CN111459140B

CN111459140B - Fermentation process fault monitoring method based on HHT-DCNN

Info

Publication number: CN111459140B
Application number: CN202010279403.XA
Authority: CN
Inventors: 高学金; 刘爽爽; 高慧慧
Original assignee: Beijing University of Technology
Current assignee: Beijing University of Technology
Priority date: 2020-04-10
Filing date: 2020-04-10
Publication date: 2021-06-25
Anticipated expiration: 2040-04-10
Also published as: CN111459140A

Abstract

The invention discloses a fermentation process fault monitoring method based on HHT-DCNN, which is used for solving the problem that the traditional fault monitoring method is easy to ignore the extraction of deep features of data, belongs to the field of fault monitoring, and comprises an off-line stage and an on-line stage. In the off-line stage, firstly, a sliding window technology is adopted to segment data to obtain a process variable sequence; then, performing Hilbert-Huang transform on the plurality of sequences to obtain a time-frequency diagram, and mining abnormal changes of variables in amplitude, frequency and phase; and finally, training a deep convolutional neural network model by using a time-frequency graph, and effectively extracting deep features of the data. In the on-line stage, m time data of a new batch are acquired from the fermentation process for preprocessing, and the new batch data is divided into a fault state and a non-fault state by adopting a trained model, so that the fault on-line monitoring is realized; the invention has the characteristics of high monitoring accuracy and strong practicability.

Description

Fermentation process fault monitoring method based on HHT-DCNN

Technical Field

The invention relates to the technical field of fault monitoring based on a neural network, in particular to a technology for monitoring faults in a fermentation process on line based on a convolutional neural network. The method of the invention is particularly applied to the fault monitoring of a typical fermentation process, namely a penicillin fermentation process and an escherichia coli fermentation process.

Background

In recent years, an intermittent process is becoming the mainstream of the market because the intermittent process can meet the demand of producing high value-added products and exceeds a continuous production mode gradually, a fermentation process is taken as a typical intermittent process, the mechanism is complex, the operation complexity is high, the product quality is easily influenced by uncertain factors, and the method has the characteristics of strong nonlinearity, dynamic property and the like and is difficult to establish a proper monitoring model for online monitoring. Therefore, in order to ensure the safe and stable operation of the fermentation process, it is necessary to establish an effective monitoring model.

The multivariate statistical method based on data driving is widely applied to the field of industrial engineering, and typical methods comprise Principal Component Analysis (PCA) and Independent Component Analysis (ICA), but the multivariate statistical method uses different kinds of thresholds as monitoring standards, so that the influence of incomplete feature extraction on monitoring results is easily ignored. Subsequently, deep learning is widely used in industrial processes with superior monitoring performance compared to conventional methods and shallow neural networks, and particularly, CNN is more and more favored by researchers with its superior feature extraction capability and is successfully used in industrial processes. However, the fermentation process disturbance actually occurs on different time scales, namely, the measured process data has time sequence correlation, and the dynamic characteristics of the data can directly influence the monitoring result of the process. Studies by scholars have discussed this problem, but studies have assumed that latent variable features extracted at different time periods are independent states, unrelated between their variable previous values and other variables. In addition, the slow characteristic analysis mining time-varying dynamic information in the batch processing process depends on complex nuclear calculation and cannot automatically learn the inherent time-varying characteristics. Therefore, before online monitoring, the problems of dynamic and nonlinear feature extraction of the fermentation process need to be considered, and the improvement of the model monitoring precision is very necessary.

Disclosure of Invention

In order to make up the defects of the prior art in dealing with the problems of dynamics and nonlinearity in the fermentation process, the invention provides a fermentation process monitoring method based on HHT-DCNN.

The fermentation process is a microbial population growth process, population quantity can change according to approximately the same curve, fluctuation of the parameters can also cause fluctuation of the microbial growth curve, corresponding amplitude, phase and frequency can change, and process data often have strong nonlinear relation and dynamic time sequence correlation relation. Aiming at the fermentation process with dynamic and nonlinear properties, the time-frequency characteristics of process data variables are considered, and the process data variables are analyzed by a time-frequency analysis method (HHT), so that abnormal change information of variable sequences on amplitude, frequency and phase is effectively mined; and then, a model is built by utilizing a convolutional neural network to adaptively extract fault characteristics to monitor the process, and the precision and monitoring performance of the model are improved to ensure the safe and stable operation of the fermentation process. Based on this, the following technical scheme and implementation steps are adopted:

A. an off-line modeling stage:

1) collecting historical data: collecting historical data under the normal working condition of a fermentation process, wherein the historical data is I batch data under the same process in the same fermentation process, and X ═ X (X ═ X)₁，X₂，…X_I)^TWherein X is_IRepresents the ith batch data. Each batch contains k sampling instants, each sampling instant collecting j process variables, i.e. X_i＝(x_i，1，x_i，2，…x_i，k) Wherein x is_i，kRepresenting data acquired at the ith sampling instant, X_i，k＝(x_i，k，1，x_i，k，2，…x_i，k，j) Wherein x is_i，k，jA measured value representing a jth process variable at a kth sampling time in the ith batch;

2) data preprocessing: aiming at the dynamics of the fermentation process, a sliding window technology is adopted to carry out sliding interception on all measured variables in sampling time, for one batch, the size of a sliding window is made to be m multiplied by n, m is m sampling moments of interception, j is n represents all measured variables, the sliding step length is 1, and process variable data are segmented to obtain a plurality of two-dimensional matrix sequences until all batch data sequences are obtained. All the batch data are expanded into a two-dimensional matrix according to variables and expressed as follows:

the plurality of two-dimensional matrix sequences for sliding window split batch 1 data acquisition are represented as follows:

3) dividing a two-dimensional matrix obtained by sliding a sliding window into a training sample and a testing sample, carrying out Hilbert-Huang transformation on the matrix, and capturing abnormal change information of the sequence on amplitude, frequency and phase. Firstly, decomposing a single variable in a sliding window acquisition matrix into a limited number of Intrinsic Mode Functions (IMFs) by using Empirical Mode Decomposition (EMD); then, each decomposed IMF is subjected to spectrum analysis by using Hilbert transform, and the instantaneous frequency of each IMF can be obtained; finally, collecting Hilbert spectrums of all IMFs of j variables to obtain a time-frequency graph representing the three-dimensional distribution of the variables in time, frequency and energy;

4) and designing a network model. The depth of the network is from one layer of convolution to four layers of convolution layers, the monitoring time of the deeper network is greatly prolonged, more time is consumed, and the monitoring accuracy is not obviously improved. So the experiment is designed within its proper range and the network with the best fault monitoring performance is used as the experimental model. For the determination of network parameters, the parameters are selected according to the data characteristics of the fermentation process and the accuracy and time consumption of the training model, and the parameters mainly comprise the size number of each convolution layer convolution kernel and the output length of a full connection layer. On the network structure, batch normalization is added after the second layer of convolution before the activation layer, so that the model training precision is improved;

5) inputting a test sample to perform model test, wherein the output result is 0 or 1, and no fault and fault are corresponded;

6) if the accuracy of the test result is higher, outputting the test result for online fault monitoring; and if not, redesigning the network depth and parameters for training.

B. And (3) an online monitoring stage:

7) obtaining data X at a new sampling instant_i(m×n)；

8) Obtaining a required sample by adopting an off-line model data processing mode for the obtained variable data;

9) inputting the online obtained samples into a model test for testing, classifying by using a trained model, outputting to be 0 or 1, and correspondingly judging to be no fault and fault;

10) after the monitoring is finished, if the accuracy of the test result is higher, the online data can be used as a historical data training model.

Advantageous effects

Compared with the prior art, the method adopts the sliding window to dynamically process the two-dimensional data expanded along the variable direction before the convolutional neural network training, and carries out multivariate time-frequency analysis. Hilbert-Huang transform is performed on process variables with time-frequency domain information, including complete time-frequency domain time-varying information of process data. The CNN has excellent multivariable processing and feature recognition capability, so that a complete transformation time-frequency diagram can be used as the input of the CNN to automatically learn high-level variable features so as to obtain better model training performance. The method can effectively improve the accuracy of fault monitoring, and can achieve the fault monitoring accuracy of more than 93% by carrying out experiments on actual production data.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a schematic view of a variable-expanded two-dimensional data sliding window capture;

FIG. 3 is a time-frequency diagram of a Hilbert-Huang transform method for time-frequency analysis of data;

FIG. 4 is a diagram of a model architecture for the process of the present invention;

FIG. 5 is a graph of the results of CNN monitoring a penicillin fermentation process simulation data test set;

FIG. 6 is a graph showing the results of a simulation data test set of the fermentation process of penicillin by DCNN;

FIG. 7 is a graph showing the results of a simulation data test set for penicillin fermentation process according to the present invention;

