A method for noise reduction, dimension reduction and feature extraction of production section data

Abstract

The present invention discloses a method for denoising, dimensionality reduction and feature extraction of production section data, comprising the following steps: (1) collecting data from the production section, and expressing the feature as [X, Y, 1], wherein X is the sample capacity and Y is the sample dimension; (2) encoding and decoding the data collected in step (1) through a denoising autoencoder network to achieve data denoising; (3) dimensionality reduction and feature extraction of the denoised data in step (2) through principal component analysis, and the feature expression after dimensionality reduction is [X, Z, 1]. In view of the difficult problems of the production line data collected by the production section having high noise, large capacity and diverse features, the present invention provides for the first time a high-dimensional strong noise production line data preprocessing method that integrates denoising autoencoder (DAE) and principal component analysis (PCA). The method of the present invention can effectively realize the cleaning of workshop industrial data and lay a foundation for subsequent data analysis and decision-making in the workshop.

Description

Technical Field

[0001] The present invention relates to the field of data cleaning, and in particular to a method for performing noise reduction, dimension reduction and feature extraction on data collected from a production section by integrating a noise reduction autoencoder (DAE) and a principal component analysis (PCA). Background Art

[0002] The smart workshop integrates advanced technologies such as network technology, modern sensor technology, and automation technology. A large number of intelligent devices such as sensors and data acquisition devices are installed in the production line. The production workshop has become a gathering center for information flow, material flow, and control flow. In the production process of products, a large amount of production, environment, status, process, and equipment operation data are generated at an unprecedented speed, thus forming workshop big data. However, due to the characteristics of workshop data such as high noise, large capacity, and diverse features, the collected data will affect the subsequent analysis and decision-making of the workshop. Therefore, before the workshop carries out subsequent work, it is necessary to reduce noise, reduce dimension, and extract features from the collected data. After dimensionality reduction, the main feature data is reduced, and the indicators that need to be analyzed are reduced without reducing the loss of the original indicator information; providing a basis for the next step of data analysis and decision-making in the workshop.

[0003] The present invention is proposed for this purpose. Summary of the invention

[0004] The present invention provides a method for denoising, dimensionality reduction and feature extraction of data collected from production sections by integrating denoising autoencoder (DAE) and principal component analysis (PCA). Aiming at the problem that the collected production line data has a lot of noise, a large capacity and various features, the present invention first performs data cleaning from the aspect of data denoising, and uses a denoising autoencoder network to encode and decode the original noisy data to reconstruct the original noise-free data, which can make the features more stable; then the principal component analysis method is used to extract features from the data from which the noise has been removed, and the orthogonal transformation is used to convert the observed data represented by linearly correlated variables into a few data represented by linearly independent variables, and the linearly independent variables are called principal components, so as to achieve data dimensionality reduction. The present invention lays a data foundation for data analysis and decision-making in intelligent workshops.

[0005] The technical solution of the present invention is as follows:

[0006] A method for denoising, reducing dimensions and extracting features from production section data, characterized in that it comprises the following steps:

[0007] (1) Collect data from the production process, and express the features as [X, Y, 1], where X is the sample size and Y is the sample dimension;

[0008] (2) encoding and decoding the data collected in step (1) through a denoising autoencoder (DAE) network to achieve data denoising;

[0009] (3) The denoised data in step (2) is subjected to dimensionality reduction and feature extraction through principal component analysis (PCA). The feature expression after dimensionality reduction is [X, Z, 1].

[0010] Preferably, in step (2), the data needs to be subjected to probabilistic noise corruption before the data is encoded and decoded, and the probabilistic noise corruption uses the following two-term distribution: Where n is the data sample size, p is the probability of event A occurring in each experiment, event A represents the random destruction of elements in the original input according to a certain proportion, and i represents the number of times event A occurs after n experiments.

[0011] Preferably, the steps of encoding and decoding the data are as follows: input the obtained noise-damaged data into the input layer of the denoising autoencoder model, and perform the encoding operation using the following formula: y=sigmoid(wx+b); wherein sigmoid is an activation function, w is the connection weight from the input layer to the hidden layer, x is the normalized input data, b is the bias of the hidden layer, and y is the output value after the sample feature is encoded;

[0012] Then, the original noisy data is decoded by denoising autoencoding, and the decoding operation is performed using the following formula: z = sigmoid (w'y + b'); where sigmoid is the activation function, w' is the weight from the hidden layer to the output layer, y is the output value after the above sample feature encoding, b' is the bias of the output layer, and z is the output value after the sample feature is decoded.

[0013] Preferably, the step of using the principal component analysis method in step (3) to reduce the dimension and extract features of the denoised data is as follows: using the output data z of the autoencoder as input data, using the principal component analysis method, and using orthogonal transformation to convert the observed data represented by linearly dependent variables into a small number of data represented by linearly independent variables, and then performing data dimension reduction and feature extraction on the linearly independent variables.

