Intelligent manufacturing plant process quality monitoring method mixed with vae and deep neural network

Abstract

To address the technical problem that traditional distributed process quality monitoring methods often ignore or suppress some fault information, leading to the inability to identify early and localized faults in production systems, this invention provides a smart manufacturing plant process quality monitoring method that combines VAE (Visual Automation) and deep neural networks. This invention divides preprocessed data into multiple sub-units to construct a distributed manufacturing process, breaking down the originally complex manufacturing scenario into multiple smaller data-dimensional manufacturing units for distributed monitoring. Then, this invention establishes a VAE-LSTM process monitoring network through nonlinear mapping to extract the features of the complex manufacturing system's data distribution. The VAE network model is improved using a Long Short-Term Memory (LSTM) deep neural network to handle the temporal data between various operational units in the actual factory process. This enables accurate location of the specific sub-unit where the fault occurs and precise identification of the fault location, thereby achieving quality monitoring at each stage of the smart manufacturing plant process.

Description

Technical Field

[0001] This invention belongs to the field of distributed process monitoring technology in industrial intelligent manufacturing, specifically relating to a method for monitoring the process quality of intelligent manufacturing plants that combines VAE and deep neural networks. Background Technology

[0002] In modern industrial production processes, especially in intelligent, large-scale, and multi-unit production processes, implementing real-time process monitoring methods can promptly detect faults, reduce damage to industrial instruments, and effectively improve production efficiency.

[0003] Traditional monitoring methods, including multivariate statistical methods and dynamic modeling-based quality monitoring methods, cannot explain the states and relationships between units in the production process, and generally suffer from low detection accuracy and efficiency. Distributed process quality monitoring reduces the complexity of quality monitoring by dividing the entire production process into multiple sub-units, and then analyzes whether faults have occurred in the production process by monitoring the status of each sub-unit. Therefore, the adoption of distributed process quality monitoring is particularly important in intelligent, large-scale, multi-unit production processes.

[0004] However, when monitoring the process of complex industrial manufacturing systems, traditional deep learning-based distributed process monitoring methods establish global monitoring for the entire process but ignore local features in the system. There are also methods that establish monitors for local units but ignore the temporal correlation between units. Because the monitoring methods ignore or suppress some fault information, they cannot identify early faults and local faults in the production system. Summary of the Invention

[0005] To address the technical problem that traditional distributed process quality monitoring methods often ignore or suppress some fault information, resulting in the inability to identify early and local faults in the production system, this invention provides a smart manufacturing plant process quality monitoring method that combines VAE and deep neural networks.

[0006] The technical solution of this invention is:

[0007] The intelligent manufacturing plant process quality monitoring method that combines VAE with deep neural networks is unique in that it includes the following steps:

[0008] Step 1: Build a manufacturing system simulation model of the smart manufacturing factory and collect simulation data of the manufacturing process, including one-to-one corresponding manufacturing process variables and quality observation variables. Introduce s types of faults as data labels to divide the manufacturing process variables into normal production data and production data with faults. Use the normal production data and its corresponding set of quality observation variables to build a training set, and use the production data with faults and its corresponding set of quality observation variables to build a test set.

[0009] Step 2: Preprocess the data in the training and test sets by removing non-numeric characters and then standardizing them;

[0010] Step 3: Divide both the preprocessed training and test sets into multiple sub-units;

[0011] Step 4: Build and train the VAE-LSTM process monitoring network model;

[0012] Step 5: Distributed process quality monitoring;

[0013] Step 5.1: After preprocessing the manufacturing process data collected in actual production, divide it into multiple manufacturing process variable sub-units and their corresponding quality monitoring variable sub-units, and input them into the VAE-LSTM process monitoring network model trained in Step 4. The VAE-LSTM process monitoring network model outputs lnp(x i |LV) and lnp(y i |LV);p(x i |LV) and lnp(y i |LV) samples for each latent variable LV with respect to x i and y i The posterior distribution of x; i For each sub-unit, process variable y is manufactured. i For x i Corresponding quality monitoring variables;

[0014] Step 5.2 The output lnp(x) i |LV) is compared with the preset threshold of each manufacturing process variable in the corresponding manufacturing process variable sub-unit.

[0015] If the output is lnp(x) i If |LV) is greater than a preset threshold of a certain manufacturing process variable, it indicates that a manufacturing process corresponding to that manufacturing process variable has failed. Based on the manufacturing process sub-unit to which the manufacturing process variable belongs and its data tag, the faulty manufacturing process variable sub-unit and fault type are located, thereby realizing distributed process quality monitoring.

[0016] Furthermore, in step 1, a manufacturing system simulation model of the intelligent manufacturing factory is built using MATLAB.

[0017] Furthermore, the method for dividing the preprocessed training set into multiple sub-units in step 3 is as follows:

[0018] Step 3.1: Calculate the correlation coefficient matrix of the manufacturing process variable dataset in the preprocessed training set;

[0019] Step 3.2: Connect each manufacturing process variable in the preprocessed training set with the manufacturing process variable with the largest correlation coefficient to form the original manufacturing process variable relationship network;

[0020] Step 3.3: Calculate the edge betweenness number of each edge in the current manufacturing process variable relationship network;

[0021] Step 3.4: Determine whether there is a unique maximum edge betweenness in the current manufacturing process variable relationship network. If yes, proceed to step 3.5; otherwise, proceed to step 3.7.

