Penicillium culture process prediction method and system based on metabolic kinetics decoupling and hybrid expert network

Abstract

The invention discloses a penicillium culture process prediction method and system based on metabolic kinetics decoupling and a hybrid expert network, and the method comprises the steps: constructing a growth trend-metabolic rhythm-random disturbance decoupling framework, and obtaining a random disturbance sequence based on the decoupling framework; constructing a hybrid expert network of four parallel expert networks, and performing multi-step prediction on the random disturbance sequence based on the hybrid expert network to obtain four expert prediction results; constructing fermentation state vectors, inputting the fermentation state vectors into a gating network to obtain weights corresponding to the four experts, and performing weighted average on expert prediction results based on the weights to obtain disturbance term multi-step prediction results; and based on the multi-step prediction result of the disturbance term, combining the extrapolated trend term and rhythm term to obtain a final prediction result. According to the method, specialized processing and self-adaptive fusion of different dynamic characteristics are functionally realized, so that the prediction precision, robustness and interpretability of the penicillium culture process can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of penicillium culture prediction, and particularly relates to a penicillium culture process prediction method and system based on metabolic kinetics decoupling and mixed expert network. BACKGROUND

[0002] Penicillin, as an important antibiotic, its production mainly depends on the microbial fermentation process of penicillium (Penicillium chrysogenum). The process is an extremely complex nonlinear, time-varying, multivariable strongly coupled biochemical reaction system. In the fermentation process, there is a complex dynamic relationship between key state variables such as biomass, penicillin yield, substrate (such as glucose, nitrogen source) concentration, pH value, dissolved oxygen (DO) and so on. The accurate prediction of the future state of these key variables is of great significance for optimizing the feeding strategy, regulating the environmental parameters, improving the production efficiency and reducing the production cost.

[0003] At present, the prediction methods for fermentation process mainly include mechanism model and data-driven model.

[0004] 1. Mechanism model: based on biochemical reaction kinetics (such as Monod equation, Luedeking-Piret equation) to construct mathematical equation set describing the growth of bacteria, substrate consumption and product synthesis. The advantage of this kind of model is that it has clear physical and biological significance and strong explanation. However, its shortcomings are also very prominent: first, the metabolic network of penicillium is extremely complex, and it is almost impossible to establish a complete mechanism model that can accurately describe all reactions; second, the model contains a large number of kinetic parameters that are difficult to measure online or change with the degradation of the strain, which makes parameter identification difficult, resulting in poor model generalization ability and difficulty in adapting to batch differences in actual production.

[0005] 2. Data-driven model: With the development of artificial intelligence technology, methods using machine learning or deep learning models (such as artificial neural networks ANN, support vector machines SVM, recurrent neural networks RNN, long short-term memory networks LSTM, etc.) to learn the dynamic characteristics of historical fermentation data have been increasingly concerned. Such models do not require in-depth understanding of complex biochemical mechanisms and can fit the relationship between input (such as historical state, control variable) and output (future state) in a "black box" or "gray box" manner. However, existing data-driven models still face the following challenges when applied to penicillium cultivation process prediction: (1) Non-stationarity problem: Fermentation process has obvious stages (lag phase, logarithmic phase, stationary phase, decline phase), and the dynamic characteristics of each stage are different. A single structure of deep learning model is often difficult to adapt to the non-stationary data distribution of all stages at the same time, and is prone to produce large prediction bias at the stage transition. (2) Multi-scale feature coupling: The fermentation data contains dynamic information of multiple time scales, such as long-term trend representing the slow proliferation of bacteria, medium and short-term rhythmic fluctuations related to cell metabolic activity, and instantaneous pulse response caused by feeding, pH regulation and other operations. Existing models usually mix these features together for processing, making it difficult to effectively separate and learn, resulting in insufficient ability to capture key details. (3) Insufficient model generalization and robustness: In actual production, due to factors such as raw material batch difference, inoculum fluctuation, environmental disturbance, etc., there are significant batch-to-batch differences in the fermentation process. Pure data-driven models are prone to overfitting when the training data is not fully covered, and the prediction performance for unseen conditions or extreme events (such as contamination, equipment failure precursor) is poor. SUMMARY

[0006] The present application aims to solve the problems of the prior art and provides the following solution:

[0007] The penicillium cultivation process prediction method based on metabolic kinetics decoupling and mixed expert network comprises the following steps:

[0008] A growth trend-metabolic rhythm-random disturbance decoupling framework is constructed, and a random disturbance sequence is obtained based on the decoupling framework;

[0009] A mixed expert network of four parallel expert networks is constructed, and a multi-step prediction is performed on the random disturbance sequence based on the mixed expert network to obtain four expert prediction results;

[0010] A fermentation state vector is constructed and input into a gating network to obtain weights corresponding to the four experts, and a weighted average of the expert prediction results is performed based on the weights to obtain a disturbance item multi-step prediction result;

[0011] Based on the disturbance item multi-step prediction result, the trend item and the rhythm item are extrapolated to obtain the final prediction result.

