Control chart quality data time sequence prediction method for multi-variety small-batch production

Abstract

The invention discloses a control chart quality data time sequence prediction method for multi-variety small-batch production, and belongs to the field of intelligent manufacturing. The method mainly comprises the following steps: firstly, establishing a TimeGAN timing confrontation model to amplify small sample quality data, and analyzing a quality data amplification effect by using a T-SNE algorithm; and an LSTM time sequence prediction model is established to carry out time sequence prediction on the amplified data sample. Through the method, a quality control link can be advanced, and quality data in a period of time in the future can be controlled. And meanwhile, the problem of scarcity of quality data in multi-variety small-batch production is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent manufacturing and intelligent quality monitoring, and in particular relates to a control chart quality data time series prediction method for multi-variety small batch production. Background Art

[0002] The traditional product quality control process is to detect a certain data of the produced product, draw it as a quality control chart, and then manually judge the current production status. This method is suitable for conventional production processes, but it is not suitable for multi-variety and small-batch production processes. Multi-variety and small-batch production processes are discrete production, and there will be inserted orders and the production line will be adjusted due to product replacement. Therefore, the intelligent control of product quality in this production process faces the following difficulties: first, it is difficult to provide a large amount of high-quality training data for the algorithm model; second, the cross-production of multiple types of products makes it difficult to use traditional SPC technology to draw quality data into control charts to control production. Therefore, the present invention proposes a control chart quality data time series prediction method for multi-variety small-batch production to solve the data scarcity problem caused by small sample problems and perform quality data time series prediction. Summary of the invention

[0003] The purpose of this invention is to solve the problem of small samples in the process of multi-variety small batch production and the general quality control process can only control the products that have been produced. A time series adversarial generative network is introduced to solve the problem of small quality data samples, and a time series algorithm is used to predict future product quality data. A control chart quality data time series prediction method for multi-variety small batch production is proposed.

[0004] The present invention proposes a control chart quality data time series prediction method for multi-variety small batch production. To achieve the above purpose, the present invention adopts the following technical solutions:

[0005] Step 1: Draw a short-term single value control chart based on the collected quality data, processing procedures, production equipment and other factors.

[0006] Step 2: Use TimeGAN to enhance the quality data collected.

[0007] Step 3: Use the LSTM time series prediction algorithm to predict the trend of quality data.

[0008] Compared with the existing technology, the implementation of this method mainly includes the following steps: First, establish the TimeGAN time series adversarial model to amplify the small sample quality data, and use the T-SNE algorithm to analyze the quality data amplification effect. And establish the LSTM time series prediction model to perform time series prediction on the amplified data samples. This method can advance the quality control link and realize the control of quality data for a period of time in the future. At the same time, it also effectively solves the problem of scarcity of quality data in multi-variety small batch production. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Flowchart of this method.

[0010] Figure 2 List of examples of TimeGAN adversarial production networks.

[0011] Figure 3 The following are the visual diagrams of the nine modes of amplification.

[0012] Figure 4 The following are the timing prediction results for 9 modes. DETAILED DESCRIPTION

[0013] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings:

[0014] The main part of this method is divided into three parts. First, the collected product quality data is classified and a short-term single value control chart is drawn. Second, the collected small sample data is amplified using TimeGAN time series adversarial generation, and the enhancement effect is visualized using T-SNE. Finally, the LSTM model is used to perform time series prediction on the control chart quality data to obtain the final result. The following steps are used for specific discussion:

[0015] Step 1: Draw a short-term single value control chart based on the collected quality data, processing procedures, production equipment and other factors.

[0016] a) Classify the collected data according to the production process, production equipment and production time of the product;

[0017] b) Use Z control chart to control the quality of collected products.

[0018] Step 2: Build the TimeGAN model.