[0022] Step 3.5: Find the edge with the largest betweenness and remove it to obtain the manufacturing process variable relationship network after removing the largest edge;

[0023] Step 3.6: Take the set of manufacturing process variables in the manufacturing process variable relationship network obtained in Step 3.5 after removing the maximum edge as a sub-unit, and calculate the modularity of the current manufacturing process variable relationship network. If the current value is greater than the value in the previous iteration, it means that the current sub-unit meets the requirements. Store the current value and return to Step 3.3; if the current value is less than or equal to the value in the previous iteration, it means that the current sub-unit does not meet the requirements. Discard it and return to Step 3.3.

[0024] The formula for calculating modularity is as follows:

[0025]

[0026] Where i represents the sequence number of the subunit; e i This represents the proportion of the number of edges formed by non-common points in the manufacturing process variable relationship network corresponding to the i-th sub-unit to the total number of edges in the original manufacturing process variable relationship network; a i This represents the proportion of the total number of edges in the manufacturing process variable relationship network corresponding to the i-th sub-unit to the total number of edges in the original manufacturing process variable relationship network;

[0027] Step 3.7: The process division is complete, ultimately yielding multiple manufacturing process variable sub-units X1, X2, ..., X n Correspondingly, based on the manufacturing process variable sub-units X1, X2, ..., X obtained from the division... nBy establishing the correspondence between various manufacturing process variables and quality observation variables, multiple quality observation variable sub-units Y1, Y2, ..., Y can be obtained. n ;

[0028] In step 3, the method for dividing the preprocessed test set into multiple sub-units is the same as that for the training set.

[0029] Furthermore, the VAE-LSTM process monitoring network model described in step 4 includes an LSTM encoder, an X-decoder, a Y-decoder, a latent variable calculation module, a posterior distribution calculation module, and an activation function module;

[0030] LSTM encoder acquires input data (X) i ,Y i The distribution pattern of ) is determined, and noise ε is simultaneously input into the latent variable calculation module. (s) Then, the latent variable calculation module calculates the latent variable LV based on the received data. Next, the LSTM encoder samples the latent variable LV according to the distribution pattern of the acquired data and outputs the sampling result LV. (s) The inputs are fed into the corresponding X-decoder and Y-decoder, which decode the received latent variable sample data to obtain the mean f and ... g The data is input to the posterior distribution calculation module, and simultaneously, the observed value e of the manufacturing process variable generated based on the noise factor is input to the posterior distribution calculation module. i and the observed value t that affects the quality variable i Standard deviation ∑ e and ∑ t The posterior distribution calculation module calculates the posterior distribution based on the mean f and g of the sampled data and the standard deviation ∑ of the noise factor. e and ∑ t Perform the calculation to obtain the posterior distribution p(x|LV)=N(f(LV),∑ e ) and p(y|LV)=N(g(LV),∑ t The posterior distribution is input into the activation function module, and the output of the activation function module is the output of the VAE-LSTM process monitoring network model.

[0031] Furthermore, in step 4, the expectation-maximization algorithm, variational lower bound, local Gaussian theorem, and sigmoid activation function are used to optimize the VAE-LSTM process monitoring network model until the maximum number of training iterations is reached, thereby obtaining the optimal set of network parameters of the VAE-LSTM process monitoring network model and the trained VAE-LSTM process monitoring network model.

[0032] A computer-readable storage medium having a computer program stored thereon; characterized in that the computer program is used to execute the above-described intelligent manufacturing plant process quality monitoring method when run.

[0033] An electronic device, including a processor, a memory, and a computer program; characterized in that: the computer program is executed by the processor to perform the above-described intelligent manufacturing plant process quality monitoring method.

[0034] The beneficial effects of this invention are:

[0035] 1. Due to the complex data format in intelligent manufacturing plants, it is impossible to monitor product quality simply through a single variable. Therefore, this invention divides the preprocessed data into multiple sub-units to construct a distributed manufacturing process. The originally complex manufacturing scenario is divided into multiple manufacturing units with smaller data dimensions for distributed monitoring. Then, this invention establishes a VAE-LSTM process monitoring network through nonlinear mapping to extract the characteristics of the data distribution in complex manufacturing systems. The VAE network model is improved using a Long Short-Term Memory (LSTM) deep neural network to handle the temporal data between various operational units in the actual factory process. This allows for accurate location of the specific sub-unit where a fault occurs and precise identification of the fault location, thereby achieving quality monitoring at each stage of the intelligent manufacturing plant process. Timely fault detection reduces the communication costs and risks associated with centralized process monitoring methods. Furthermore, the final results of the case experiments in Table 2 show that this invention has a higher fault detection rate and a lower false alarm rate than VAE-DNN, PCA, and KPCA methods. Analysis of Table 2 also shows that this invention can achieve a fault detection rate of up to 100% and a false alarm rate of up to 0% for some fault types.

