Fermentation process soft measurement method and device based on deep learning
By constructing the SENet-GRU model, processing multivariable and time series information of the fermentation process, the problem of low prediction accuracy of soft measurement methods in biofermentation engineering is solved, and more efficient fermentation process prediction is achieved.
Patent Information
- Application Number
- CN202411098404.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-29
AI Technical Summary
Existing soft measurement methods have low prediction accuracy for fermentation process variables in biofermentation engineering, especially because of the failure to effectively process time series and dynamic information.
Using a soft measurement method of fermentation process based on deep learning, the SENet-GRU model is constructed by collecting and preprocessing the sample data of the fermentation process, and multivariate and nonlinear data are processed using sliding time windows and SE modules, and time series information is extracted in combination with the GRU network. , improve prediction accuracy.
It significantly improves the accuracy of the prediction of product concentration in the fermentation process, improves model performance and efficiency, and is better than traditional models in terms of prediction effects.
Smart Images

Figure CN119049559A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industrial microbial manufacturing, and in particular to a soft measurement method and device for a fermentation process based on deep learning. Background Art
[0002] The biological fermentation process is an important part of the field of bioengineering and is widely used in many industries such as food, biomaterials, and biopharmaceuticals. With the increase in social demand and the improvement of awareness of resource and environmental protection, the biological fermentation industry urgently needs to improve production efficiency and reduce production costs. Monitoring the fermentation process is the basis for controlling the production process. Therefore, one of the keys to achieving efficient control of the fermentation production process is to monitor key variables or indicators. At present, online sensors such as temperature and pH are widely used, but variables such as product concentration in the fermentation process cannot usually be measured online by sensors. Generally, manual sampling is required and offline measurements are performed using special instruments or experiments, which leads to the problem of untimely monitoring.
[0003] Soft sensor technology is a method that uses mathematical modeling and computer technology to infer or estimate process variables that cannot be directly measured based on known process variable data (such as temperature, pressure, flow, etc.). It can reduce costs and improve product quality in the production process and has important use value. In the field of bio-fermentation engineering, the existing soft sensor technology modeling methods are mainly divided into two types: modeling methods based on mechanism models and modeling methods based on data-driven.
[0004] The modeling method based on the mechanism model is mainly based on the growth mechanism of microorganisms. However, since the main body of the biological fermentation process is the life activities of microorganisms, it involves intracellular metabolic biochemical reactions involving multiple substances such as DNA, mRNA and protein molecules. The fermentation process is complex and highly nonlinear, which brings difficulties to the establishment of mechanism models. In addition, the mechanism and kinetic parameters of the fermentation process of different fungi are different, and the differences in fermentation environments or inoculated strains will also lead to differences in the characteristics of the production process. The established fermentation mechanism model is often only for the fermentation production process of a certain type of bacteria, and the model is less versatile.
[0005] Data-driven modeling methods include statistical reasoning and machine learning techniques. Regression analysis, support vector machines, and artificial neural networks are commonly used in biological fermentation processes. Such soft sensor models can capture and process nonlinear relationships in process data to a certain extent, but only consider the relationship between input features and corresponding outputs during model training, without explicitly processing dynamic information with time series in the data. The fermentation process has certain time series characteristics, such as the concentration of the product is not only related to the bacterial concentration at the current moment, but also to the bacterial concentration and process status at the past moment. Therefore, the accuracy of such traditional machine learning models is subject to certain limitations.
[0006] In summary, in the field of bio-fermentation engineering, the models constructed by existing soft sensing methods have the problem of low prediction accuracy of variables. Summary of the invention
[0007] Therefore, the technical problem to be solved by the present invention is to overcome the problem that the model constructed by the soft measurement method in the prior art has low prediction accuracy for variables.
