	US 12,236,670 B1
	Classification method for product quality accidents and its system based on convolutional neural networks
Jingxing Liao, Beijing (CN); Jingna Yang, Beijing (CN); and Yuwei Lu, Beijing (CN)
Assigned to China National Institute of Standardization, Beijing (CN)
Filed by China National Institute of Standardization, Beijing (CN)
Filed on Jul. 19, 2024, as Appl. No. 18/778,875.
Int. Cl. G06V 10/771 (2022.01); G06Q 10/0637 (2023.01); G06T 7/00 (2017.01); G06V 10/82 (2022.01)

CPC G06V 10/771 (2022.01) [G06Q 10/06375 (2013.01); G06T 7/001 (2013.01); G06V 10/82 (2022.01); G06T 2207/30108 (2013.01)]

10 Claims

1. A classification method for product quality accidents based on convolutional neural networks, characterized in that:

collecting the sample data of product quality accidents, labeling the products that suffered quality accidents as the products with quality accidents, respectively, and thus the accident severity evaluation indices of various products with quality accidents can be generated by computation;

extracting the image feature elements of each product with quality accident, and the evaluation values of accident relevance of various image feature elements can be obtained by analyzing based on the accident severity evaluation index for each product with quality accident, after that, a valid feature element screening is performed according to the evaluation values of accident relevance of various image feature elements to obtain various valid feature elements;

then, a product quality accident classification model can be generated through training various valid feature elements, after that, getting a target product, and thus the product quality accident is classified according to the product quality accident classification model;

the specific analysis process for obtaining the said accident severity evaluation index for each product with quality accident consists of:

extracting the initial images and images of quality accidents of each product with quality accident based on the sample data of product quality accidents, and then comparing the initial images with the images of quality accidents of each product with quality accident, and thus a quantitative index for the degree of damage of each product with quality accident can be generated by analyzing;

obtaining the cost data of repairing each product with quality accident and calculating the time it took to repair, and thus a quantitative index for the impact degree of each product with quality accident can be generated after processing and computation;

then, the accident severity evaluation index for each product with quality accident can be generated by a comprehensive analysis;

the specific analysis process for obtaining the evaluation values of accident relevance of various image feature elements consists of:

obtaining the image entropy value and the image contrast index for each product with quality accident through processing based on the image of quality accident of each product with quality accident, and thus the feature quantification values of various image feature elements of each product with quality accident can be obtained by a comprehensive computation;

based on the accident severity evaluation index for each product with quality accident and the feature quantification values of various image feature elements of each product with quality accident, the evaluation values of accident relevance of various image feature elements can be obtained by computation the specific calculation expression of the accident severity evaluation index for each product with quality accident is:

wherein, δ_irefers to the accident severity evaluation index for the i^thproduct with quality accident, e refers to a natural constant, φ_S→irefers to the quantitative index for the damage degree of the i^thproduct with quality accident, and φ_Y→irefers to the quantitative index for the impact degree of the i^thproduct with quality accident, ξ₁refers to the accident severity impact factor corresponding to the given quantitative index for the damage degree, as well as ξ₂refers to the accident severity impact factor corresponding to the given quantitative index for the impact degree;

the specific calculation expression of the feature quantification values of various image feature elements of each product with quality accident is:

wherein, χ_irrefers to the feature quantification value of the r^thimage feature element of the i^thproduct with quality accident, e refers to the natural constant, H(s)_irefers to the image entropy value of the i^thproduct with quality accident H(D)_irefers to the image contrast index for the i^thproduct with quality accident, τ_S→rrefers to the quantitative feature impact factor of the given r^thimage feature element corresponding to the image entropy value, and τ_D→rrefers to the quantitative feature impact factor of the given r^thimage feature element corresponding to the image contrast index;

the specific calculation expression of calculating the quantitative index for the degree of damage of each product with quality accident is:

wherein, φ_S→irefers to the quantitative index for the degree of damage of the i^thproduct with quality accident, P_S→ijrefers to the pixel value of the j^thpixel shows in the image of quality accident of the i^thproduct with quality accident, P_C→ijrefers to the pixel value of the j^thpixel shows in the initial image of the i^thproduct with quality accident, and ΔP refers to the given allowable deviation pixel value, and ζ refers to the given quantitative correction factor of the damage degree of product, i refers to the number of each product with quality accident, and for i=1,2,3, . . . , n, n refers to the total number of products with quality accidents, j refers to the number of each pixel point, for j=1,2,3, . . . , m, m refers to the total number of pixel points;

the specific calculation expression of calculating the quantitative index for the impact degree of each product with quality accident is:

wherein, refers φ_Y→ito the quantitative index for the impact degree of the i^thproduct with quality accident, e refers to a natural constant, R_X→irefers to the cost of repairing of the i^thproduct with quality accident, R_J→irefers to the time cost of repairing the i^thproduct with quality accident, R_X→0refers to the predefined critical cost of repairing, and R_J→0refers to the predefined critical duration of repairing, ψ₁refers to the quantified weight factor of the impact degree corresponding to the given cost of repairing, ψ₂refers to the quantified weight factor of the impact degree corresponding to the given time cost of repairing, and ψ₃refers to the quantified weight factor of the impact degree corresponding to the given unit value of cost of repairing, as well as ψ₄refers to the quantified weight factor of the impact degree corresponding to the given unit value of time cost of repairing;

the specific calculation expression of calculating the evaluation values of accident relevance of various image feature elements is:

wherein, F_rrefers to the evaluation value of the accident correlation degree of the r^thimage feature element, δ_irefers to the quantitative index for the impact degree of the i^thproduct with quality accident, χ_irrefers to the feature quantification value of the r^thimage feature element of the i^thproduct with quality accident, δ refers to the evaluation index for average accident severity of the product with quality accident, for

χ_rrefers to the average feature quantification value of the r^thimage feature element, and for

ç refers to the given accident correlation degree correction factor.