A method and system for monitoring multiple processes of a cylinder dryer

Abstract

The present application relates to a kind of multi-block process monitoring method and system of cylinder dryer, belong to tobacco monitoring field, for the problem of inaccurate detection of abnormal working condition of cut tobacco drying, the present application includes the test sample data set is divided into multiple sub-blocks according to variable characteristics;Based on the hidden feature of each sub-block extracted from the trained auto-encoding model, the hidden feature of each sub-block is used as a new observation data set, multiple sub-PCA models are established, and the monitoring statistics of each sub-block in each sub-PCA model are calculated;The monitoring results of each sub-block are fused to obtain the final monitoring result;Determine whether the monitoring result exceeds the preset control limit threshold, if it exceeds, then carry out abnormal processing;Wherein the trained auto-encoding model is processed by standardization normal data set;According to variable characteristics, process variables are divided into multiple sub-blocks;Each sub-block trains an auto-encoding model to obtain a trained auto-encoding model;The method improves certain monitoring effect.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of drum dryer piece process monitoring method and system, belong to tobacco monitoring field. BACKGROUND

[0002] Leaf drying is an important process of cigarette production, the purpose of leaf drying is to dry the tobacco with higher humidity to the water content required for rolling into cigarette by drying machine, the quality of product is high, the problem in process is usually difficult to remedy, the significance of process monitoring research on drum leaf drying is that, can find fault in time, reduce the generation probability of defective product, help to improve the qualified rate of cigarette product quality, reduce the invalid loss of cigarette raw material, to achieve the purpose of reducing consumption.

[0003] Leaf drying process is controlled by multiple variables, traditional single variable statistical process control (SPC) method has mature application in its quality monitoring, mainly using rainbow chart and process capability index to monitor the state of equipment;Since single variable monitoring does not consider the correlation between process variables, and data presents the characteristics of diversification, single variable statistical method is more and more unable to meet the process monitoring demand;The process monitoring method based on data driving that is popular at present is not limited by the difficulties of modeling and expert knowledge accumulation, relies on the massive data generated by industrial process, and can achieve the purpose of process monitoring, so this kind of method is favored by many scholars.

[0004] With the development of artificial intelligence, deep learning method is also more and more applied to industrial process monitoring field, deep learning technology can mine useful information from a large amount of data through supervised or unsupervised learning, compared with kernel method, has better nonlinear feature extraction ability;In addition, integrated modeling strategy may cause data information confusion and reduce the precision of fault detection for large-scale industrial process. SUMMARY

[0005] The purpose of the present application is to provide a kind of drum dryer piece process monitoring method and system, to solve the problems that traditional principal component analysis (PCA) process monitoring method does not have nonlinear expression ability and global modeling cannot accurately detect fault.

[0006] To achieve the above-mentioned purpose, the scheme of the present application includes:

[0007] A kind of drum dryer piece process monitoring method of the present application, comprising the following steps:

[0008] 1) test sample data set is divided into multiple sub-blocks according to variable characteristics;

[0009] 2) extracting the hidden features of each sub-block based on the trained auto-encoding model, taking the hidden features of each sub-block as a new observation data set, establishing multiple sub-PCA models, and calculating the monitoring statistics of each sub-block in each sub-PCA model;

[0010] 3) fusing the monitoring results of each sub-block to obtain the final monitoring result;

[0011] 4) judging whether the monitoring result exceeds the preset control limit threshold, and if so, performing abnormal processing;

[0012] The trained auto-encoding model processes the normal data set through standardization; according to the variable characteristics, the process variables are divided into multiple sub-blocks; and the auto-encoding model is trained for each sub-block to obtain the trained auto-encoding model.

[0013] By dividing the variables into blocks, the hidden features of each sub-block are extracted; the PCA method is used to model the feature data of each sub-block, and the monitoring results of multiple sub-spaces are fused. Actual production data verification and comparison experiments show that the method has better fault detection performance.

[0014] Further, in the training process of the trained auto-encoding model, the control limit calculated by the PCA model established after the normal data set is divided into blocks is used as the control limit threshold in step 4).

