered correct, that is, there are 3 defects in the model, and when more than 10 points are marked for defects, there are defects.
Under the above settings, 99.90% of the models were correctly judged, which indicates that the recognition results have a high degree of confidence.
At the same time, the figure below shows the identification confusion matrix for each defect type in the test set, in which the confusion matrix is approximated as a diagonal matrix, which also shows that the method in this paper is accurate and effective.
The gear also has intersection defects, which makes it difficult to identify the point category, which can be divided into self-intersection of the same defect feature and intersection of different defect features, the figure below shows the intersection results of pitting holes and the intersection results of broken teeth and wear, respectively.
In Gear-PCNet, this intersection result may require many relevant samples to aid the training of the network, but it can be satisfied in Gear-PCNet++.
Virtual manufacturing is of great significance in the detection of defects in gear parts, and as a power transmission element widely used in industrial machinery, the defect detection of mechanical gears is essential to ensure product quality and safety.
Although existing 3D object detection methods have been extensively studied, the virtual manufacturing of gear defects still faces key challenges, as capturing gear defects in real time during the manufacturing process can prevent failures and reduce losses.
In the actual industrial environment, gear defects are unavoidable, and their failure leads to about 80% of problems in mechanical transmission systems.
This not only affects personal safety, but also causes production and financial losses, and the introduction of defect detection technology in mechanical systems is of crucial importance.
The virtual manufacturing gear defect detection technology based on point cloud data has broad application prospects in improving the reliability and performance of mechanical systems.
For future research, more data augmentation methods and more effective feature extraction strategies can be further explored to advance this field.