[0008] In order to solve the above technical problems, the present invention provides a fermentation process soft sensing method based on deep learning, comprising:
[0009] Collecting sample data of the fermentation process, wherein the sample data includes auxiliary variable data and target variable data;
[0010] Preprocess the sample data, including:
[0011] Connect the sample data end to end in time sequence along the batch direction and expand them into two-dimensional data;
[0012] Normalize the expanded sample data;
[0013] Construct a sliding time window, and divide the normalized sample data into training sample data according to the sliding time window;
[0014] Construct a fermentation soft-sensing model and use the training sample data to train the fermentation product concentration prediction model;
[0015] The auxiliary variable data of the fermentation process are collected in real time according to the sliding time window, and after preprocessing, they are input into the trained fermentation soft sensor model to obtain the predicted data of the target variable.
[0016] Preferably, the sample data includes three dimensions: time series, batch, and variable.
[0017] Preferably, the target variable is the fermentation product concentration, and the auxiliary variables include aeration rate, stirring speed, feed rate, dissolved oxygen concentration, total volume of fermentation broth, carbon dioxide concentration, pH value and fermentation temperature.
[0018] Preferably, the expanded sample data is normalized using the Min-Max method.
[0019] Preferably, when the normalized sample data is divided into training sample data according to the sliding time window, the time window containing sample data at two different batches of time points is deleted.
[0020] Preferably, the fermentation soft sensor model is a SENet-GRU model, comprising a SENet and a GRU network connected in sequence.
[0021] Preferably, the SENet comprises an input layer, a first convolutional layer, a SE module, a second convolutional layer, a fully connected layer and an output layer which are connected in sequence.
[0022] Preferably, a sequence expansion layer and a flattening layer are arranged between the output layer of SENet and the GRU network; the sequence expansion layer restores the output of SENet into three-dimensional sequence data having three dimensions of features, time and batch, and inputs it into the flattening layer; the flattening layer reformats the three-dimensional sequence data so that the three-dimensional sequence data meets the requirements of the GRU network input.
[0023] Preferably, a Bayesian optimization algorithm is used to determine the hyperparameters of the SENet-GRU model.
[0024] The present invention also provides a fermentation process soft measurement device based on deep learning, comprising:
[0025] A collection module, used for collecting sample data of the fermentation process, wherein the sample data includes auxiliary variable data and target variable data;
[0026] The preprocessing module is used to preprocess the sample data, including:
[0027] Connect the sample data end to end in time sequence along the batch direction and expand them into two-dimensional data;
[0028] Normalize the expanded sample data;
[0029] Construct a sliding time window, and divide the normalized sample data into training sample data according to the sliding time window;
[0030] Model building and training module, used to build a fermentation soft-sensing model and train the fermentation product concentration prediction model using training sample data;
[0031] The target variable prediction module is used to collect auxiliary variable data of the fermentation process in real time according to the sliding time window, and input it into the trained fermentation soft measurement model after preprocessing to obtain the predicted data of the target variable.
[0032] The above technical solution of the present invention has the following beneficial effects compared with the prior art:
[0033] The present invention discloses a soft-sensing method for a fermentation process based on deep learning. When training a model, it is considered that the sample data of an intermittent fermentation process also has batch information compared to general time series data. The three-dimensional sample data of time series, batch, and variable are expanded in the batch direction, and a sliding time window is constructed to extract the time series information in the window, so as to make full use of the sample information of multiple batches to train the fermentation soft-sensing model and improve the performance and efficiency of the model. In addition, the present invention abandons the pooling layer in the convolutional neural network, designs a simpler two-layer convolutional network, and introduces the SE module between the two convolutional layers to form SENet, so as to process the multivariate and nonlinearity in the intermittent fermentation data and extract the data features; and uses the GRU network to extract the time series information, and constructs the SENet-GRU model as the fermentation soft-sensing model according to the characteristics of the dynamic and nonlinearity of the fermentation process, so as to improve the prediction accuracy of the target variable by the soft-sensing of the fermentation process. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:
[0035] Figure 1 It is a flow chart of a soft measurement method of a fermentation process based on deep learning of the present invention;
[0036] Figure 2 is a schematic diagram of data preprocessing, where Figure 2 (a) is a schematic diagram of batch expansion. Figure 2 (b) is a schematic diagram of sample data after two time batches are expanded;
[0037] Figure 3 It is a structural diagram of the SENet model constructed by the present invention;
[0038] Figure 4 This is a single GRU unit structure diagram involved in the present invention;
[0039] Figure 5 It is the structure diagram of the SENet-GRU model proposed in the present invention;
[0040] Figure 6 is a comparison chart of the implementation effect of the present invention on the penicillin fermentation data set, wherein Figure 6 (a) in the figure is the prediction fitting curve of the Bayes-GRU model. Figure 6 (b) in the figure is the predicted fitting curve of LeNet. Figure 6 (c) in the figure is the prediction fitting curve of SENet. Figure 6(d) is the prediction fitting curve diagram of the present invention. DETAILED DESCRIPTION
[0041] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.