FIG. 8 is a graph showing the results of CNN monitoring a test set of production data of Escherichia coli fermentation;

FIG. 9 is a graph showing the results of a test set of production data of DCNN on Escherichia coli fermentation;

FIG. 10 is a graph showing the results of a test set of production data obtained by the method of the present invention in Escherichia coli fermentation;

Detailed Description

Penicillin fermentation process

Penicillin is an important antibiotic with high efficiency, low toxicity and wide clinical application, and the preparation process is a typical nonlinear and dynamic intermittent production process. In the experiment, a PenSim2.0 simulation platform is used for simulating the penicillin fermentation process and is used as a data source, the sampling time interval is 0.5h, and 10 process variables are selected, as shown in Table 1. Generating 5 batches of normal working condition data and 5 batches of fault working condition data as training samples in a simulation mode; and 3 batches of fault data of different training samples are used as test samples to verify the effectiveness of the method. The type, magnitude, start and stop times of the fault are shown in table 2.

TABLE 1 variables used in modeling

TABLE 2 Fault setup conditions

Based on the above, the invention is applied to the fermentation process simulation platform, and the specific implementation steps are as follows:

A. an off-line modeling stage:

step 1: 10 batches of normal data X to be generated_10×10×800Expanding along the direction of the variable, and dynamically intercepting data (see the invention content) by adopting a sliding window technology, wherein a batch of acquired data is shown as follows;

step 2: and performing Hilbert-Huang transform on the data dynamically intercepted by the sliding window to obtain a time-frequency diagram. Specifically, the steps of performing empirical mode decomposition and hilbert transform on a plurality of variables are as follows:

firstly, decomposing each variable in data obtained by a sliding window into a finite number of Intrinsic Mode Functions (IMFs) by using an EMD method, wherein each decomposed IMF component comprises a local characteristic signal of a period of time scale of an original signal; then, each decomposed IMF is subjected to spectrum analysis by using Hilbert transform, and the instantaneous frequency of each IMF can be obtained; finally, collecting Hilbert spectrums of all IMFs of j variables to obtain a time-frequency graph representing the three-dimensional distribution of the variables in time, frequency and energy;

and step 3: and establishing a fault monitoring model by adopting a designed deep convolution neural network. In particular, the deep convolutional neural network of the present invention includes: 1 input layer, two convolutional layers, a pooling layer, a convolutional layer with a BP layer, a pooling layer, a full-connection layer and an output layer. The convolution layer kernel size is 5 × 5, the step size is 1, and the number of convolution kernels is respectively: 64. 64 and 32, the window size of the pooling layer is 2 x 2, the step length is 2, the output length of the full connection layer is 20, the output layer has a sample matrix class of a 'softmax' function, and the vector of the output length of the model is 2, is represented as 0 and 1, and is divided into two classes of no fault and fault. The activating function selects a rectifying linear unit (ReLU) function, and a gradient descent method is used for updating the weight and the bias.

And 4, step 4: inputting a test sample to perform model test;

and 5: if the accuracy of the test result is higher, outputting the test result for online fault monitoring; otherwise, redesigning the network for training.

B. And (3) an online monitoring stage:

step 1: data X of 10 process variables at 32 sampling moments of the current fermentation process are collected_32×10；

Step 2: performing time-frequency analysis on the data in an off-line stage data processing mode;

and step 3: the time-frequency diagram input model is used for testing the monitoring result, and the monitoring result is divided into 0,1 and 1 to indicate the fault type;

and 4, step 4: after the monitoring is finished, if the accuracy of the test result is higher, the online data can be used as a historical data training model.

The steps are the specific application of the method in the field of penicillin fermentation simulation platform fault monitoring. To verify the validity of the proposed method, 3 batches of fault data were tested. The experimental results are shown in fig. 5 to 7, and higher degree of overlap of the expected test value and the actual test value indicates higher monitoring accuracy.

Fig. 5, 6, and 7 are diagrams illustrating the CNN method, the DCNN method, and the monitoring results of the method for fault batches provided herein, respectively. It can be seen from fig. 5 that the overlap between the actual test value and the expected value is relatively low, many error points occur, and the accuracy is below 90%; as can be seen from FIG. 6, the test value and the expected value have some non-coincidence points, and the misclassification condition is low, which indicates that the monitoring accuracy is higher than that of the CNN method, and the accuracy is between 90% and 95%; compared with the method, the method disclosed by the invention has the advantages that wrong division and missed division are basically avoided, the test accuracy is up to more than 95%, and the test accuracy is higher and closer to the actual production process by adopting the deep convolution neural network for model training after comprehensively considering the sliding window and the time-frequency analysis.

