A product quality feature data monitoring method based on key feature extraction

Abstract

The application relates to a product quality feature data monitoring method based on key feature extraction, which comprises the following steps: in a product production process, quality monitoring is performed on the product according to the importance ranking of features; the importance ranking of features is obtained through the following steps: 1) obtaining several types of feature data of the product and performing pretreatment; 2) clustering each type of feature data through a clustering algorithm to obtain a clustering center of each type of feature data; 3) performing dimension reduction processing on each type of feature data through a trained autoencoder to obtain reduced dimension data; 4) calculating the feature weight of each type of feature through a ReliefF algorithm according to the reduced dimension data and the clustering center; and 5) ranking the importance of each type of feature according to the feature weight. Compared with the prior art, the application can extract the key quality characteristics of the product, strengthen the monitoring and optimization of the production link corresponding to the key quality characteristics, and improve the product fault diagnosis accuracy and the detection time.

Description

TECHNICAL FIELD

[0001] The application relates to a product quality monitoring technology, in particular to a product quality feature data monitoring method based on key feature extraction. BACKGROUND

[0002] With the intelligent development of process industrial production and processing technology, the process production process gradually develops towards automation, intelligence and innovation, and data gradually becomes an important factor for enterprises to improve productivity and production quality. Quality control in product production process is an important link of enterprise production, and early prediction of quality problems and fault classification can help enterprises optimize production process and reduce operation and maintenance cost. For production links prone to problems and high maintenance cost, key features affecting product quality can be analyzed in advance, so as to facilitate staff to supervise and control the process, effectively reduce the efficiency of fault occurrence and ensure the working efficiency of production line. However, most enterprises only record data in the data monitoring process without analyzing the characteristics and importance of the data, which leads to low utilization efficiency of the data although the data quantity is large, and the product production quality cannot be effectively improved according to the existing data. Due to the complexity of modern production process and the large number of product parameters, the mutual coupling influence between process parameters increases the difficulty of quality improvement. Under the complex process and production conditions, it is difficult to determine the key quality characteristics of hot rolling process, which leads to the fact that the enterprise cannot control the key process production links and parameters. In order to improve the quality control efficiency and reduce the cost of enterprises, it is particularly important to extract key quality characteristics, but the existing researches are mostly concentrated on the determination of optimal parameters and fault diagnosis, and the research on hot rolling quality parameters is relatively less.

[0003] Although there are some related quality characteristic analysis researches before, most of them are concentrated in the field of rotating machinery fault, and the research on process type process production is less, and many researchers do not consider the influence of product quality data on the final product quality. Secondly, the high-dimensional data sample analysis is not thorough enough, and the key quality feature determination method used is relatively single, which leads to the limitation of the research on key quality characteristics of process type industrial products. SUMMARY

[0004] The purpose of the application is to overcome the defects of the prior art and provide a product quality feature data monitoring method based on key feature extraction, which can intuitively extract the key quality characteristics of the product, can be targeted to strengthen the monitoring and optimization of the production links corresponding to the key quality characteristics, improve the product fault diagnosis accuracy and detection time, and can avoid risks in advance, reduce maintenance cost and avoid unnecessary economic loss.

[0005] The purpose of the application can be achieved by the following technical solutions:

[0006] A product quality feature data monitoring method based on key feature extraction, comprising:

[0007] In the product production process, the quality of the product is monitored according to the importance ranking of the features;

[0008] The process of obtaining the importance ranking of the features comprises:

[0009] 1) Obtain several types of feature data of the product and pre-process them;

[0010] 2) Cluster each type of feature data by using a clustering algorithm to obtain the cluster center of each type of feature data;

[0011] 3) Perform dimension reduction processing on each type of feature data by using a trained autoencoder to obtain reduced dimension data;

[0012] 4) Calculate the feature weight of each type of feature by using the ReliefF algorithm according to the reduced dimension data and the cluster center;

[0013] 5) Rank the importance of each type of feature according to the feature weight;

[0014] If the feature ranking is performed according to all the feature data, it will greatly increase the calculation amount and also lead to inaccurate feature ranking results. Therefore, by performing dimension reduction on the feature data and integrating the high-dimensional data, since the best data point of the collected feature data cannot be determined, the clustering algorithm is used to cluster each type of feature data, the obtained cluster center is taken as the best data point, the feature weight of each type of feature is calculated by using the ReliefF algorithm according to the reduced dimension data and the cluster center, and the importance of each type of feature is ranked according to the feature weight. The key quality characteristics of the product can be directly extracted, the production links corresponding to the key quality characteristics can be targetedly monitored and optimized, the product fault diagnosis accuracy and the detection time are improved, the risk can be avoided in advance, the maintenance cost is reduced, unnecessary economic losses are avoided, and the effect is more obvious for process products.