[0036] 2. In the probabilistic generation model of this invention, after data preprocessing, manufacturing variables are first divided into process variables and quality variables. Other variables are defined as "other variables," and noise factors are introduced into the model to account for some influences in the process. Finally, a nonlinear mapping is used to establish the probabilistic generation model. Analysis of the remaining variables and noise factors makes the VAE-LSTM process monitoring network more accurate and consistent with reality.

[0037] 3. The VAE-LSTM process monitoring network in this invention can also be used to monitor other types of data, such as images and sounds, and has a wide range of applications.

[0038] 4. To optimize the monitoring process of the VAE-LSTM process monitoring network and maximize its log-likelihood estimate, this invention uses the expectation-maximization (EM) algorithm to calculate and maximize it, and employs variational lower bounds to calculate the maximum likelihood estimate, thereby enabling the VAE-LSTM process monitoring network to reach the optimal state. At the same time, the algorithm is improved by introducing KL divergence to give the loss function of this invention, which can further optimize the VAE-LSTM process monitoring network in nonlinear manufacturing systems.

[0039] 5. Considering the difficulty in determining the expected value in nonlinear processes, this invention presents a sampling calculation method for the model. Furthermore, based on feature extraction using the LSTM model, a formulaic distribution is generated to specifically represent each process. The final model framework presented in the construction of the VAE-LSTM process monitoring network also differs. This demonstrates that the VAE-LSTM process monitoring network of this invention can adapt to more complex intelligent manufacturing production processes.

[0040] 6. This invention detects fault information by fusing data from various sub-units and combining two statistical measures. Based on the corresponding calculated thresholds, it can quantitatively determine the level of faults detected by the VAE-LSTM process monitoring network. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the distributed sub-unit division of the TE process.

[0042] Figure 2 This is a flowchart of the distributed subunit partitioning algorithm.

[0043] Figure 3 It is the result of the division of distributed sub-units.

[0044] Figure 4 This is a structural diagram of the VAE-LSTM process monitoring network.

[0045] Figure 5 This is a flowchart for distributed process quality monitoring.

[0046] Figure 6 This is example 1 of the distributed process quality monitoring result diagram.

[0047] Figure 7 This is example 2 of the distributed process quality monitoring result diagram. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] The intelligent manufacturing plant process quality monitoring method based on the hybrid VAE and deep neural network proposed in this invention includes the following steps:

[0050] Step 1: Data Collection;

[0051] This embodiment uses Figure 1 Taking the TEP (Tennessee Eastman Process) industrial manufacturing process as an example, data is collected through simulation of the TEP manufacturing system. Specifically, after building a simulation model of the TEP manufacturing system using MATLAB, the Simulink simulation toolbox in MATLAB is used to configure the environment and set parameters (including at least the sampling period, sampling length, output format, runtime, and disturbance factor). The simulation collects manufacturing process data, including manufacturing process variables related to product production and quality observation variables that correspond one-to-one with these variables and are related to product quality. Then, 21 known fault types (obtained from historical fault data) are introduced to label and classify the manufacturing process variables in the collected simulation data, ultimately resulting in a normal production dataset and a fault-prone production dataset. A training set (480*52 pixels) is constructed using the normal production dataset and its corresponding set of quality observation variables, and a test set (960*52 pixels) is constructed using the fault-prone production dataset and its corresponding set of quality observation variables.

[0052] Assume the manufacturing process variable of TEP is x. i ∈R m , i = 1, 2, ..., N; R m For the set of manufacturing process variables; N represents R m The number of process variables contained in the model. When the model input is the training set, R... m x represents the set of manufacturing process variables in the training set. i Represents the manufacturing process variables in the training set; when the model input is the test set, R... m x represents the set of manufacturing process variables in the test set. i This represents the manufacturing process variables in the test set.

[0053] Assume the quality observation variable of TEP is y. i ∈R p , i = 1, 2, ..., N; R p For the set of manufacturing process variables; N represents R m The number of quality observation variables contained in the model. When the model input is the training set, R p y represents the set of quality observation variables in the training set. i R represents the quality observation variables in the training set; when the model input is the test set, R... p x represents the set of quality observation variables in the test set. iThe quality observation variables represent those in the test set. Step 2: Preprocess the data in the training and test sets respectively;

[0054] Step 2.1: Determine whether there are special characters and stop characters or other non-numeric characters (such as English letters, / / , *, punctuation marks, etc.) in the training set and test set. If so, delete the special characters and stop characters or other non-numeric characters.

[0055] Step 2.2: Standardize the data in the training and test sets according to the following formula to ensure the uniformity of data dimensions in the input model and prevent overfitting and peaking phenomena in the VAE-LSTM process monitoring network model subsequently built in this invention. The data standardization formula is as follows:

[0056]

[0057] Among them, X nor X represents the result after data standardization; X represents the variable data in the original simulation dataset; X mean Represents the mean and X of the data in the original simulation dataset. std This represents the standard deviation of the data in the original simulation dataset; the original simulation dataset is the training set or test set obtained in step 1.