[0012] Preferably, the decoupled framework is:

[0013]

[0014] wherein, denotes a time series of key process variables, denotes a growth trend component, denotes a metabolic rhythm component, denotes a random disturbance sequence.

[0015] Preferably, the growth trend component is:

[0016]

[0017]

[0018] wherein, denotes a first set of learnable parameters of the model, denotes a maximum variable value that the fermentation process can reach, denotes a coefficient related to the initial state, denotes a growth / variation rate, denotes a time point at which half of the maximum rate is reached.

[0019] Preferably, the metabolic rhythm component is:

[0020]

[0021]

[0022] wherein, denotes a second set of learnable parameters of the model, denotes a preset number of rhythm components, denotes an amplitude of the i-th rhythm, denotes a center of a main time window of the i-th rhythm activity, denotes a width of the main time window of the i-th rhythm activity, denotes a frequency of the i-th rhythm, denotes a phase of the i-th rhythm.

[0023] Preferably, the hybrid expert network comprises: a spatial convolution network, a gated recurrent unit network, a biochemical kinetics constraint network, and a self-attention mechanism network;

[0024] The spatial convolution network extracts local patterns and morphological features in the random disturbance sequence by using a plurality of one-dimensional convolution kernels to obtain a first expert prediction result;

[0025] The gating recurrent unit network is used to capture long-range temporal dependencies and memory effects in the random disturbance sequence, to obtain a second expert prediction result;

[0026] The biochemical kinetics constraint network embeds biochemical reaction rate equations to constrain the prediction of the random disturbance sequence, to obtain a third expert prediction result; wherein the biochemical reaction rate equation is , represents the predicted concentration of the key substrate, represents the real-time concentration of the key substrate, and C represents other process variables, and U represents control variables;

[0027] The self-attention mechanism network captures non-continuous and long-distance event correlations by calculating the correlation weights between each time point in the random disturbance sequence, to obtain a fourth expert prediction result.

[0028] Preferably, the fermentation state vector is:

[0029]

[0030] wherein, represents the one-hot encoding of the current fermentation phase determined according to the slope of the growth trend component, 、 、 、 represents the measurement value of each key variable at the current time, represents the recent change rate of the key variable.

[0031] Preferably, the method for obtaining the multi-step prediction result of the disturbance term comprises:

[0032]

[0033] wherein, represents the multi-step prediction result of the disturbance term, represents the weight corresponding to the four experts, represents the prediction result of the four experts, h represents the index of a single prediction step, and H represents the total number of multi-step prediction.

[0034] Preferably, the method for obtaining the final prediction result comprises:

[0035]

[0036] wherein, represents the final prediction result, represents the trend term obtained by forward deduction of the learned parameter model by h steps, represents the rhythm term obtained by forward deduction of the learned parameter model by h steps.

[0037] The application also provides a penicillium culture process prediction system based on metabolic kinetics decoupling and mixed expert networks, which applies the method and comprises a decoupling framework modeling module, a mixed network prediction module, a weighting module and a result integration module.

[0038] The decoupling framework modeling module is used to construct a growth trend-metabolic rhythm-random disturbance decoupling framework and obtain a random disturbance sequence based on the decoupling framework.

[0039] The mixed network prediction module is used to construct a mixed expert network of four parallel expert networks and perform multi-step prediction on the random disturbance sequence based on the mixed expert network to obtain four expert prediction results.

[0040] The weighting module is used to construct a fermentation state vector and input the fermentation state vector into a gating network to obtain weights corresponding to the four experts, and perform weighted average on the expert prediction results based on the weights to obtain a disturbance item multi-step prediction result.

[0041] The result integration module obtains a final prediction result based on the disturbance item multi-step prediction result and combines extrapolated trend items and rhythm items.

