Product quality detection method in complex industrial process fusing prior knowledge

A product quality detection method in a complex industrial process fusing prior knowledge, comprising the following steps:

• S1:
• acquiring samples of product production in the complex industrial process;
• wherein the samples comprise production process data of each process acquired by each detection sensor in a production process and a product quality label marked after the product production is completed;
• S2:
• preprocessing the acquired samples, and constructing a training data set by the preprocessed samples, wherein the preprocessing comprises data cleaning, data alignment and normalization;
• wherein S2 specifically comprises:
• S2.1:
• cleaning the acquired samples by a data cleaning algorithm, and removing samples comprising sampling anomalies and not-a-number values in the acquired production process data to obtain cleaned samples;
• S2.2:
• performing the data alignment and the data normalization on the cleaned samples and constructing the training data set;
• specifically, first, the production process data in the cleaned samples is subjected to data alignment operation, so that the production process data of each aligned sample have a same length m, then, the aligned production process data is subjected to normalization operation, and finally, production process data Xp corresponding to each process after preprocessing is obtained, p being a serial number of processes and p=1, 2, . . . , n, and n being the number of the corresponding processes in the production process;
• therefore, each preprocessed sample is represented by Xi={X1, X2, . . . , Xn}, and i is a serial number of the preprocessed samples, so that the training data set
• D
• =
• {
• X
• i
• ,
• y
• i
• }
• i
• =
• 1
• N
• is finally obtained, N being a number of the preprocessed samples, and yi being the product quality label corresponding to the preprocessed samples Xi;
• S3:
• constructing a product quality detection model in the complex industrial process fusing the prior knowledge;
• wherein the product quality detection model in the complex industrial process fusing the prior knowledge comprises a plurality of feature extraction networks, a feature fusion module and a quality detection module;
• the feature extraction networks are configured to perform feature extraction on the production process data in each process to obtain a feature representation corresponding to the production process data of each process;
• a number of the feature extraction networks equals a number of the corresponding processes in the production process, one feature extraction network corresponds to one process;
• the feature fusion module is configured to splice input feature representation corresponding to the production process data of each process to obtain a spliced feature representation, then take mutual influence of each process as the prior knowledge, and fuse the spliced feature representation by the prior knowledge to finally obtain the feature representation of a corresponding preprocessed sample;
• the prior knowledge is represented by an adjacent matrix, the adjacent matrix A=[apq]n×n represents an influence relationship among n processes in the production process, apq∈{0,1} indicates whether a pth process generates influence on production of a qth process, 0 indicates that no influence is generated, and 1 indicates that influence is generated;
• the quality detection module is a two-layer full-connection network and is configured to classify input feature representations of the preprocessed samples and output predicted product quality;
• S4:
• training the product quality detection model in the complex industrial process fusing the prior knowledge by the training data set to obtain a trained product quality detection model in the complex industrial process fusing the prior knowledge;
• wherein S4 specifically comprises:
• S4.1:
• respectively inputting the production process data of each process in the preprocessed samples in the training data set into a corresponding feature extraction network to obtain the feature representation corresponding to the production process data of each process;
• a feature extraction process of the feature extraction networks is as follows:
• z
• p
• =
• f
• p
• (
• X
• p
• )
• ,
• wherein zp is the feature representation of the pth process, and fp(•) represents a feature extraction network corresponding to the pth process;
• a multi-scale feature extraction module in each feature extraction network adopts convolutional kernels with various sizes to perform feature parallel extraction;
• specifically, convolutional kernels Wt1 and Wt2 of 1×k1 and 1×k2 are adopted to extract features on time scales, convolutional kernels Ws1 and Ws2 of k1×1 and k2×1 are respectively adopted to extract features on variable scales, and k1 and k2 are sizes of the convolutional kernels;
• the multi-scale feature extraction module is as follows:
• z
• p_out
• =
• W
• s
• ⁢
• 1
• ×
• (
• W
• t
• ⁢
• 1
• ×
• z
• p_in
• )
• +
• W
• s
• ⁢
• 2
• ×
• (
• W
• t
• ⁢
• 2
• ×
• z
• p_in
• )
• ,
• wherein zp_in represents an input feature of the feature extraction module, zp_out represents an output feature, Wt1 is a convolutional kernel of a first convolutional network in a first branch, Ws1 is a convolutional kernel of a second convolutional network in the first branch, Wt2 is a convolutional kernel of a first convolutional network in a second branch, and Ws2 is a convolutional kernel of a second convolutional network in the second branch;
• S4.2:
• analyzing a production technology in the complex industrial process to obtain the mutual influence among the processes to serve as the prior knowledge;
• S4.3:
• performing feature fusion on the feature representation corresponding to an extracted production process data of each process by the obtained prior knowledge to finally obtain the feature representation of a corresponding sample, and predicting the product quality by the quality detection module to obtain the predicted product quality;
• specifically, the feature fusion based on the prior knowledge in S4.3 is shown as the following formula:
• y
• ˆ
• =
• f
• c
• (
• A
• ⁢
• con
• ⁢
• cat
• ⁡
• (
• z
• 1
• ,
• z
• 2
• ,
• …
• ,
• z
• n
• )
• )
• ,
• wherein ŷ represents a predicted value of the product quality by the product quality detection model in the complex industrial process fusing the prior knowledge, concat(•) represents splicing operation of the feature representation of each process, fc(•) represents the quality detection module, A represents the adjacent matrix, and zn represents the feature representation of the nth process;
• S4.4:
• after reaching set training rounds or converging a loss function, ending the training to obtain the trained product quality detection model in the complex industrial process fusing the prior knowledge;
• specifically, the loss function in S4.4 is a cross entropy loss function, and a loss between a product quality predicted value and a product quality real value is calculated by a cross entropy loss LCE, as shown in the following formula:
• L
• CE
• =
• -
• ∑
• j
• =
• 1
• c
• y
• j
• ⁢
• log
• ⁡
• (
• y
• ˆ
• j
• )
• ,
• wherein c represents a number of product quality types, j is a serial number of the product quality types, yj is the product quality real value, and ŷj is the product quality predicted value;
• S5:
• deploying a trained quality detection model in the complex industrial process fusing the prior knowledge for product quality detection in the complex industrial process to obtain product quality.