[0058] Step 3: Divide both the training set and the test set obtained after preprocessing in Step 2 into multiple sub-units;

[0059] The method for dividing the training and test sets into multiple sub-units is the same, such as... Figure 2 As shown, the following explanation uses the method of dividing the training set as an example. The specific steps are as follows:

[0060] Step 3.1: Calculate the correlation coefficient matrix in the training set obtained after preprocessing;

[0061] Step 3.2: Connect each manufacturing process variable in the preprocessed training set with its manufacturing process variable with the largest correlation coefficient to form the original manufacturing process variable relationship network;

[0062] Step 3.3: Calculate the edge betweenness number of each edge in the current manufacturing process variable relationship network. Since the total number of shortest paths in the same manufacturing process variable relationship network is fixed, it is only necessary to calculate the number of shortest paths through each edge.

[0063] Step 3.4: Determine whether there is a unique maximum edge betweenness in the current manufacturing process variable relationship network. If yes, proceed to step 3.5; otherwise, proceed to step 3.7.

[0064] Step 3.5: Find the edge with the largest betweenness number and remove it to obtain the manufacturing process variable relationship network after removing the largest edge;

[0065] Step 3.6: Take the set of manufacturing process variables in the manufacturing process variable relationship network obtained in Step 3.5 after removing the maximum edge as a sub-unit, and calculate the modularity of the current manufacturing process variable relationship network. If the current value is greater than the value in the previous iteration (the initial value is 0 in the first iteration), it means that the current sub-unit meets the requirements. Store the current value and return to Step 3.3; if the current value is less than or equal to the value in the previous iteration, it means that the current sub-unit does not meet the requirements. Discard it and return to Step 3.3.

[0066] The formula for calculating modularity is as follows:

[0067]

[0068] Where i represents the sequence number of the subunit; e i This represents the proportion of the number of edges formed by non-common points in the manufacturing process variable relationship network corresponding to the i-th sub-unit to the total number of edges in the original manufacturing process variable relationship network; a i This represents the proportion of the total number of edges in the manufacturing process variable relationship network corresponding to the i-th sub-unit to the total number of edges in the original manufacturing process variable relationship network.

[0069] Step 3.7: The process division is complete, ultimately yielding multiple manufacturing process variable sub-units X1, X2, ..., X n That is, the set of manufacturing process variables R in the training set. m Divided into multiple subsets R1 m R2 m ,...,R n m .

[0070] Accordingly, based on the manufacturing process variable sub-units X1, X2, ..., X obtained from the division... n By establishing the correspondence between various manufacturing process variables and quality observation variables, multiple quality observation variable sub-units Y1, Y2, ..., Y can be obtained. n That is, the set of quality observation variables R in the training set. p Divided into multiple subsets R1 p R2 p ,...,R n p .

[0071] Figure 3 Table 1 shows the final sub-unit division results of this embodiment. The variable labels connected by solid lines between two sub-units that do not intersect indicate that they exist in both sub-units.

[0072] Table 1. Example of sub-unit partitioning results.

[0073]

[0074]

[0075] Step 4: Build and train a VAE-LSTM process monitoring network model to monitor the various manufacturing process variable sub-units obtained in Step 3;

[0076] Step 4.1: The VAE-LSTM process monitoring network model is built as follows. Figure 4 As shown, it includes an LSTM encoder, an X-decoder, a Y-decoder, a latent variable calculation module, a posterior distribution calculation module, and an activation function module.

[0077] (X i ,Y i After being input into the VAE-LSTM process monitoring network model, the input data (X) is first obtained through the LSTM encoder. i ,Y i The distribution pattern of ) is determined, and noise ε is input into the latent variable calculation module. (s) Then, the latent variable calculation module calculates the latent variable LV based on the received data. Next, the LSTM encoder samples the latent variable according to the distribution pattern of the acquired data and outputs the sampling result LV. (s) The inputs are fed into the corresponding X-decoder and Y-decoder, which decode the received latent variable sample data to obtain the mean f and ... g The data is then input into the posterior distribution calculation module, along with the observed values ​​e of the manufacturing process variables generated based on the noise factor. i and the observed value t that affects the quality variable i Standard deviation ∑ e and ∑ t The posterior distribution calculation module calculates the posterior distribution based on the mean f and g of the sampled data and the standard deviation ∑ of the noise factor. e and ∑ t Perform the calculation to obtain the posterior distribution p(x|LV)=N(f(LV),∑ e ) and p(y|LV)=N(g(LV),∑ t The posterior distribution is input into the activation function module, and the output of the activation function module is the output of the VAE-LSTM process monitoring network model of the present invention.

[0078] in:

[0079] Latent variable (LV) refers to the factors that independently affect product quality in the distributed process quality monitoring workflow. s is the number of times the LSTM encoder samples the latent variable, where s = 1, 2, ..., S, and S is the maximum number of samples. (s) This is the latent variable sampled in the s-th sampling. Latent variable LV (s) From noise ε (s) And determined by the mean μ, ε (s) It follows a unit Gaussian distribution p(ε)=N(0,1).