[0019] a) Construction of latent space: TimeGAN first samples in the latent low-dimensional continuous space through a random process;

[0020] b) Time series generation: These sampling points are fed into the generator, whose task is to map the points in the latent space back to the time series space and generate new time series data;

[0021] c) Use of the discriminator: The generated data and real data are fed into the discriminator together, and the discriminator distinguishes the real data from the fake data generated by the generator;

[0022] d) Adversarial training process: The generator and the discriminator are trained adversarially. The generator tries to generate more and more realistic time series data to deceive the discriminator, while the discriminator strives to improve its ability to distinguish between real and fake data. This process is iterated until the generator can generate fake data that is indistinguishable from real data;

[0023] e) Preservation of temporal dependency: TimeGAN specifically designs its network structure to ensure that the generated time series data can preserve the dynamic characteristics of time, which is achieved by introducing temporal correlation into the model;

[0024] f) Choice of model structure: In the implementation of TimeGAN, recurrent neural networks (RNN) or long short-term memory networks (LSTM) are often used as the model structures of the generator and discriminator because they can well capture the temporal dependencies in time series data.

[0025] Step 3: Establish an LSTM time series prediction model.

[0026] a) The number of input nodes is 40;

[0027] b) The number of output nodes is 10;

[0028] c) The number of neuron nodes in the hidden layer is determined to be 800 through multiple experiments;

[0029] d) Optimization algorithm: Use the Adam optimization algorithm, which is an optimization algorithm with adaptive learning rate and can automatically adjust the learning rate;

[0030] e) Maximum number of iterations: Set the maximum number of training iterations to 500;

[0031] f) Minimum batch size: Set the number of samples per batch to 10, which is the number of samples used each time the parameters are updated;

[0032] g) Prevent gradient explosion: Set the gradient threshold to 1. When the norm of the gradient exceeds the threshold, the gradient will be clipped to prevent gradient explosion.

[0033] h) Initial learning rate: Set the initial learning rate to 0.005, which is the learning speed of the network in the initial stage;

[0034] i) Learning rate scheduling: Use piecewise linear learning rate scheduling strategy;

[0035] j) Learning rate decrease cycle: Set the learning rate decrease cycle to 125, that is, the learning rate will decrease every 125 iterations;

[0036] k) Learning rate decrease factor: Set the learning rate decrease factor to 0.2, that is, each time the learning rate decreases, the learning rate will be multiplied by 0.2;

[0037] l) Validation data: The input and output data of the training set are used as validation data to evaluate the model performance during the training process;

[0038] m) Verification frequency: Set verification to be performed every 5 iterations to monitor the training progress and performance of the model.

[0039] The LSTM model was constructed with the above parameter settings, and a quality data time series prediction model with 40-dimensional input and 10-dimensional output was established.

[0040] The beneficial effects of the present invention are as follows: the present invention introduces the TimeGAN algorithm and establishes a quality generation model based on time series prediction in combination with LTSM, which effectively solves the problem of small samples in small batch production of multiple varieties and moves the quality control link forward.

[0041] Implementation Case 1

[0042] During the production process, product quality data was collected and 1,800 sets of quality data control charts were drawn. First, a data augmentation model was established. The main parameter settings of TimeGAN are shown in Table 1.

[0043] Table 1 TimeGAN parameter settings

[0044]

[0045] The TimeGAN network generated 1,800 sets of data. The comparison between the generated sequences and the real samples is shown in Figure 2. Figure 2 As shown in the figure, the horizontal axis is the time step, which is equivalent to the order of production product quality data, and the vertical axis represents the corresponding standardized value of quality data. Observing the comparison of the numerical values ​​and fluctuations of the two types of data, the two types of data are not equal in value, but have the same data characteristics, namely random distribution, within this time window, which shows that the generated data has learned the static characteristics and temporal characteristics of the real data. In addition, this difference in the values ​​of the two types of data is more in line with the purpose of research. To solve the problem of small samples, we cannot simply fill in a large number of repeated samples, but generate more new data after learning the characteristics of historical samples.

[0046] In order to verify the generation effect, T-SNE is used to perform two-dimensional dimensionality reduction visualization operations on the real data and the generated data, such as Figure 3As shown in the figure. Since the collected data set is composed of nine types of pattern data, the data set contains nine types of data. This paper performs T-SNE processing on these nine types of real data and the corresponding generated data. The red points in the figure represent the low-dimensional mapping of the real quality data, and the black points represent the low-dimensional mapping of the generated data. Observe the distribution of the dimensionality reduction of the real and generated data. The closer the distribution of the two points is, the more likely it is that the generated data and the real data have the same data characteristics.

