Method for realizing small sample automobile simulation data clustering based on key perspective fusion

Abstract

A clustering method for small-sample automotive simulation data based on key perspective fusion is proposed. This method involves constructing a multi-view neural network feature extraction model to extract image features from each viewpoint in the sample set. The extracted feature vectors are then used as input, and the K-means algorithm is applied to obtain local clustering results. Subsequently, a multi-source decision fusion algorithm is used to fuse the local clustering results from different viewpoints to form a fused category representation. Global categories that meet the criteria are then selected from this fused category representation. Finally, the global categories are optimized under the guidance of the dominant viewpoint. This invention effectively integrates clustering results from different viewpoints and performs dominant optimization based on the key viewpoint, thereby improving the consistency and accuracy of the clustering results.

Description

Technical Field

[0001] This invention relates to a technology in the field of automobile manufacturing, specifically a method for clustering small-sample automobile simulation data based on key perspective fusion. Background Technology

[0002] In current automobile manufacturing processes, clustering simulation data can automatically identify structural features under different operating conditions, thereby optimizing design parameters and improving vehicle safety and reliability. However, existing clustering methods are often based on a single perspective and have poor adaptability to complex datasets, especially when dealing with multi-perspective simulation data. The main problems are as follows: inconsistencies between perspectives: clustering results from different perspectives may differ significantly or even conflict, leading to poor accuracy and stability of the overall clustering results; the importance of key perspectives is overlooked: in multi-perspective data analysis, some perspectives may have higher weight and dominance in the clustering results. If key perspectives are not effectively emphasized and optimized, the final performance of the clustering results will be affected. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a clustering method for small-sample automotive simulation data based on key perspective fusion. This method effectively integrates clustering results from different perspectives and performs dominant optimization on key perspectives, thereby improving the consistency and accuracy of clustering results.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a method for clustering small-sample automotive simulation data based on key perspective fusion. The method involves constructing a multi-view neural network feature extraction model to extract image features from each viewpoint in the sample set. The extracted image feature vectors are then used as input, and the K-means algorithm is applied to obtain local clustering results. Subsequently, a multi-source decision fusion algorithm is used to fuse the local clustering results from different viewpoints to form a fused category representation. Global categories that meet the criteria are then selected from the fused category representation. Finally, the obtained global categories are optimized under the guidance of the dominant viewpoint.

[0006] The multi-view neural network feature extraction model includes a low-level feature extraction unit and a high-level feature extraction unit. The low-level feature extraction unit performs feature extraction processing based on the convolutional layer, batch normalization layer, activation function layer and pooling layer in the pre-trained ResNet50 model according to the original image information, and outputs a low-level feature representation of the image. The high-level feature extraction unit extracts high-level semantic features of the image based on the image feature map extracted by the low-level feature extraction, and generates the final feature representation by using high-order residual blocks in the pre-trained ResNet152 model.

[0007] The sample set Where: M is the size of the sample set, and each sample... Images containing three different perspectives: X, Y, and Z.

[0008] The K-means algorithm refers to dividing the dataset into K clusters, iteratively selecting cluster centers and assigning samples to the nearest cluster center until a termination condition is met, minimizing the distance between points within each cluster and maximizing the distance between different clusters. Specifically: ,in: . Let be the feature vector of the i-th data point; Let j be the feature vector of the j-th cluster center; For data points With cluster center The Euclidean distance between them; Let be the feature value of the i-th data point in the l-th dimension; Let be the feature value of the j-th cluster center in the l-th dimension.

[0009] The local clustering results, i.e. the category labels for each viewpoint, specifically refer to the range of category label values ​​for each viewpoint and the value of K in the K-means algorithm.

[0010] In this invention, K is preferably 3, and the clustering result for each perspective takes a range of values. .in It represents three perspectives.

[0011] The fusion category, the set of fusion representations of the local categories of all samples, i.e., the multi-view information of the fusion samples, is denoted as... Where 'a' represents the number of fusion categories.