[0080] The decoder and encoder have similar structures. For the X-decoder LSTM, when the input to the LSTM is LV, the output f(LV) is a constant with mean. The corresponding output of the Y-decoder LSTM is g(LV). The LSTM network represents distributed process data through its network structure. Considering the difficulty in determining the expected value of the sampling process, the sampling process in this invention is approximated as follows:

[0081]

[0082] Where S is the total number of samples of the latent variables by the LSTM encoder.

[0083] The X-decoder LSTM and the Y-decoder LSTM are represented as follows:

[0084] X = f(LV) + e + z

[0085] Y = g(LV) + t

[0086] In the formula, f(LV) represents R n →R m That is, a nonlinear mapping function describing how to map the latent variable set LV to the manufacturing process variable set; g(LV) represents R n →R p Z is used to describe how to generate quality observations from the latent variable set LV; variables that have no substantial impact on quality and manufacturing process (such as manufacturing process variables labeled 32 and 36 that do not exist in the final sub-cells obtained in the previous step) are constructed as Z. j Z is a set of variables that have no practical impact on quality and manufacturing process. j The data in the figure; e is a noise factor affecting variables related to the manufacturing process and follows a zero-mean Gaussian distribution, i.e., e ~ N(0, ∑ e In order to account for the observed values ​​that affect the quality variables, this invention introduces a noise factor t to represent them. This noise includes operational changes, process fluctuations, and some feedback activities in the process.

[0087] Step 4.2: Define the parameters of the VAE-LSTM process monitoring network model constructed in Step 4.1:

[0088] The sub-unit length val_num is determined by the number of sub-units divided in step 3;

[0089] Maximum length of manufacturing process variable data: 960 (maximum number of dimensions in the dataset);

[0090] The model optimization function is optim.Adam(model.parameters(), lr = learning_rate);

[0091] The activation function chosen in this invention is ln(1+e). x () is used as the sigmoid activation function;

[0092] The loss function used is KL, which means that the KL (Kullback-Leibler) divergence is used to measure the similarity between the distributions of a random vector with respect to a certain random measurement.

[0093] The data length in the linear layer of the network is 256;

[0094] 2000 iterations;

[0095] The data length and number of iterations in the linear layers of the network are determined based on the optimal convergence point of the model during training.

[0096] Step 4.3: Process the sub-unit process variables X1, X2, ..., X2 processed in Step 3. n and quality observation variables Y1, Y2, ..., Y n The VAE-LSTM process monitoring network model constructed in step 4.1 is input sequentially for training and testing. The optimal parameter values ​​of the cross-entropy, maximum likelihood estimation, and loss function of the VAE-LSTM process monitoring network model are obtained. It can also solve the problem of missing data and analyze the correlation between complex data.

[0097] Step 4.4: Repeat step 4.3 and use the Expectation-Maximization (EM) algorithm, variational lower bound, local Gaussian theorem, and sigmoid activation function to optimize the VAE-LSTM process monitoring network model until the maximum number of training iterations is reached. Then, obtain the optimal set of network parameters of the VAE-LSTM process monitoring network model and the trained VAE-LSTM process monitoring network model.

[0098] The specific process is as follows:

[0099] Step 4.4.1: Optimize the VAE-LSTM process monitoring network model to maximize its log-likelihood.

[0100] Assume manufacturing process variable subunit X i Quality observation variable subunit Y i All of them follow the standard normal distribution of the latent variable set LV, i.e., p(LV) = N(0,1). Then, according to the standard continuous model, assuming that each pair of inputs (x... i ,y i Independent and uniformly distributed, S can be used N Let (X,Y) represent the log-likelihood, and use the Expectation-Maximization (EM) algorithm to maximize S. N Log-likelihood of (X,Y);

[0101] Log-likelihood S N The expression for (X,Y) is as follows:

[0102]

[0103] According to the calculation process of maximum likelihood estimation, due to the marginal distribution p(x) i ,y i The complexity of ) makes it difficult to directly maximize S. N (x i ,y i The opposite variational lower bound LS N (x i ,y i It is easier to perform a maximization estimate, where: LS N (x i ,y i )≤S N (x i ,y i However, the manufacturing process variable subunit X i With quality observation variable subunit Y i Since they are independent of the latent variable set LV, the maximum likelihood can be calculated using the maximum variational lower bound based on the expectation-maximization algorithm, resulting in the following formula:

[0104]

[0105] Among them, E q(LV) It is a marginal distribution p(x) i ,y i The expectation of a random vector is given by the KL (Kullback-Leibler) divergence, which measures the random distribution q and marginal distribution p(x) with respect to a random measurement. i ,y iThe similarity between them. The VAE-LSTM process monitoring network model obtains the posterior distribution p(LV|x) for computing the decoder output data in the last training iteration of the E-step of the EM algorithm. i ,y i Then maximize LS N (x i ,y i And update the variational lower bound parameter during the M-step computation of the EM algorithm to achieve S N (x i ,y i Maximize, and finally repeat the EM algorithm until the log-likelihood S is reached. N The estimated parameters of (X,Y) converge.