[0042] Reference Figure 1 As shown, the present invention proposes a fermentation process soft sensing method based on deep learning, which is used to predict the product concentration of the fermentation process online. The steps include:
[0043] S1. Collecting sample data of a fermentation process, wherein the sample data includes auxiliary variable data and target variable data.
[0044] Specifically, the sample data includes three dimensions: time series, batch, and variable. The target variable is the fermentation product concentration, and the auxiliary variables should be selected as variables that have a certain correlation with the target variable and are easy to obtain. Therefore, the auxiliary variables selected in this embodiment include ventilation rate (L / h), stirring speed (r / min), feed rate (L / h), dissolved oxygen concentration (mmol / L), total volume of fermentation liquid (L), carbon dioxide concentration (mmol / L), pH value and fermentation temperature (K) and other variables.
[0045] S2. In order to facilitate the learning of the fermentation soft sensor model, the sample data is preprocessed and the three-dimensional sample data is expanded into two-dimensional data along the batch direction.
[0046] The fermentation process is usually carried out in batches. Each batch of fermentation has an independent operating procedure, raw material input, and fermentation time, which makes the production data between batches independent. However, the fermented products of different batches are the same and have the same variables, so they are also consistent. Therefore, for each batch, sample data is collected at the same interval. These data have the same structure but are independent of each other. In the intermittent fermentation process, because the fermentation data structure of each batch is parallel, the sample data of different batches can be connected end to end in time order to form a set of sample data with only two dimensions. In order to make full use of the data information between different batches, it is necessary to perform pre-processing steps such as batch expansion on the data. Refer to Figure 2 As shown, the specific steps are as follows:
[0047] S201. The fed-batch fermentation process is carried out in batches. The sample data is first expanded in the batch direction, and the total number of batches is recorded as I. The number of auxiliary variables collected is determined according to analysis and needs, and is recorded as J. Sample data is collected at a fixed time interval Δt in each batch until the fermentation is completed, and the total number of samples collected is recorded as K.
[0048] Batch expansion diagram reference Figure 2 As shown in (a), the 1 to K sample data of the second batch are connected to the Kth data of the first batch, and the sequence of the new set of data becomes 1 to 2K. Similarly, after all batches of sample data are connected end to end in time sequence, a set of expanded data with the batch dimension of the time series data reduced to 1 is formed, which is recorded as X1(J×IK). The expanded data X1 is equivalent to a set of data without batch dimension.
[0049] S202, normalize the expanded data X1. Use the Min-Max method to normalize the data of each auxiliary variable, and the calculation formula is as follows:
[0050]
[0051] in, is the normalized value of the jth variable of the i-th sample in the expanded data, is the sampling value of the jth variable of the i-th sample in the expanded data, and are the maximum and minimum values of the j-th variable in the expanded data, i=1,2,…,IK, j=1,2,…,J.
[0052] S203: construct a sliding time window, and divide the normalized sample data into training sample data according to the sliding time window.
[0053] The data of the fermentation process has obvious time series information, but when processing, it is necessary to consider that the state at a certain time point will have less impact on subsequent time points as time goes by, so a sliding time window can be constructed to select sample data of several sampling points before the time point to be predicted, and to train and predict the model by combining the time series information and the variable state information of the sample data. This can not only capture the information of the time points with greater correlation, but also reduce the amount of calculation during network training.