[0047] observe Figure 3 It can be seen that T-SNE dimensionality reduction distributes the two types of data samples, which to a certain extent shows the difference in distribution between real data and generated data. Among them, the normal mode (NOR), cycle mode (CYC), layered mode (STR), rising trend (UT), falling trend (DT), and upward step mode (US) are basically distributed in the same state, and the generated data has the characteristics of real data. In the three types of data, system mode (SYS), downward step mode (DS), and mixed mode (MIX), the quality data generated by TimeGAN has a large overlap with the real data in distribution. Although there are some out-of-line points, TimeGAN essentially generates data with the same characteristics, rather than simple numerical replication. Therefore, using TimeGAN for data enhancement can effectively solve the problem of insufficient quality data samples in the process of multi-variety small batch generation, and can effectively provide sufficient learning samples for the subsequent quality data generation model.

[0048] The data augmented dataset obtained from TimeGAN is used as the training and test data for the LSTM algorithm. Nine patterns are randomly selected from the test set and their prediction results are visualized, as shown in the figure below. Figure 4 As shown in the figure, it is found that the generative model learns the changing trend of each type of pattern from a large amount of historical data, realizes the time series prediction of quality data, and lays a foundation for timely control of product quality.

Claims

1. A method for predicting control chart quality data time series for multi-variety small batch production, characterized in that: The following steps are involved: Step 1: Draw a short-term single value control chart based on the collected quality data according to the processing procedures, production equipment and other factors; Step 2: Use TimeGAN to generate adversarial networks and amplify the collected quality data control charts; Step 3: Use the LSTM time series prediction algorithm to predict the changing trend of the quality control chart data.

2. The method for predicting control chart quality data time series for multi-variety small batch production according to claim 1, characterized in that: Step 1 includes drawing a short-term single value control chart based on the collected quality data according to the processing procedures and production equipment factors; a) Classify the collected data according to the production process, production equipment and production time of the product; b) Use Z control chart to control the quality of collected products.

3. The method for predicting control chart quality data time series for multi-variety small batch production according to claim 1, characterized in that: Step 2 includes,the construction of latent space: TimeGAN first samples in the latent low-dimensional,continuous space through a random process; a) Generation of time series: The sampling points are fed into the generator, which maps the points in the latent space back to the time series space to generate new time series data; b) Use of the discriminator: The generated data and real data are sent to the discriminator together, and the discriminator distinguishes the real data from the fake data generated by the generator; c) Adversarial training process: The generator and discriminator are trained adversarially; The generator tries to generate more and more realistic time series data to deceive the discriminator, while the discriminator strives to improve its ability to distinguish between real and fake data; This process is iterated until the generator is able to produce fake data that is indistinguishable from real data; d) Preservation of temporal dependency: TimeGAN specifically designs its network structure to ensure that the generated time series data can preserve the dynamic characteristics of time, which is achieved by introducing temporal correlation into the model; e) Choice of model structure: In the implementation of TimeGAN, recurrent neural network (RNN) or long short-term memory (LSTM) network is used as the model structure of the generator and discriminator.

4. The method for predicting control chart quality data time series for multi-variety small batch production according to claim 1, characterized in that: Step 3 includes, inputting the number of nodes as 40; a) The number of output nodes is 10; b) The number of neuron nodes in the hidden layer is determined to be 800 through multiple experiments; c) Optimization algorithm: Use the Adam optimization algorithm, which is an optimization algorithm with adaptive learning rate and can automatically adjust the learning rate; d) Maximum number of iterations: Set the maximum number of training iterations to 500; e) Minimum batch size: Set the number of samples per batch to 10, which is the number of samples used each time the parameters are updated; f) Prevent gradient explosion: Set the gradient threshold to 1. When the norm of the gradient exceeds the threshold, the gradient will be clipped to prevent gradient explosion. g) Initial learning rate: Set the initial learning rate to 0.005, which is the learning speed of the network in the initial stage; h) Learning rate scheduling: Use piecewise linear learning rate scheduling strategy; i) Learning rate decrease cycle: Set the learning rate decrease cycle to 125, that is, the learning rate will decrease every 125 iterations; j) Learning rate decrease factor: Set the learning rate decrease factor to 0.2, that is, each time the learning rate decreases, the learning rate will be multiplied by 0.2; k) Validation data: The input and output data of the training set are used as validation data to evaluate the model performance during the training process; l) Verification frequency: Set verification to be performed every 5 iterations to monitor the training progress and performance of the model.

Images