[0106] Step 4.4.2: Generate the posterior distribution of each latent variable collected by the LSTM encoder with respect to (x,y);

[0107] The VAE-LSTM process monitoring network model established in this invention is a probabilistic generation model. In nonlinear systems, the generation formula is difficult to fit. Similarly, the posterior distribution p(LV|x,y) of the latent variable samples with respect to (x,y) does not have a specific analytical form. Therefore, this invention uses the local Gaussian theorem to obtain the posterior distribution of the latent variable samples.

[0108] The local Gaussian theorem formula is as follows:

[0109] p(LV|x,y)=N(μ(x,y),Λ(x,y))

[0110] Where μ is the mean; Λ(x,y) is the covariance matrix, and the orthogonality of LV can be represented by restricting the diagonal of the covariance matrix;

[0111] Using the formula of the local Gaussian theorem mentioned above, the posterior distributions of each latent variable sample with respect to x and y can be obtained:

[0112] p(x|LV)=N(f,∑ e )

[0113] p(y|LV)=N(g,∑ t ).

[0114] Step 4.4.3: Input the posterior distribution obtained in step 4.4.2 into the activation function module of the VAE-LSTM process monitoring network model for calculation. When the number of iterations reaches the set value, use the calculated value obtained by the activation function module as the output of the entire model.

[0115] This invention selects ln(1+e) xThe posterior distribution is input into the sigmoid activation function. The system checks if the number of iterations has reached the set number in step 4.2. If not, it returns to step 4.4.1 to continue training, and simultaneously uses the output loss function to determine if the model training process is gradually converging. If the number of iterations has reached the set number in step 4.2, the model training ends, and the output of the sigmoid activation function, lnp(x), is then executed. i |LV) and lnp(y i |LV) is the output of the VAE-LSTM process monitoring network module.

[0116] Step 5: Distributed process quality monitoring;

[0117] Step 5.1 After preprocessing the manufacturing process data collected in actual production, the same preprocessing and sampling method as in Step 2 is used. The data is then divided into multiple manufacturing process variable sub-units using the method in Step 3, along with their corresponding quality observation variable sub-units. These are then input into the VAE-LSTM process monitoring network model trained in Step 4. The model's final output is lnp(x). i |LV) and lnp(y i |LV).

[0118] Step 5.2 The output lnp(x) of the VAE-LSTM process monitoring network model... i |LV) is compared with the preset threshold of the corresponding manufacturing process variable in its corresponding manufacturing process variable sub-unit. If lnp(x) i If |LV) is greater than a preset threshold of a certain manufacturing process variable, it indicates that a failure has occurred in the manufacturing process corresponding to that manufacturing process variable. Furthermore, since there is a correlation between the manufacturing process variable and the failure type (the simulation data was labeled by introducing the failure type in step 1), it is also possible to locate the failure type that occurred in the manufacturing process, thereby realizing distributed process quality monitoring.

[0119] As a preferred option, based on the above fault diagnosis and location, the output lnp(y) can also be... i |LV) is compared with the preset threshold of each quality observation variable in the corresponding quality observation variable sub-unit. If lnp(y) i If |LV) is greater than a preset threshold for a certain quality observation variable, it indicates that a failure has occurred in the manufacturing stage corresponding to the manufacturing process variable corresponding to that quality observation variable. In this case, by outputting lnp(x)... i |LV) and lnp(y i The fault is determined and its type is located as soon as either |LV) is compared with the corresponding preset threshold. This improves the reliability and stability of fault monitoring and reduces the probability of missed fault detection.

[0120] Performance verification of the VAE-LSTM process monitoring network model of this invention:

[0121] Step 1: Based on distributed process quality monitoring, design a distributed fault monitor to quantitatively detect the fault detection level of this invention;

[0122] This invention combines two indicators T in the PCA monitoring process. 2 We use SPE (Statistical Process Optimization) to quantitatively analyze the failure status of variables in distributed processes. Among them, the statistic T... 2 Used to measure the change of the sample vector in the principal space (i.e., T). 2 Used to monitor normal manufacturing processes, the SPE metric measures the change in the projection of the sample vector onto the residual space (i.e., SPE is used to detect the occurrence of faults). The statistic T... 2 and Comparison, comparing SPE with Comparison: If one of the final results exceeds the statistical control limit... If so, it is determined that there is a fault in the manufacturing process. Figure 5 The specific design flow for a distributed process monitor is presented and elaborated in detail below:

[0123] Step 1.1: Collect local historical data p of TEP manufacturing process variables to construct a normal historical dataset, and normalize the data in the normal historical dataset according to the mean and variance;

[0124] Step 1.2: Divide the preprocessed dataset from Step 1.1 into multiple sub-units according to the partitioning method in Step 3, and then calculate the number N of latent variables generated by the probability model in the local historical data p based on the relationship between the sub-units.