[0054] The window length of the sliding time window is k, and the sliding step size is Δt.
[0055] It should be noted that the sample data of consecutive time points are combined to obtain the data within the sliding time window, and when the batch is expanded, the sample data of the connecting part of two batches has no time continuity. Therefore, in order to ensure the validity of the data time series information, when the normalized sample data is divided into training sample data according to the sliding time window, the time window containing the sample data of two different batch time points needs to be deleted. Figure 2 (b) is a schematic diagram of sample data after two time batches are expanded. The two colors represent two batches, where the time windows x'1 = {x1, x2, x3}, x'2 = {x2, x3, x4} and x' n-2={x n-2 ,x n-3 ,x n} all come from the same batch and are valid data; while the time window x' invalid The gray and white sample data are from two different batches. Since the time window containing sample data from two different batches is invalid data, x' invalid It is invalid data and should be deleted during training.
[0056] S3. Construct a fermentation soft-sensing model and use the training sample data to train the fermentation product concentration prediction model.
[0057] In recent years, deep learning in machine learning technology has developed rapidly, with the ability to analyze and reason about complex problems, providing a new path for biofermentation technology. Among them, convolutional neural networks (CNN) are often used to extract hidden relationships and features in data, and researchers have also continuously developed various variants of convolutional neural networks to improve performance. For example, by adding SE modules to form SENet, importance weights can be added to the feature channels of the network, improving the ability of the original convolutional neural network framework to extract features. Unlike traditional machine learning models that process static relationships, recurrent neural networks (RNNs) have the ability to process input sequence time series information, and can extract dynamic information in the fermentation process. Basic recurrent neural networks will have problems with gradient disappearance or gradient explosion during sequence training, and thus LSTM (Long Short-Term Memory, LSTM) and GRU (Gate Recurrent Unit, GRU) have been developed. Both can solve the problems of traditional recurrent neural networks, and the network structure of GRU is simpler than LSTM and more efficient in processing data.
[0058] In this embodiment, the fermentation soft sensor model is a SENet-GRU model, including two network model structures of SENet and GRU networks, which will be further described below with reference to the accompanying drawings.
[0059] The structural diagram of SENet is as follows Figure 3 As shown, it includes a convolutional neural network and a SE module. The convolutional neural network includes two convolutional layers and does not contain a pooling layer. The SE module is added between the two convolutional layers to form SENet. The addition of the SE module can explicitly extract the importance of each feature in the feature channel of the convolutional neural network, so that the output of the network is adjusted according to such importance to obtain a more accurate output result.
[0060] S301. According to the data characteristics of the fermentation process, a convolutional neural network structure is constructed as the basis of SENet.
[0061] The structure of the convolutional neural network was created for application in the field of images. However, since the size and scale of the fermentation process data are far different from image data, the network structure needs to be adjusted according to the characteristics of the data. The size of the image input is usually 224x224~448x448. The pooling layer in the convolutional neural network structure is set up to avoid the high-dimensional data formed after multiple convolution operations, which causes computational difficulties for the network. The input dimension of the sample data of the fermentation process, that is, the number of variables, is generally around 10, which is much smaller than the image data input, so the pooling layer in the general convolutional neural network structure is removed. Such a design can not only make the network structure simpler, but also avoid the information loss that may be caused by the pooling operation. In addition, in the convolutional neural network structure, the convolutional layer is used to extract features from the data, but the more layers, the better the effect. This also requires balancing multiple factors such as data scale, convolutional layer extraction capability, and training difficulty to select the number of convolutional layers.
[0062] Through experimental comparison and verification, the convolutional neural network constructed by the present invention includes an input layer, two convolutional layers, a fully connected layer and an output layer.
[0063] S302. Build SENet.