[0015] By comparing with the control limit calculated by the model established after the normal data set is divided into blocks, the detection result can be more accurate.

[0016] Further, before being divided into multiple sub-blocks in step 1), the test sample is preprocessed to obtain a standardized data set as a test sample data set.

[0017] Further, by preprocessing the test sample to obtain a standardized data set, the detection result is more accurate.

[0018] Further, in step 2), the hidden features of each sub-block are extracted by the auto-encoder AE.

[0019] Further, in step 2), the monitoring statistics of each sub-block at least include Q statistics and T 2 statistics.

[0020] Further, in step 3), the statistics of each sub-block are fused by Bayesian.

[0021] Further, in step 4), the PCA model further evaluates the monitoring performance of the model by detection rate and false detection rate.

[0022] A multi-block process monitoring system of a drum dryer includes a processor configured to perform steps 1) to 4) of a multi-block process monitoring method of a drum dryer. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is the AE network structure;

[0024] Figure 2 is a flow chart of AE-MPCA process monitoring in this embodiment;

[0025] FIG3(a), FIG3(b), and FIG3(c) are process monitoring diagrams under different methods of this embodiment. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0027] Cut tobacco drying is a key process in cigarette production. As traditional cut tobacco drying equipment, the drum dryer has a complex internal structure and exhibits strong coupling between multiple variables. To address the limitations of traditional principal component analysis (PCA) process monitoring methods, which lack nonlinear representation capabilities and global modeling, and cannot accurately detect faults, a multi-block modeling approach driven by autoencoder feature extraction and multi-block principal component analysis (AE-MPCA) is proposed. First, the variables are partitioned into blocks, and an autoencoder (AE) is used to extract deep features of the variables involved in each sub-block. Second, PCA is used to model the feature data of each sub-block. Finally, Bayesian inference (BI) is used to fuse the monitoring results from multiple subspaces. Validation and comparative experiments using actual production data demonstrate that the proposed approach achieves average detection rates of 0.9176 and 0.8896 in the residual space, respectively, demonstrating superior fault detection performance.

[0028] The method is described in detail in the following embodiment of a method for monitoring multiple blocks of a drum tofu drying machine.

[0029] An embodiment of a method for monitoring multiple blocks of a drum tofu drying machine:

[0030] Before implementing this method, the algorithm principles to be used (autoencoder AE and principal component analysis PCA) are introduced:

[0031] Autoencoder AE (autoender, autoencoder):

[0032] like Figure 1 As shown, the automatic encoder is an unsupervised learning algorithm, including a three-layer network structure. The encoder takes the input information x = [x1, x2, ..., x n ]∈R n×j (where n is the number of samples and j is the number of variables), compressed into hidden features h∈R n×m , and then use this feature to reconstruct the output x′∈R n×jThe model can be trained by minimizing the reconstruction error of the input and output data, gradually improving the accuracy of the model.

[0033] The encoding process of input X is represented as:

[0034] h = f(W e x + b e ) (1)

[0035] The decoding process of hidden layer features is represented as:

[0036] x' = f(W d h + b d ) (2)

[0037] where f is the activation function, W e , W d are the weight matrices for encoding and decoding, and b e , b d are the biases for encoding and decoding. AE optimizes the parameter set C = {W e , W d , b e , b d} multiple times through the backpropagation algorithm until the reconstruction error is minimized, making the output of the model as close to the input as possible.

[0038] Principal Component Analysis (PCA):

[0039] As one of the basic multivariate statistical process monitoring methods, PCA projects data into a low-dimensional space by orthogonal linear transformation to extract uncorrelated latent variables to represent the changes in data. For a standardized data matrix X ∈ R n×u , where n is the number of samples and u is the number of variables, the matrix X can be decomposed into the following form:

[0040] X = TP T + E (3)

[0041] where T is the principal component score matrix, P is the loading matrix, and E is the residual matrix, which decomposes the original data space into principal component space and residual space, and the number of principal components k is obtained by cumulative variance contribution rate (CPV ≥ 85%).