B Escherichia coli fermentation process

The experimental data for monitoring the escherichia coli fermentation process failure are from actual factory workshops, and the practical production environment often has a plurality of uncertain factors, so that the experimental data has practical significance for algorithm research.

The experiment selects the fermentation data of escherichia coli from a certain biological pharmaceutical factory in Beijing. The fermentation tank capacity is 15L, the fermentation time is 6-7 hours, and the sampling interval is 5 min. Due to the difference between actual data batches, the experiment adopts 78-time sample data with equal length between batches, and selects 7 main process variables with variable names as shown in table 3.

TABLE 3 variables used in modeling

FIGS. 8, 9 and 10 are graphs showing the results of monitoring the production data of E.coli by the CNN method, DCNN and the methods herein, respectively. It can be seen from the figure that compared with the CNN and DCNN methods which have lower accuracy of the monitoring result of the actual data, the method of the invention has certain superiority and can still keep higher monitoring accuracy, because the dynamic characteristics of the production data are better represented during modeling, the deep convolution neural network is trained, the high-level characteristics of the data are extracted in a self-adaptive manner, the model is more accurate, and the on-line monitoring performance is greatly improved.

In order to more vividly compare the effectiveness of the prior method and the effectiveness of the invention applied to the fault monitoring in the escherichia coli process, the method of the invention compares the monitoring effect lists of four methods as follows:

TABLE 4 Fault monitoring results of three methods

As can be readily seen from Table 4, the method of the present invention has improved accuracy and shows better monitoring results than other methods.

Claims

1. A fermentation process monitoring method based on HHT-DCNN is characterized by comprising two stages of 'off-line modeling' and 'on-line modeling', and comprises the following specific steps:

A. an off-line modeling stage:

1) collecting historical data: collecting historical data under the normal working condition of a fermentation process, wherein the historical data is I batch data under the same process in the same fermentation process, and X ═ X (X ═ X)₁，X₂，…X_I，)^TWherein X is_IRepresenting the ith batch of data, each batch containing k sampling instants at which j process variables, i.e., X, are acquired_i＝(x_i，1，x_i，2，…x_i，k) Wherein x is_i，kRepresenting data acquired at the ith sampling instant, X_i，k＝(x_i，k，1，x_i，k，2，…x_i，k，j) Wherein x is_i，k，jA measured value representing a jth process variable at a kth sampling time in the ith batch;

2) data preprocessing: aiming at the dynamics of the fermentation process, adopting a sliding window technology to perform sliding interception on all measured variables in sampling time, enabling the size of a sliding window to be mxn for one batch, enabling m to be m sampling moments of interception, enabling n to be j to represent all measured variables, enabling the sliding step length to be 1, and segmenting process variable data to obtain a plurality of two-dimensional matrix sequences until all batch data sequences are obtained; all the batch data are expanded into a two-dimensional matrix according to variables and expressed as follows:

3) dividing a two-dimensional matrix obtained by sliding a sliding window into a training sample and a testing sample, carrying out Hilbert-Huang transformation on the matrix, and capturing abnormal change information of a sequence on amplitude, frequency and phase: firstly, decomposing a single variable in a sliding window acquisition matrix into a limited number of Intrinsic Mode Functions (IMFs) by using Empirical Mode Decomposition (EMD); then, each decomposed IMF is subjected to spectrum analysis by using Hilbert transform, and the instantaneous frequency of each IMF can be obtained; finally, collecting Hilbert spectrums of all IMFs of j variables to obtain a time-frequency graph representing the three-dimensional distribution of the variables in time, frequency and energy;

4) adopting a deep convolutional neural network to construct a fault monitoring model, inputting a time-frequency graph into the network model for training, and adjusting the weight and the offset of the network by utilizing gradient descent;

5) inputting a test sample to perform model test, and outputting results of no fault and fault;

6) if the test result meets the requirement, outputting the test result for online fault monitoring; if the network depth does not meet the requirement, the network depth is readjusted, and the training is repeated;

B. and (3) an online monitoring stage:

7) obtaining data X at a new sampling instant_i(m×n)；

8) Processing the obtained online monitoring data by adopting the steps 2) and 3) of the offline modeling stage;

9) and inputting the processed on-line monitoring data into the trained fault monitoring model for monitoring.