[0064] The SE module is added between the two convolutional layers of the convolutional neural network constructed in S401 to form a SENet that can consider the importance of features during the convolution process. The SENet includes an input layer, a first convolutional layer, a SE module, a second convolutional layer, a fully connected layer, and an output layer connected in sequence.
[0065] The first convolution layer includes m1 convolution kernels of size [3, 1], and the second convolution layer includes m2 convolution kernels of size [3, m1].
[0066] Considering that the feature matrix after the first convolution layer still retains part of the physical meaning of the input, the SE module is selected after the first convolution layer so that the network can learn the importance of each feature channel at this stage and obtain feature matrices with different importance. The output of the SE module is then convolved for the second time to fully extract the feature information.
[0067] Among them, the SE module performs the following three steps:
[0068] (1) Squeeze: Through global average pooling, the global information U with dimensions W×H×C is compressed into a new feature vector with dimensions 1×1×C, making it easier to process the relationship between channels. The output of the front part of the SE module is recorded as U=[u1,u2,…,u c ], then the calculation equation for the squeeze operation of the cth element is as follows:
[0069]
[0070] exist Figure 3 In the example shown, H is equal to n, W is equal to 1, and C is equal to m1. After the above processing, a global feature representation is obtained, and the data dimension becomes 1×1×m1.
[0071] (2) Excit ion: Perform a series of nonlinear mappings on the feature vectors obtained by the squeeze operation to obtain the weight value of each channel, which represents the importance of different channels. This can be achieved by using a gating mechanism with a sigmoid function. The calculation equation is as follows:
[0072] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))
[0073] Where z is the result of the Squezze step, δ is the ReLU layer, and W1 is the dimension A real matrix of dimension W2. The dimension of the result z obtained by the Squezze step is 1×1×C, and the dimension of the data obtained after multiplying with W1 is After the ReLU operation, the dimension of the data remains unchanged. Then, it is multiplied by W2, and the resulting data dimension is 1×1×C. Figure 3 Medium, F ex The output s is a 1×1×m1 matrix, which represents the weight of each feature channel.
[0074] (3) Scale: Apply the weights obtained by the Excitation operation to the corresponding channel features. The size of the weights represents the importance of these features. Such weights can be used to calibrate the importance of the original input after the Scale operation. Usually, a multiplication step is used to achieve this goal, that is, the channel is multiplied by the weight coefficient. The calculation equation is as follows:
[0075]
[0076] S303, constructing a SENet-GRU model as a fermentation soft sensor model.
[0077] The biological fermentation data is input data with time series. In order to make full use of the information, a GRU network is added after SENet to extract the time information. Since SENet built on a convolutional neural network processes static data, and the GRU network processes time series data, special attention should be paid to the shape conversion of the data when inputting the output of SENet into the GRU network. Specifically, the data is adjusted to a form that each structure can handle by setting the following network layers:
[0078] A sequence expansion layer and a flattening layer are set between the output layer of SENet and the GRU network; the sequence expansion layer restores the output of SENet into three-dimensional sequence data with three dimensions of features, time, and batch, and inputs it into the flattening layer; the flattening layer reformats the three-dimensional sequence data so that the three-dimensional sequence data meets the requirements of the GRU network input.
[0079] The structure diagram of a single unit in the GRU network is as follows Figure 4 As shown in the figure, it has two gate structures, the reset gate and the update gate, which can handle the problems related to time series data. t To reset the door state, z t To update the door state, reset the door to calculate the parameter r in the range of (0,1) t , and apply it to the state information h of the previous unit t-1 , get the hidden state of the current unit make The update gate is responsible for receiving and integrating the information of the reset gate and the previous unit to update the state of the current unit h t The network parameters and state update formulas are as follows:
[0080]
[0081] r t j =σ(W r x t +U r h t-1 ) j
[0082]
[0083] Among them, W z with U z are the weight matrices of the input and hidden states in the update gate, respectively. r and U r are the weight matrices of the input and hidden states in the reset gate, respectively. h and U hThe weight matrices from the hidden state adjusted by the input and reset gates to the candidate hidden state, respectively.