[0125] Step 1.3: Input the variable data of each sub-unit, which has been divided in Step 1.2, into the VAE-LSTM model, and use the data trained in the VAE-LSTM process monitoring network model to calculate the weight matrix P and variance matrix V of the sub-unit variable data. p ;

[0126] Step 1.4: Determine the statistic T 2 SPE and its control limits

[0127] Construct the statistics T of each sub-unit variable data calculated in step 1.3 according to the following formula. 2 SPE, where T 2 It can monitor the output lnp(x) of the VAE-LSTM process monitoring network model. i |LV) and lnp(y i|LV) changes in the latent space, SPE can monitor lnp(x) i |LV) and lnp(y i |LV) changes in the residual space; the latent space refers to the space generated by latent variables in the process of generating probability distributions; the residual space refers to the space obtained after removing the latent space from the principal component space; the SPE index can monitor the normal region, thereby through the statistic T 2 Complementing the SPE index, it enables simultaneous monitoring of both faulty and normal areas, ultimately allowing the calculation of fault detection rate and false alarm rate.

[0128]

[0129]

[0130] Where Λ is a diagonal matrix, Λ=diag{λ1,λ2,...,λ A V is the feature vector matrix of the monitoring data; P is the first A columns of the feature vector matrix V; These represent the control limits with a confidence level of α;

[0131] Control Limits The following calculation methods are commonly used:

[0132]

[0133]

[0134] Where is the distribution value with A and nA degrees of freedom and confidence level α; The control limits here These can also be used as references for the preset thresholds of the manufacturing process variables and the preset thresholds of the quality observation variables in step 5.2 above.

[0135] Step 1.5: Based on the statistic T determined in Step 1.4 2 , control limits of SPE, and compare T values ​​of each subunit variable. 2 Regarding the SPE statistic, due to the statistic T... 2 During calculation, a residual projection is made onto the fault space, making it more sensitive to fault detection than SPE. Therefore, typically only the statistic T is used. 2 The fault can be determined by comparing it with the control limits. Specifically, when the statistic T of the sub-unit variable... 2The presence of conditions exceeding the control limits indicates a failure in the manufacturing process. Further optimization can be achieved by comparing the SPE statistic with the control limits to determine the failure, thus improving reliability. In this invention, the latent variables all follow a standard Gaussian distribution, and all data transformations are linear processes; therefore, the statistics T for each sub-unit variable are... 2 The control limits can all be determined by χ, which has one degree of freedom. 2 (Chi-square distribution) determined;

[0136] Step 1.6: When inputting the collected data from each sub-unit into the VAE-LSTM process monitoring network model, repeat steps 1.1 to 1.5 after each set of data is input to achieve the design of a distributed fault monitor.

[0137] Step 2: Design Fault Detection Rate (FDR) and False Alarm Rate (False Alarm Rate) to measure the performance of the VAE-LSTM process monitoring network model. FDR represents the ratio of samples whose detection indicators exceed the control range to the total number of samples, while False Alarm Rate represents the ratio of false alarm data to the total number of normal data. The specific descriptions are as follows:

[0138]

[0139]

[0140] This invention provides two schematic diagrams for fault detection. Figure 6 , Figure 7 This invention detects and analyzes 15 known types of faults. Since faults 3, 9, and 15 have relatively small magnitudes, the statistical quantity T is less important during the detection process. 2 Insufficient sensitivity; therefore, this invention employs distributed process detection for the remaining 12 known types of faults. We discovered that the manufacturing process quality introduced by fault 1 is recoverable; that is, when this type of fault occurs, after a period of fluctuation in product quality, it will eventually return to normal. This type of fault includes fault 1, fault 5, and fault 7. Therefore, distributed detection can detect faults that are difficult to detect during centralized detection. Figure 6 The diagram shows random fault types, which demonstrates that the present invention can detect faults promptly and accurately, and can obtain the corresponding characteristics for each fault type. Figure 7The results obtained from process monitoring of step-type faults show that these faults can be detected in a timely manner, and the model can output the status of various faults. According to the comparison of the final fault detection results in the TE process in Table 2, the VAE-LSTM process monitoring network proposed in this invention significantly improves the fault detection rate compared to traditional PCA and KPCA methods, and also improves the fault detection accuracy compared to the VAE-DNN method. The false alarm rate of this invention is also greatly reduced; as shown in Table 2, the false alarm rate of this invention is the lowest in distributed process monitoring.

[0141] Table 2. Fault detection results during TEP (%)

[0142]

Claims

1. A method for intelligent manufacturing plant process quality monitoring by mixing VAE with deep neural networks, characterized in that, Includes the following steps: Step 1: Build a manufacturing system simulation model of the smart manufacturing factory and collect simulation data of the manufacturing process, including one-to-one corresponding manufacturing process variables and quality observation variables. Introduce s types of faults as data labels to divide the manufacturing process variables into normal production data and production data with faults. Use the normal production data and its corresponding set of quality observation variables to build a training set, and use the production data with faults and its corresponding set of quality observation variables to build a test set. Step 2: Preprocess the data in the training and test sets by removing non-numeric characters and then standardizing them; Step 3: Divide both the preprocessed training and test sets into multiple sub-units; The method for dividing the preprocessed training set into multiple sub-units is: Step 3.1: Calculate the correlation coefficient matrix of the manufacturing process variable dataset in the preprocessed training set; Step 3.2: Connect each manufacturing process variable in the preprocessed training set with the manufacturing process variable with the largest correlation coefficient to form the original manufacturing process variable relationship network; Step 3.3: Calculate the edge betweenness number of each edge in the current manufacturing process variable relationship network; Step 3.4: Determine whether there is a unique maximum edge betweenness in the current manufacturing process variable relationship network. If yes, proceed to step 3.5; otherwise, proceed to step 3.