[0012] The global category This refers to the final category of the sample. Based on the global category construction criteria, category labels that meet the conditions are selected from the fused categories. The range of global category label values ​​for each sample is related to the global category construction criteria set by the criteria, denoted as . .

[0013] The global category construction criteria include:

[0014] 1) The global categories each contain the value range of the clustering results from the main perspective direction;

[0015] 2) Under the premise of satisfying condition 1, the three fusion categories with the largest number of samples are the constructed global categories.

[0016] The aforementioned optimization of global results under the guidance of the main perspective refers to: adding the previously selected fusion categories back to the global categories according to the following conditions:

[0017] 1) If the results of non-subjective local clustering are the same as those of the global category, then the cluster is assigned to the global category;

[0018] 2) If condition 1) is not met, the fusion category will be assigned to the global category that is the same as the local category result of its own main perspective.

[0019] Technical effect

[0020] Compared with existing technologies, this invention utilizes a multi-view neural network feature extraction model and a multi-source decision fusion algorithm to learn features from multiple perspectives, effectively capturing multi-dimensional information from the data. Combined with the multi-source decision fusion algorithm, it overcomes the limitations of single-view clustering and achieves weighted fusion of information from different perspectives, thereby improving the accuracy, stability, and robustness of the clustering results. This algorithm can reasonably allocate the contributions of each perspective, ensuring high accuracy and consistency in the final clustering decision, and effectively reducing the interference of noise and outliers. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system of the present invention;

[0022] Figure 2 Build a flowchart for the global category. Detailed Implementation

[0023] like Figure 1 As shown in this embodiment, a small-sample car simulation data clustering system based on key perspective fusion is implemented. The system includes: a multi-view feature extraction module, a viewpoint local clustering module, a category fusion module, a global category construction module, and a master-view guided clustering result optimization module. Specifically: the multi-view feature extraction module receives a multi-view car simulation image dataset, constructs a multi-view neural network model based on a pre-trained convolutional neural network, and extracts features from the data of each viewpoint to provide basic feature data for subsequent clustering tasks; the viewpoint local clustering module, based on feature extraction, uses the K-means clustering algorithm to independently cluster the feature data of each viewpoint, generating local clustering results for each viewpoint; the category fusion module fuses the information of the local clustering results of each sample under different viewpoints to form a fused category for the sample; the global category construction module constructs a global category that meets the conditions from the fused categories according to the set global category construction criteria; and the master-view guided clustering result optimization module optimizes the global clustering results based on the constructed global categories under the key perspective guiding criteria.

[0024] like Figure 2As shown, this embodiment implements a clustering method for small sample car simulation data based on the above system, including:

[0025] S1. Read the multi-view simulation image dataset M represents the sample set size. This embodiment uses a multi-view simulation image dataset generated by a car manufacturer during the vehicle development process. This dataset contains 70 samples, corresponding to 70 simulation instances of the vehicle development process. Each sample... The image shows three mutually perpendicular views of the simulation model, reflecting the deformation of the vehicle's transverse and longitudinal beam structure. Among them is... First-person perspective image and It is a non-first-person perspective image.

[0026] S2, Use Input a multi-view neural network feature extraction model and output the final feature set for each view. .

[0027] S3, will Input the low-level feature extraction units of the multi-view neural network feature extraction model, and output a set of low-level feature maps for each view. Then... Input the high-level feature extraction units of the multi-view neural network feature extraction model, and output the final feature sets of each viewpoint. .

[0028] S4. Extract feature subsets from the feature set that share the same perspective. , and The K-means algorithm is used for preliminary clustering to obtain local clustering results for each viewpoint. Each sample has three local clustering results. , and .in .

[0029] S5. Integrate the local clustering results of each sample from different perspectives to obtain the fusion category, specifically including:

[0030] S51. Merge the local clustering results of each sample to form a fused representation of the local categories of that sample. .

[0031] S52. Organize the fusion representations of the local categories of all samples to form a fusion category set. .

[0032] S6. Based on the established global category construction criteria, extract the global categories that meet the conditions from the fusion categories, specifically including:

[0033] S61, on the set of fusion categories The data is classified into three fusion categories, each with... of If the values ​​are the same, it means that the samples have the same clustering results in the X direction.