[0084] The structural diagram of the fermentation soft sensor model is shown in Figure 2. Figure 5 In order to obtain a better hyperparameter combination for the SENet-GRU model, the Bayesian optimization algorithm is used to optimize it and determine the hyperparameters of the SENet-GRU model.
[0085] S4. Auxiliary variable data of the fermentation process are collected in real time according to the sliding time window, and input into the trained fermentation soft sensor model after preprocessing to obtain the predicted data of the target variable.
[0086] In summary, the present invention provides a kind of fermentation process soft measurement method based on deep learning. When training the model, it is considered that the sample data of the intermittent fermentation process also has batch information compared with the general time series data. The three-dimensional data of the time series, batch, and variable are expanded in the batch direction, and a sliding time window is constructed to extract the time series information in the window, so as to make full use of the sample information of multiple batches to train the fermentation soft measurement model, and improve the performance and efficiency of the model. In addition, the present invention abandons the pooling layer in the convolutional neural network, designs a simpler two-layer convolutional network, and introduces the SE module between the two convolutional layers to form SENet, so as to process the multivariate and nonlinearity in the intermittent fermentation data and extract the data features; and the GRU network is used to extract the time series information, and the SENet-GRU model is constructed as the fermentation soft measurement model according to the characteristics of the dynamic and nonlinearity of the fermentation process, so as to improve the prediction accuracy of the soft measurement of the fermentation process for the variable.
[0087] In order to illustrate that the soft sensing method for fermentation process based on deep learning described in the present invention is effective, the following simulation experiment is used for analysis:
[0088] 9 batches of penicillin fermentation data were selected, each batch had a reaction time of 400h, and the required variables were sampled at intervals of 1h. 8 batches of data were randomly selected to train a variety of soft sensor models including the method of the present invention, and the trained models were used to predict the remaining batch. In order to evaluate the prediction accuracy of each model, the mean absolute error (MAE) and the root mean square error (RMES) were set as evaluation indicators, and the calculation formula was as follows:
[0089]
[0090] Where y represents the observed value, that is, the target variable data, represents the predicted value, that is, the predicted data of the target variable output by the fermentation soft sensor model, N represents the number of sample data, and the smaller the calculated RMSE and MAE, the better the performance of the model.
[0091] The target variable prediction results obtained by the present invention are compared with other existing models. The comparison results are recorded in Table 1. The prediction fitting curve is shown in Table 1. Figure 6 shown. Figure 6 (a) in the figure is the prediction fitting curve of the Bayes-GRU model. Figure 6 (b) in the figure is the predicted fitting curve of LeNet. Figure 6 (c) in the figure is the prediction fitting curve of SENet. Figure 6 (d) is the prediction fitting curve diagram of the present invention.
[0092] Table 1 Comparison of prediction effects of different models
[0093]
[0094] Among them, Bayes-GRU uses the same hyperparameter optimization method for a separate GRU model, LeNet is a classic CNN network model, and SENet is a model without GRU in step 4. It can be seen from the results that compared with other models, the soft measurement method based on the SENet-GRU model proposed in the present invention has the smallest RMES and MAE values. This shows that the modification of the CNN structure by the method improves the prediction accuracy, and the introduction of GRU in SENet is also helpful to improve the prediction accuracy, which illustrates the effectiveness of the method proposed in the present invention.
[0095] Based on the above-mentioned fermentation process soft measurement method based on deep learning, this embodiment further provides a fermentation process soft measurement device based on deep learning, including:
[0096] A collection module, used for collecting sample data of the fermentation process, wherein the sample data includes auxiliary variable data and target variable data;
[0097] The preprocessing module is used to preprocess the sample data, including:
[0098] Connect the sample data end to end in time sequence along the batch direction and expand them into two-dimensional data;
[0099] Normalize the expanded sample data;
[0100] Construct a sliding time window, and divide the normalized sample data into training sample data according to the sliding time window;
[0101] Model building and training module, used to build a fermentation soft-sensing model and train the fermentation product concentration prediction model using training sample data;
[0102] The target variable prediction module is used to collect auxiliary variable data of the fermentation process in real time according to the sliding time window, and input it into the trained fermentation soft measurement model after preprocessing to obtain the predicted data of the target variable.