[0014] Preferably, the specific steps of dimensionality reduction and feature extraction are: the Y-item denoised data z output by the autoencoder are represented as column vectors A1, A2, A3.......A Y ; and merged into a data set denoised by autoencoder:

[0015] A=[A1,A2,A3.......A Y ];make

[0016]

[0017] The normalized matrix B is calculated using the autoencoder denoised dataset A.

[0018]

[0019]

[0020] Among them, m represents the sample capacity, n represents the sample characteristics, that is, Y sample characteristics; represents the mean of column j in matrix A, s j represents standard deviation;

[0021] Then the correlation coefficient matrix R is calculated from the standardized matrix B by the following formula: R = (r ij ) m×n ; i=1,2,3……m,j=1,2,3……n;then the eigenvalue λ is calculated by |R-λI|=0;where R represents the correlation coefficient matrix, r ij represents the correlation coefficient between the i-th feature and the j-th feature, I represents the unit matrix, and λ represents the eigenvalue; the corresponding eigenvalues ​​λ are sorted by size, and several eigenvalues ​​Z are taken, which are expressed as [X, Z, 1] after dimensionality reduction.

[0022] Beneficial effects of the present invention:

[0023] In order to solve the problem that the production line data collected by the production section has a lot of noise, large capacity and diverse features, the present invention provides a high-dimensional strong noise production line data preprocessing method that integrates noise reduction autoencoder (DAE) and principal component analysis (PCA) for the first time. The method of the present invention can effectively clean the workshop industrial data and lay the foundation for subsequent data analysis and decision-making in the workshop. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a high-dimensional strong noise production line data preprocessing method framework that integrates denoising autoencoder (DAE) and principal component analysis (PCA) of the present invention.

[0025] Figure 2 This is a structural diagram of the denoising autoencoding network described in the present invention.

[0026] Figure 3 It is a graph of experimental data of the thin plate drying process of the embodiment.

[0027] Figure 4 It is a comparison diagram of partial data of a certain feature after data denoising in the embodiment and the original data.

[0028] Figure 5 It is a process parameter correlation matrix diagram of the experimental data of the thin plate drying process in the embodiment.

[0029] Figure 6 It is a scree plot of process parameters of experimental data of the thin plate drying process of the embodiment.

[0030] Figure 7 It is a diagram explaining the total variance of process parameters of the experimental data of the thin plate drying process in the embodiment. DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments, but the content of the present invention is not limited to the scope of the embodiments.

[0032] Embodiment: A method for denoising, reducing dimensions and extracting features from production process data. Take the method for denoising, reducing dimensions and extracting features from experimental data of a thin plate drying process in a factory as an example. Figure 1 As shown, the following steps are included:

[0033] (1) The production line data of the thin plate drying link of a tobacco factory in 2021 was collected. The data situation is as follows: this batch of data has 245,165 sample data and 67 process parameters, that is, 67 characteristic samples. Some sample characteristics are outlet material temperature, hot air temperature, ambient temperature, ambient humidity, etc.; Figure 3 , the feature expression is [245165,67,1]; the production line data has the characteristics of high noise, large capacity and diverse features;

[0034] (2) The collected data of the thin-plate drying process is subjected to probability noise damage to generate some noise data. Perform probabilistic noise damage; where n is the data sample size, p is the probability of event A occurring in each experiment, event A represents random destruction of elements in the original input according to a certain ratio, and i represents the number of times event A occurs after n experiments;

[0035] (3) The above-mentioned noise-damaged data is input into the input layer of the denoising autoencoder network model, and the original noisy data is encoded by the denoising autoencoder network. The collected production line data is encoded by the formula y=sigmoid(wx+b). First, the production line data collected from the thin plate drying process of a tobacco factory is input as the training sample set, and the data feature is expressed as [245165,67,1]; then the training sample set is trained using a fully connected network; where sigmoid is the activation function, w is the connection weight from the input layer to the hidden layer, x is the normalized input data, that is, the original actual values ​​of the sample features such as the outlet material temperature, hot air temperature, ambient temperature, and ambient humidity, b is the bias of the hidden layer, and y is the output value after encoding the sample features such as the outlet material temperature, hot air temperature, ambient temperature, and ambient humidity;

[0036] (4) Decoding the original noisy data through the denoising DAE network, and decoding the feature output y extracted after encoding through the formula z = sigmoid (w'y + b'); selecting part of the test data from the encoded sample feature output values ​​such as outlet material temperature, hot air temperature, ambient temperature, ambient humidity, etc., and normalizing them; then splitting the noise data in the test data; finally, predicting the split data, and characterizing the prediction effect through the loss function; where sigmoid is the activation function, w' is the weight from the hidden layer to the output layer, y is the output value of the sample feature encoding such as outlet material temperature, hot air temperature, ambient temperature, ambient humidity, etc., b' is the bias of the output layer, and z is the output value of the sample feature decoding such as outlet material temperature, hot air temperature, ambient temperature, ambient humidity, etc.;