7. Step 3.5: Find the edge with the largest betweenness and remove it to obtain the manufacturing process variable relationship network after removing the largest edge; Step 3.6: Take the set of manufacturing process variables in the manufacturing process variable relationship network obtained in Step 3.5 after removing the maximum edge as a sub-unit, and calculate the modularity of the current manufacturing process variable relationship network. If the current value is greater than the value in the previous iteration, it means that the current sub-unit meets the requirements. Store the current value and return to Step 3.3; if the current value is less than or equal to the value in the previous iteration, it means that the current sub-unit does not meet the requirements. Discard it and return to Step 3.

3. Modularity The calculation formula is as follows: in, The serial number representing the sub-unit; Indicates the first The proportion of the number of edges formed by non-common points in the manufacturing process variable relationship network corresponding to each sub-unit to the total number of edges in the original manufacturing process variable relationship network; Indicates the first The proportion of the total number of edges in the manufacturing process variable relationship network corresponding to each sub-unit to the total number of edges in the original manufacturing process variable relationship network; Step 3.7: The process division is completed, ultimately resulting in multiple manufacturing process variable sub-units. Correspondingly, based on the manufacturing process variable sub-units obtained from the division... By establishing the correspondence between various manufacturing process variables and quality observation variables, multiple quality observation variable sub-units can be obtained. ; The method for dividing the preprocessed test set into multiple sub-units is the same as that for the training set; Step 4: Build and train the VAE-LSTM process monitoring network model; The VAE-LSTM process monitoring network model includes an LSTM encoder, an X-decoder, a Y-decoder, a latent variable calculation module, a posterior distribution calculation module, and an activation function module. LSTM encoder acquires input data The distribution pattern is determined, and noise is simultaneously input into the latent variable calculation module. Then, the latent variable calculation module calculates the latent variables based on the received data. Then, the LSTM encoder processes the latent variables according to the distribution pattern of the acquired data. Perform sampling and obtain the sampling results The inputs are fed into the corresponding X-decoder and Y-decoder, which decode the received latent variable sample data to obtain the mean of the sample data. and The data is then input into the posterior distribution calculation module, along with the observed values ​​of the manufacturing process variables generated based on the noise factor. and the observed values ​​that affect the quality variable Standard deviation and The posterior distribution calculation module calculates the mean of the sampled data. and and the standard deviation of the noise factor and Perform the calculation to obtain the posterior distribution. and The posterior distribution is input into the activation function module, and the output of the activation function module is the output of the VAE-LSTM process monitoring network model. Step 5: Distributed process quality monitoring; Step 5.1: After preprocessing the manufacturing process data collected in actual production, divide it into multiple manufacturing process variable sub-units and their corresponding quality monitoring variable sub-units, and input them into the VAE-LSTM process monitoring network model trained in Step 4. The VAE-LSTM process monitoring network model outputs... and ; and For each latent variable The samples are respectively about and The posterior distribution of; To create process variables in each sub-unit, To and Corresponding quality monitoring variables; Step 5.2 output Compare with the preset thresholds of each manufacturing process variable in the corresponding manufacturing process variable sub-unit. If the output If the value exceeds a preset threshold for a certain manufacturing process variable, it indicates that a fault has occurred in the manufacturing process corresponding to that manufacturing process variable. Furthermore, the fault is located by identifying the manufacturing process variable sub-unit to which the fault belongs and the fault type based on the manufacturing process sub-unit to which the manufacturing process variable belongs and its data tag, thereby achieving distributed process quality monitoring.

2. The VAE and deep neural network hybrid intelligent manufacturing plant process quality monitoring method of claim 1, wherein: In step 1, MATLAB is used to build a simulation model of the manufacturing system of the smart manufacturing factory.

3. The VAE and deep neural network hybrid intelligent manufacturing plant process quality monitoring method of claim 1, wherein: In step 4, the expectation-maximization algorithm, variational lower bound, local Gaussian theorem, and sigmoid activation function are used to optimize the VAE-LSTM process monitoring network model until the maximum number of training iterations is reached. Then, the optimal set of network parameters of the VAE-LSTM process monitoring network model and the trained VAE-LSTM process monitoring network model are obtained.

4. A computer readable storage medium having stored thereon a computer program; characterized in that: The computer program is executed to perform the intelligent manufacturing plant process quality monitoring method according to any one of claims 1-3.

5. An electronic device comprising a processor, a memory, and a computer program; characterized by: The computer program, when run by the processor, is used to execute the intelligent manufacturing plant process quality monitoring method according to any one of claims 1-3.

Images