[0103] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0104] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0105] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0106] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0107] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.
Claims
1. A soft sensing method for fermentation process based on deep learning, characterized in that: include: Collecting sample data of the fermentation process, wherein the sample data includes auxiliary variable data and target variable data; Preprocess the sample data, including: Connect the sample data end to end in time sequence along the batch direction and expand them into two-dimensional data; Normalize the expanded sample data; Construct a sliding time window, and divide the normalized sample data into training sample data according to the sliding time window; Construct a fermentation soft-sensing model and use the training sample data to train the fermentation product concentration prediction model; The auxiliary variable data of the fermentation process are collected in real time according to the sliding time window, and after preprocessing, they are input into the trained fermentation soft sensor model to obtain the predicted data of the target variable.
2. A fermentation process soft sensing method based on deep learning according to claim 1, characterized in that: The sample data includes three dimensions: time series, batch, and variable.
3. The fermentation process soft sensing method based on deep learning according to claim 1, characterized in that: The target variable was the concentration of fermentation products, and the auxiliary variables included aeration rate, stirring speed, feed rate, dissolved oxygen concentration, total volume of fermentation broth, carbon dioxide concentration, pH value and fermentation temperature.
4. The fermentation process soft sensing method based on deep learning according to claim 1, characterized in that: The Min-Max method is used to normalize the expanded sample data.
5. The fermentation process soft sensing method based on deep learning according to claim 1, characterized in that: When the normalized sample data is divided into training sample data according to the sliding time window, the time window containing sample data at two different batch time points is deleted.
6. A fermentation process soft sensing method based on deep learning according to claim 1, characterized in that: The fermentation soft sensor model is a SENet-GRU model, including a SENet and a GRU network connected in sequence.
7. A fermentation process soft sensing method based on deep learning according to claim 6, characterized in that: The SENet includes an input layer, a first convolutional layer, a SE module, a second convolutional layer, a fully connected layer and an output layer which are connected in sequence.
8. The fermentation process soft sensing method based on deep learning according to claim 7, characterized in that: A sequence expansion layer and a flattening layer are set between the output layer of SENet and the GRU network; the sequence expansion layer restores the output of SENet into three-dimensional sequence data with three dimensions of features, time, and batch, and inputs it into the flattening layer; the flattening layer reformats the three-dimensional sequence data so that the three-dimensional sequence data meets the requirements of the GRU network input.
9. A fermentation process soft sensing method based on deep learning according to claim 6, characterized in that: The hyperparameters of the SENet-GRU model are determined using the Bayesian optimization algorithm.
10. A fermentation process soft measurement device based on deep learning, characterized in that: include: A collection module, used for collecting sample data of the fermentation process, wherein the sample data includes auxiliary variable data and target variable data; The preprocessing module is used to preprocess the sample data, including: Connect the sample data end to end in time sequence along the batch direction and expand them into two-dimensional data; Normalize the expanded sample data; Construct a sliding time window, and divide the normalized sample data into training sample data according to the sliding time window; Model building and training module, used to build a fermentation soft-sensing model and train the fermentation product concentration prediction model using training sample data; The target variable prediction module is used to collect auxiliary variable data of the fermentation process in real time according to the sliding time window, and input it into the trained fermentation soft measurement model after preprocessing to obtain the predicted data of the target variable.
Citation Information
Patent Citations
Penicillin fermentation process online soft measurement method and device based on deep learning
CN116469478A
Photovoltaic output prediction method combined with channel attention mechanism and terminal
CN117039983A
1D-CNN blood glucose concentration prediction method based on activation function
CN118452912A
Cited By
Fermentation process prediction method and system based on generative algorithm, and terminal
CN119560002A
Fermentation process prediction method, system and terminal based on generative algorithm
CN119560002B