[0042] T 2 and Q statistics are calculated as follows:

[0043] T 2 = tS -1 t T = xPS -1 P T x T(4)

[0044] Q = ee T = x(I-PP T )x T (5)

[0045] where t is the principal component score vector of dimension (1 x k), the diagonal matrix S = diag(λ1, λ2,... λ R k) is composed of the first k eigenvalues of the covariance matrix of the modeling dataset X, e is the residual vector, T 2 and Q control limits are usually obtained from F distribution, weighted χ 2 distribution respectively:

[0046]

[0047]

[0048] where F k,n-k,α α, v represent the mean and variance of the modeling data Q statistics respectively.

[0049] Based on the above two algorithms, a process monitoring method based on AE-MPCA is established,

[0050] A large amount of process data obtained online in actual industrial mass production has the characteristics of nonlinearity and strong coupling between variables. Global modeling will be flooded with invalid information, which reduces the fault detection accuracy. Therefore, first of all, according to prior knowledge, the variables are divided into b sub-blocks according to appropriate rules, multiple local models are established, and the monitoring effect is improved.

[0051] Under the above multi-module modeling framework, the AE-MPCA process monitoring algorithm flow is as shown in Figure 2 First, the hidden features of each sub-block are extracted by using the autoencoder AE, and the hidden features are used as new observation variables to establish the sub-PCA module. Finally, the monitoring results of multiple subspaces are fused by Bayesian inference to obtain the whole monitoring index.

[0052] The online detection process is as follows:

[0053] S1 is as shown in Figure 2 For new test samples, the standardized data set X test is obtained by data preprocessing, and the process variables are divided into b sub-blocks by using the same blocking strategy;

[0054] S2 based on the trained autoencoder model, extract the hidden features h of each sub-block; the hidden features of each block are used as new observation data set, b sub-PCA models are established, and the monitoring statistics T i2 , Q i :

[0055] T i 2 = x i P i S i -1 P i T x i T (9)

[0056] Q i = x i (I-P i P i T )x i T (10)

[0057] The false detection rate (FDR) and the false alarm rate (FAR) are used to evaluate the monitoring performance of the model:

[0058]

[0059]

[0060] S3 The statistics and control limits of each sub-block are different, and it is difficult to directly fuse these results. Bayesian inference strategy can construct overall monitoring indicators. Taking T 2 as an example, the sample X test The failure probability of the i-th sub-block can be expressed as follows:

[0061]

[0062]

[0063] Conditional probability And The specific expression is as follows:

[0064]

[0065] In the formula: X test,i represents the test sample in the i-th sub-block, N and F represent normal and abnormal conditions, is the prior probability of normal samples, representing the confidence level α, then represents the confidence level 1-α. Then, the detection results of all sub-blocks are fused to obtain the final monitoring result as follows:

[0066]

[0067] Similarly, the final monitoring result BIC of all sub-block Q statistics can be obtained Q as follows:

[0068]

[0069] when and BIC Q When the statistic exceeds the confidence level 1-α, it is considered that the process has failed, otherwise it is considered to be normal.

[0070] It should be noted that, in this embodiment, the confidence level is used as the control limit to evaluate whether it is normal, and in other embodiments, the evaluation is not limited to using other thresholds as the control limit.

[0071] Offline modeling is required before monitoring:

[0072] Through offline modeling, the trained autoencoder model and control limits required for the monitoring process are obtained to verify whether the sample data in the online detection meets the standards. The specific steps are as follows:

[0073] H1 normal data set X after standardization;

[0074] H2 divides the process variables into b sub-blocks according to the variable characteristics;

[0075] Each sub-block of H3 trains the autoencoder model separately, extracts the latent feature h of each sub-block according to formula (1), and obtains b feature data sets h1,h2,...,h b , the characteristic data represents the original data information to the maximum extent;

[0076] H4 will be h1,h2,...,h b As the new data set of each sub-block, according to formula (3), the PCA model of each sub-block can be expanded as

[0077] X i =T i P i T +E i ,i∈[1,b] (8)

[0078] And calculate the control limits of each sub-PCA model statistic according to formulas (6)-(7) Q i,lim ,Finally, the overall monitoring index is constructed through the Bayesian inference strategy, ,that is, the set control limit, which is compared with the online monitoring results.,If the set control limit is exceeded, the system will issue an alarm.