[0015] Further, the type of the autoencoder comprises a convolutional autoencoder, a denoising autoencoder, a stacked autoencoder, and a deep autoencoder.

[0016] Further, the determination method of the type of the autoencoder comprises:

[0017] The trained several types of autoencoders are respectively used to perform dimension reduction processing on the best feature data set, the error between the minimum reconstruction term of each type of autoencoder and the original data is calculated, the dimension reduction effects of different types of autoencoders are compared, and the autoencoder with the smallest error is selected as the final adopted autoencoder;

[0018] Adopt multiple self-encoders to carry out dimension reduction simultaneously, select the self-encoder with the minimum error as the final adopted self-encoder, which is beneficial to minimize the error, maximize the data characteristics of the feature data, and improve the accuracy of the final calculated feature weight.

[0019] Further, the step 4) comprises:

[0020] Obtain sample data of several types of features of the product to form a training set;

[0021] Randomly select a group of sample data from the training set, find the nearest neighbor sample from other sample data of the same type as the group of sample data, and find the nearest neighbor sample from other sample data of different types as the group of sample data;

[0022] Calculate each feature weight according to the following formula:

[0023]

[0024]

[0025] Wherein, A represents the feature type, P k is the proportion of the same feature type sample in the total sample number in the sample data, V i and V i,j,k are the feature values of x i and x i,j,k , max(A) and min(A) are the maximum and minimum values of all samples in A, diff(x i , x i,j,k ) is the feature difference of two different samples, m is the total number of features, and n is the number of nearest neighbor samples.

[0026] Further, the preprocessing process comprises:

[0027] Standardize and discretize the feature data.

[0028] Further, when standardizing and discretizing the feature data, clean and reduce the feature data, delete the feature record when there is a missing value in the feature data, ensure the consistency of the input data dimension, clean the data and eliminate the feature record containing abnormal values;

[0029] Preprocess the feature data to improve the accuracy of the feature data and improve the accuracy of the calculated feature weight.

[0030] Further, the clustering algorithm is a K-means clustering algorithm, which has the advantages of simple principle, easy implementation, fast convergence speed, better clustering effect, and strong interpretability.

[0031] Further, the encoding process of the autoencoder adopts a sigmoid function, and the feature reconstruction vector J is:

[0032]

[0033] wherein z is the reconstruction vector, x is the training set sample, and m is the number of training set samples.

[0034] Further, the loss function for training the autoencoder is mean square error.

[0035] Further, the gradient descent method is used to optimize the loss function.

[0036] Compared with the prior art, the present application has the following beneficial effects:

[0037] (1) In the product production process, the quality of the product is monitored according to the importance ranking of the features. If the feature ranking is performed according to all feature data, the calculation amount will be greatly increased, and the feature ranking result will be inaccurate. Therefore, the high-dimensional data is integrated by dimension reduction. Since the best data point of the collected feature data cannot be determined, the clustering algorithm is used to cluster each type of feature data, and the obtained clustering center is used as the best data point. Then, according to the dimension reduction data and the clustering center, the feature weight of each type of feature is calculated by the ReliefF algorithm, and the importance of each type of feature is ranked according to the feature weight. The key quality characteristics of the product can be directly extracted, and the production process corresponding to the key quality characteristics can be targetedly monitored and optimized, thereby improving the product fault diagnosis accuracy and detection time, avoiding unnecessary economic losses, and reducing maintenance costs.

[0038] (2) The types of the autoencoder include a convolutional autoencoder, a denoising autoencoder, a stacked autoencoder, and a deep autoencoder. The trained autoencoders of several types are used to perform dimension reduction on the best feature data set. The error between the minimum reconstruction term of each type of autoencoder and the original data is calculated, the dimension reduction effects of different types of autoencoders are compared, the autoencoder with the minimum error is selected as the final adopted autoencoder, and multiple autoencoders are used for dimension reduction. The autoencoder with the minimum error is selected as the final adopted autoencoder, which is beneficial to minimizing the error and maximizing the data characteristics of the feature data, and improves the accuracy of the feature weight obtained by the final calculation.