[0037] After the above encoding and decoding operations, the original data without noise can be reconstructed, which can make the features more stable. The denoising autoencoder network structure is as follows: Figure 2 As shown in the figure, the left half is the encoding area, which extracts the data features, and the right half is the decoding area, which reconstructs the data so that the reconstructed data is close to the original data. Figure 4 As shown by Figure 4 It can be seen that the denoised data is smoother and has fewer burrs than the original data;

[0038] (5) Based on the above, the denoised data is subjected to principal component analysis to reduce the dimensionality of the data. Specifically, the 67 denoised data z output by DAE are represented as column vectors A1, A2, A3…A67, and merged into the DAE denoised data set A = [A1, A2, A3…A 67 ]:make The standardized matrix B is calculated using the DAE denoised data set A.

[0039] In the above matrix and formula, m represents the sample capacity, and n represents the sample characteristics, i.e., 67 sample characteristics such as material temperature, hot air temperature, ambient temperature, and ambient humidity; represents the mean of the sample features of column j in matrix A, s j represents standard deviation;

[0040] (7) The correlation coefficient matrix R is calculated from the standardized matrix B by the following formula: R = (r ij ) m×n ; i=1,2,3……m,j=1,2,3……n;then the eigenvalue λ is calculated by |R-λI|=0;where R represents the correlation coefficient matrix, r ijrepresents the correlation coefficient between the i-th feature and the j-th feature, I represents the unit matrix, and λ represents the eigenvalue; the corresponding eigenvalues ​​are sorted by size, and 15 main eigenvalues ​​are taken, which is the feature expression after dimensionality reduction as [245165,15,1].

[0041] Based on the above, the correlation matrix of principal component analysis method based on SPSS analysis is as follows Figure 5 As shown, the gravel diagram is as follows Figure 6 As shown, the total variance explained is Figure 7 As shown in the figure, 15 main features are finally obtained after dimensionality reduction, and the feature expression after dimensionality reduction is [245165,15,1].

[0042] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for noise reduction, dimension reduction and feature extraction of production section data, characterized in that: The following steps are involved: (1) Collect data from the production process, and express the features as [X, Y, 1], where X is the sample size and Y is the sample dimension; (2) performing encoding and decoding operations on the data collected in step (1) through a denoising autoencoder network to achieve data denoising; (3) Dimension reduction and feature extraction are performed on the denoised data in step (2) by principal component analysis, and the feature expression after dimensionality reduction is [X, Z, 1]; The specific steps of dimensionality reduction and feature extraction are: the Y-term denoised data z output by the autoencoder is represented as column vectors A1, A2, A3...A Y ; and merge into a data set after autoencoding denoising: A=[A1,A2,A3.......A Y ]; make The normalized matrix B is calculated using the autoencoder denoised dataset A. Among them, m represents the sample capacity, n represents the sample characteristics, that is, Y sample characteristics; represents the mean of column j in matrix A, s j represents standard deviation; Then the correlation coefficient matrix R is calculated from the standardized matrix B by the following formula: R = (r ij ) m×n ; i=1,2,3……m,j=1,2,3……n;then the eigenvalue λ is calculated by |R-λI|=0;where R represents the correlation coefficient matrix, r ij represents the correlation coefficient between the i-th feature and the j-th feature, I represents the unit matrix, and λ represents the eigenvalue; the corresponding eigenvalues ​​λ are sorted by size, and several eigenvalues ​​Z are taken, which are the features after dimensionality reduction expressed as [X, Z, 1].

2. The method according to claim 1, characterized in that Step (2) Before encoding and decoding the data, the data needs to be subjected to probabilistic noise corruption, and the probabilistic noise corruption uses the following two-term distribution: Where n is the data sample size, p is the probability of event A occurring in each experiment, event A represents the random destruction of elements in the original input according to a certain proportion, and i represents the number of times event A occurs after n experiments.

3. The method according to claim 2, characterized in that The steps of encoding and decoding the data are as follows: input the obtained noise-damaged data into the input layer of the denoising autoencoder model, and use the following formula for encoding: y = sigmoid (wx + b); where sigmoid is the activation function, w is the connection weight from the input layer to the hidden layer, x is the normalized input data, b is the bias of the hidden layer, and y is the output value after the sample feature is encoded; Then, the original noisy data is decoded by denoising autoencoding, and the decoding operation is performed using the following formula: z = sigmoid (w'y + b'); where sigmoid is the activation function, w' is the weight from the hidden layer to the output layer, y is the output value after the above sample feature encoding, b' is the bias of the output layer, and z is the output value after the sample feature is decoded.

4. The method according to claim 3, characterized in that The steps of using the principal component analysis method to reduce the dimension and extract features of the denoised data in step (3) are as follows: using the output data z of the autoencoder as input data, using the principal component analysis method, and using orthogonal transformation to convert the observed data represented by linearly dependent variables into data represented by a small number of linearly independent variables, and then obtaining linearly independent variables for data dimensionality reduction and feature extraction.

Images