[0079] An embodiment of a multi-block process monitoring system for a drum tofu drying machine:

[0080] An embodiment of a multi-block process monitoring system of a drum dryer includes a processor configured to execute an embodiment of a method of multi-block process monitoring of a drum dryer. The embodiment of the method of multi-block process monitoring of a drum dryer is described in detail above, and will not be repeated here.

[0081] Case: First, 20 batches of historical data of YA brand of SX cigarette in May 2021 were selected for process monitoring research, with a sampling interval of 10 s, from May 6, 2021 to May 18, 2021. Through data screening, 11 batches of normal batch data were selected for monitoring modeling, and 5 batches of normal working condition data were added to the disturbance as test samples X test1 ; reselect 3 batches of normal working condition data from May 24 to May 25, 2021 to add disturbance as test samples X test2 , after preprocessing three groups of data to obtain modeling data set X train (7236x13) and test samples X test1 (2582x13), X test2 (2268x13), respectively using three different methods to monitor the process, wherein the block strategy is shown in Table 2:

[0082] Table 2 Block strategy

[0083]

[0084] The abnormal alarm rate FDR(%) of the three different methods is shown in Table 3 and Figures 3(a), 3(b), and 3(c):

[0085] Table 3

[0086]

[0087] Result analysis:

[0088] Comprehensive Figure 2 , Figures 3(a), 3(b), and 3(c) and Table 3, it can be found that for the statistics of the case, the AE-MPCA algorithm has a higher detection rate. For the statistics of the case, the PCA detection rate is only 32.53%, and the AE-PCA detection effect is improved by 45% compared with PCA, which proves that the autoencoder network has good nonlinear feature extraction capability; further, when the multi-block modeling and autoencoder feature extraction are combined, the detection rate of the AE-MPCA algorithm proposed in this paper is as high as 91.67%, which is improved by 14% compared with the AE-PCA algorithm, which shows that the multi-block modeling strategy can effectively extract the local features of the system and improve the abnormal detection accuracy.

Claims

1. A multi-block process monitoring method for a drum dryer, characterized in that The method comprises the following steps: 1) dividing a test sample data set into multiple sub-blocks according to variable characteristics in a cut tobacco process; 2) extracting hidden features of each sub-block based on a trained auto-encoding model, the hidden features of each sub-block being extracted by an auto-encoder AE feature extraction to extract nonlinear features of each sub-block; taking the hidden features of each sub-block as a new observation data set, adopting a block modeling strategy to establish multiple sub-PCA models, and calculating monitoring statistics of each sub-block in each sub-PCA model; 3) fusing monitoring results of each sub-block to obtain a final monitoring result; 4) judging whether the monitoring result exceeds a preset control limit threshold, and if so, performing abnormal processing; The trained auto-encoding model is obtained by standardizing a normal data set; The process variables are divided into multiple sub-blocks according to variable characteristics; and an auto-encoding model is trained for each sub-block to obtain a trained auto-encoding model.

2. The drum dryer multi-piece process monitoring method according to claim 1, characterized in that In the training process, the control limit calculated by the PCA model established after the normal data set is divided into blocks is used as the control limit threshold in step 4).

3. The drum sieve dryer multi-block process monitoring method according to claim 1, characterized in that, Before being divided into multiple sub-blocks in step 1), the test sample is preprocessed to obtain a standardized data set as the test sample data set.

4. The drum sieve dryer multi-block process monitoring method according to claim 1, characterized in that, Step 2) the monitoring statistics of each sub-block at least include Q statistics and T 2 statistics.

5. The drum sieve dryer multi-block process monitoring method according to claim 1, characterized in that, In step 3), the statistics of each sub-block are fused by Bayesian fusion.

6. The drum sieve dryer multi-block process monitoring method according to claim 1, characterized in that, In step 4), the PCA model further evaluates the monitoring performance of the model by detection rate and false detection rate.

7. A multi-block process monitoring system for a drum dryer, comprising a processor, wherein, The processor is configured to execute steps 1) to 4) of the multi-block process monitoring method of the drum cut tobacco machine according to claim 1.

Images