[0039] (3) The feature data is cleaned and reduced in data when the feature data is standardized and discretized, the feature record is deleted when the feature data has missing values, the input data dimension is ensured to be consistent, the data is cleaned and the feature record containing abnormal values is removed, the feature data is preprocessed, the accuracy of the feature data is improved, and the accuracy of calculating the feature weight is improved;

[0040] (4) The clustering algorithm is a K-means clustering algorithm, which has simple principle, easy implementation, fast convergence speed, better clustering effect and strong interpretability. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The flowchart of the acquisition process of the importance ranking of features;

[0042] Figure 2 The schematic diagram of part of production data collected for a hot rolling production line;

[0043] Figure 3 The structure diagram of the autoencoder;

[0044] Figure 4 The schematic diagram of the visualization result of the input data of the autoencoder;

[0045] Figure 5 The scatter plot of the data obtained after dimensionality reduction by the autoencoder. DETAILED DESCRIPTION

[0046] The present application will be described in detail below in combination with the drawings and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present application, and detailed implementation and specific operation process are given, but the protection scope of the present application is not limited to the following examples.

[0047] A product quality feature data monitoring method based on key feature extraction, comprising:

[0048] In the product production process, the product is monitored for quality according to the importance ranking of features;

[0049] As Figure 1 , the acquisition process of the importance ranking of features includes:

[0050] 1) Obtain a plurality of types of feature data of a product and pre-process them;

[0051] 2) Cluster each type of feature data by a clustering algorithm to obtain a cluster center of each type of feature data;

[0052] 3) Reduce the dimension of each type of feature data by a trained autoencoder to obtain reduced dimension data;

[0053] 4) According to the dimension reduction data and the clustering center, the feature weight of each type of feature is calculated by the ReliefF algorithm;

[0054] 5) The importance of each type of feature is sorted according to the size of the feature weight;

[0055] If the feature sorting is performed according to all the feature data, not only the calculation amount will be greatly increased, but also the feature sorting result will be inaccurate. Therefore, the high-dimensional data is integrated by dimension reduction. Since the best data point of the collected feature data cannot be determined, the clustering algorithm is used to cluster each type of feature data, and the obtained clustering center is used as the best data point. Then, the feature weight of each type of feature is calculated by the ReliefF algorithm according to the dimension reduction data and the clustering center. The importance of each type of feature is sorted according to the size of the feature weight. The key quality characteristics of the product can be directly extracted. The production links corresponding to the key quality characteristics can be targetedly monitored and optimized. The product fault diagnosis accuracy and the detection time are improved. The risk can be avoided in advance. The maintenance cost is reduced. Unnecessary economic losses are avoided. Especially for process products, the effect is more obvious.

[0056] The structure of the autoencoder is as shown in Figure 3 The types of the autoencoder include a convolutional autoencoder, a denoising autoencoder, a stacked autoencoder and a deep autoencoder.

[0057] The determination method of the type of the autoencoder is as follows:

[0058] The trained several types of autoencoders are used for dimension reduction processing of the best feature data set. The error between the minimum reconstruction term of each type of autoencoder and the original data is calculated. The dimension reduction effects of different types of autoencoders are compared. The autoencoder with the minimum error is selected as the finally used autoencoder.

[0059] The autoencoders of multiple types are used for dimension reduction. The autoencoder with the minimum error is selected as the finally used autoencoder. This is beneficial to minimize the error, maximize the data characteristics of the feature data, and improve the accuracy of the feature weight obtained by the final calculation.

[0060] Step 4) includes:

[0061] Sample data of several types of features of the product is obtained to form a training set;

[0062] A group of sample data is randomly selected from the training set. The nearest neighbor sample is searched from other sample data of the same type as the group of sample data. The nearest neighbor sample is searched from other sample data of different types from the group of sample data;

[0063] The feature weight of each feature is calculated according to the following formula:

[0064]

[0065]

[0066] wherein A represents a feature type, P k is the proportion of the same feature type sample in the total number of samples in the sample data, V i and V i,j,k are the feature values of x i and x i,j,k , max(A) and min(A) are the maximum and minimum values of all samples in A, diff(x i , x i,j,k ) is the feature difference of two different samples, m is the total number of features, n is the number of nearest neighbors, and the value range of the feature weight is [-1, 1].

[0067] The preprocessing process includes:

[0068] standardizing and discretizing the feature data.

[0069] When the feature data is standardized and discretized, the feature data is cleaned and reduced. When the feature data has missing values, the feature record is deleted to ensure that the input data dimension is consistent. The data is cleaned and the feature record containing abnormal values is removed.

[0070] The feature data is preprocessed to improve the accuracy of the feature data and improve the accuracy of calculating the feature weight.

[0071] The clustering algorithm is a K-means clustering algorithm, which is simple in principle, easy to implement, fast in convergence, has better clustering effect, and has strong interpretability.

[0072] The encoding process of the autoencoder uses a sigmoid function, and the feature reconstruction vector J is:

[0073]

[0074] wherein z is the reconstruction vector, x is the training set sample, and m is the number of training set samples.

[0075] The loss function for training the autoencoder is the mean square error, and the gradient descent method is used to optimize the loss function.