[0034] S62. Sort the fusion categories of each cluster according to the sample size, and use the fusion category with the largest sample size for each cluster. Defined as a global category.

[0035] S7. Based on the established global category construction criteria, from Extract the fusion categories that meet the criteria as the global category. .

[0036] S8. Under the guidance of the main perspective, optimize the global results by assigning the remaining fused categories to three global categories to construct a complete global category. Assign a global category to each sample. Finally, the clustering results are fused and optimized using multi-perspective information.

[0037] To further illustrate the effectiveness of this method, three methods were compared: (1) clustering using only the main view X image; (2) clustering using only the Y view image; and (3) clustering using only the Z view image. The Land coefficients of the above methods were calculated to evaluate the clustering performance of each method.

[0038] Table 1

[0039]

[0040] As shown in Table 1, for the clustering problem of multi-view car simulation images, compared with using only a single view, the performance index of the Land coefficient of this method is improved by 0.0662~0.1308, which shows that the present invention significantly enhances the clustering prediction ability of multi-view car simulation images.

[0041] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for clustering small-sample automotive simulation data based on key perspective fusion, characterized in that, After extracting image features from each perspective in the sample set by constructing a multi-view neural network feature extraction model, the final feature vector of the extracted image is used as input. After obtaining local clustering results by K-means algorithm, the local clustering results under different perspectives are fused to form a fused category representation by multi-source decision fusion algorithm. Global categories that meet the conditions are obtained by global category screening from the fused category representation. Then, the global results are optimized under the guidance of the main perspective. The multi-view neural network feature extraction model includes a low-level feature extraction unit and a high-level feature extraction unit. The low-level feature extraction unit performs feature extraction based on the original image information, using convolutional layers, batch normalization layers, activation function layers, and pooling layers in a pre-trained ResNet50 model, outputting a low-level feature representation of the image. The high-level feature extraction unit, based on the image feature map extracted from the low-level feature extraction, extracts high-level semantic features of the image through high-order residual blocks in a pre-trained ResNet152 model, generating the final feature representation. The sample set Where: M is the size of the sample set, and each sample... Images containing three different perspectives: X, Y, and Z; The fusion category, the set of fusion representations of the local categories of all samples, i.e., the multi-view information of the fusion samples, is denoted as... Where: a is the number of fusion categories; The global category This refers to the final category of the sample. Based on the global category construction criteria, category labels that meet the conditions are selected from the fused categories. The range of global category label values ​​for each sample is related to the global category construction criteria set by the criteria, denoted as . ; The global category construction criteria include: 1) The global categories each contain the value range of the clustering results from the main perspective direction; 2) Provided that condition 1) is met, the three fusion categories with the largest number of samples are the constructed global categories; The aforementioned optimization of global results under the guidance of the main perspective refers to: adding the previously selected fusion categories back to the global categories according to the following conditions: 1) If the results of non-subjective local clustering are the same as those of the global category, then the cluster is assigned to the global category; 2) If condition 1) is not met, the fusion category will be assigned to the global category that is the same as the local category result of its own main perspective.

2. The method for clustering small-sample car simulation data based on key perspective fusion according to claim 1, characterized in that, The K-means algorithm refers to dividing the dataset into K clusters, iteratively selecting cluster centers and assigning samples to the nearest cluster center until a termination condition is met, minimizing the distance between points within each cluster and maximizing the distance between different clusters. Specifically: ,in: , Let be the feature vector of the i-th data point; Let j be the feature vector of the j-th cluster center; For data points With cluster center The Euclidean distance between them; Let be the feature value of the i-th data point in the l-th dimension; Let be the feature value of the j-th cluster center in the l-th dimension.

3. The method for clustering small-sample car simulation data based on key perspective fusion according to claim 2, characterized in that, The local clustering results, i.e. the category labels for each viewpoint, specifically refer to the range of category label values ​​for each viewpoint and the value of K in the K-means algorithm.

Images