[0076] Taking the 1580 hot rolling process of Baowu Group as an example, the model running environment of the product quality feature data monitoring method based on key feature extraction is Python 3.6, Pycharm-Tensorflow framework, Nvidia 1080Ti GPU, windows10 system, 30 features related to hot rolling quality are selected, including tension, stress, steel type, outlet temperature, inlet temperature, stress, speed, torsion force and loop angle, 8316 sample data are collected for each feature, and pre-processing is performed, Figure 2 The schematic diagram of part of the production data collected for the hot rolling production line is shown in FIG. 1, Figure 2 The first column in FIG. 1 is the production line number, and the remaining columns represent the production data categories. Screw-down represents the collected pressure data, Force represents stress, and Speed represents the process processing speed.

[0077] The pre-processed feature data is divided into a training set and a test set, and the division ratio is 7:3. The training set and the test set are used to train and test the self-encoder, respectively.

[0078] Figure 4 The visualization schematic diagram of the sample data of two types of features input into the self-encoder is shown in FIG. 2. The total number of sample data of each type of feature is 8316, the number of categories is 2, and the visualization number iteration is 1. After dimensionality reduction by the self-encoder, the data scatter plot of each type of feature data is shown in FIG. 2. Figure 5

[0079] The embodiment proposes a product quality feature data monitoring method based on key feature extraction, integrates high-dimensional feature data, discovers information beneficial to production behind the feature data, so that quality inspection personnel can take targeted measures to optimize the production process, strengthen the management and control of key production links, and also can perform risk avoidance and other operations in advance, reduce maintenance cost, and avoid unnecessary economic loss.

[0080] The above describes in detail the preferred embodiments of the application. It should be understood that those skilled in the art can make many modifications and changes without creative labor according to the concept of the application. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment based on the existing technology according to the concept of the application should be within the protection scope determined by the claims.​

Claims

1. A product quality feature data monitoring method based on key feature extraction, characterized by, The application relates to a quality monitoring method for products in a production process. The method comprises the following steps: 1) obtaining and preprocessing several types of feature data of the product; 2) clustering each type of feature data by using a clustering algorithm to obtain a clustering center of each type of feature data, wherein the clustering algorithm is a K-means clustering algorithm; 3) performing dimension reduction processing on each type of feature data by using a trained autoencoder to obtain reduced dimension data; 4) calculating a feature weight of each type of feature data by using a ReliefF algorithm according to the reduced dimension data and the clustering center; 5) performing importance ranking on each type of feature according to the feature weight. The step 4) comprises the following steps: obtaining sample data of several types of features of the product to form a training set; randomly selecting a group of sample data from the training set, finding a nearest neighbor sample from other sample data of the same type as the group of sample data, and finding a nearest neighbor sample from other sample data of a different type from the group of sample data; calculating the feature weight according to the following formula: The determination method of the autoencoder type is as follows: wherein, denotes a feature type, is the proportion of the same feature type sample in the total number of samples in the sample data, and are respectively and the feature values of and are the maximum value and the minimum value of all samples in the feature difference of two different samples, is the total number of features, is the number of nearest neighbor samples; performing dimension reduction processing on the best feature data set by using several types of trained autoencoders respectively, comparing the dimension reduction effects of the autoencoders of different types by calculating the error between the minimum reconstruction term of each type of autoencoder and the original data, and selecting the autoencoder with the minimum error as the final adopted autoencoder. The types of the autoencoder include a convolutional autoencoder, a denoising autoencoder, a stacked autoencoder and a deep autoencoder.

2. The product quality feature data monitoring method based on key feature extraction according to claim 1, characterized in that, The preprocessing process comprises the following steps: 3.The product quality feature data monitoring method based on key feature extraction of claim 1, wherein, performing standardization and discretization operations on the feature data. When performing the standardization and discretization operations on the feature data, data cleaning and data reduction are performed on the feature data, when there are missing values in the feature data, the feature record is deleted, the input data dimension is ensured to be consistent, the data is cleaned and the feature record containing abnormal values is removed. 4.The product quality feature data monitoring method based on key feature extraction of claim 3, wherein, The loss function for training the autoencoder is a mean square error.

5. The product quality feature data monitoring method based on key feature extraction according to claim 1, characterized in that, The encoding process of the self-encoder adopts a sigmoid function, and the feature reconstruction vector is: wherein, z is a reconstruction vector, x is a training set sample, is the number of training set samples.

6. The product quality feature data monitoring method based on key feature extraction according to claim 1, characterized in that, The gradient descent method is used to optimize the loss function.

7. The product quality feature data monitoring method based on key feature extraction according to claim 6, characterized in that, ​

Images