[0042] Compared with the prior art, the application has the following beneficial effects:

[0043] The application decomposes complex, non-trivial fermentation process time series into three sub-sequences with different physical meanings and dynamic characteristics through an innovative “growth trend-metabolic rhythm-random disturbance” (GMS) decoupling framework, then uses an adaptive weighted heterogeneous expert network to finely model the most difficult to predict random disturbance item, and finally reconstructs and outputs a high-precision multi-step prediction result. Through the above design, the application not only effectively decouples and hierarchically models complex biological processes in structure, but also functionally realizes specialized processing and adaptive fusion of different dynamic characteristics, thereby significantly improving the prediction accuracy, robustness and interpretability of the penicillium culture process. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings in the following description only constitute some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0045] Figure 1 The method flowchart of the embodiment of the application. DETAILED DESCRIPTION

[0046] With reference to the drawings and specific embodiments, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0047] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0048] Embodiment one

[0049] In this embodiment, as shown in the following formula, based on the penicillium culture process prediction method of metabolic kinetics decoupling and mixed expert network, the following steps are included: Figure 1

[0050] S1. Construct a growth trend-metabolic rhythm-random disturbance decoupling framework, and obtain a random disturbance sequence based on the decoupling framework.

[0051] In this embodiment, for the time series of any key process variable (such as biomass concentration), the value at time t is decomposed into the sum of three components, and the obtained decoupling framework is as follows:

[0052]

[0053] wherein, represents the time series of the key process variable, represents the growth trend component, represents the metabolic rhythm component, represents the random disturbance sequence.

[0054] The growth trend component describes the macroscopic growth curve of the bacteria in the entire fermentation period, which usually presents an "S" shape, and is specifically as follows:

[0055]

[0056]

[0057] wherein, represents the first set of learnable parameters of the model, represents the maximum variable value that can be reached in the fermentation process, represents the coefficient related to the initial state, represents the growth / change rate, ​​denote the time points when the rate reaches half of the maximum rate. These parameters are dynamically generated by a small neural network according to initial fermentation conditions (such as inoculum size, initial substrate concentration) and historical data segments, so that the trend term can adapt to the initial differences of different batches.

[0058] The metabolic rhythm component captures the medium and short-term fluctuations under the macroscopic growth trend, caused by the periodic conversion of the internal metabolic state of the cells (such as respiratory oscillation, changes in key enzyme activity). Instead of using a fixed Fourier basis, the invention designs a rhythm model composed of multiple learnable, phase-variable Gaussian kernel functions, specifically:

[0059]

[0060]

[0061] where, denote the second set of learnable parameters of the model, denote the preset number of rhythm components, denote the amplitude of the i-th rhythm, denote the center of the main time window of the i-th rhythm activity, denote the width of the main time window of the i-th rhythm activity, denote the frequency of the i-th rhythm, denote the phase of the i-th rhythm. This form can flexibly capture metabolic oscillations that appear or disappear at different stages of fermentation, which are not strictly periodic.

[0062] is a random disturbance sequence, which is the residual term after subtracting the growth trend and metabolic rhythm from the original sequence, This component contains unstructured, high-frequency dynamic information caused by feeding operations, small fluctuations in pH and temperature, measurement noise, and other unmodeled complex biological effects. It is the most difficult but crucial part of the prediction.

[0063] S2. Construct a hybrid expert network of 4 parallel expert networks, and based on the hybrid expert network, respectively, multi-step prediction is performed on the random disturbance sequence to obtain 4 expert prediction results.

[0064] In this embodiment, a MoE structure containing four parallel expert networks is designed, which is specifically used for multi-step prediction of the random disturbance sequence The input is the historical disturbance sequence window and the historical windows of other related process variables (such as pH, DO, substrate concentration) and control variables (such as feeding rate). Each expert network has a different inductive bias, extracting information from different angles.

[0065] The hybrid ensemble of experts includes a spatial convolution network, a gated recurrent unit network, a biochemical kinetics constrained network, and a self-attention mechanism network.

[0066] The spatial convolution network extracts local patterns and morphological features in the random disturbance sequence using multi-layer one-dimensional convolution kernels, such as sharp impulse responses or rapid oscillations caused by feed supplementation, to obtain the first expert prediction result.

[0067] The gated recurrent unit network, as a variant of RNN, is used to capture long-range temporal dependencies and memory effects in the random disturbance sequence to obtain the second expert prediction result. The gated recurrent unit network is suitable for learning slow-changing autocorrelations in the disturbance term.

[0068] The biochemical kinetics constrained network embeds biochemical reaction rate equations to constrain the prediction of the random disturbance sequence to obtain the third expert prediction result; wherein the biochemical reaction rate equation is , represents the predicted concentration of the key substrate (e.g. glucose), represents the real-time concentration of the key substrate (e.g. glucose), C represents other process variables, and U represents control variables. The network does not directly predict , but predicts unknown parameters or correction terms in the rate equation, and then obtains the prediction result by numerical integration. This expert network ensures that the prediction follows the basic material balance and kinetic constraints in the short term.

[0069] The self-attention mechanism network captures non-continuous, long-distance event associations by calculating the correlation weights between each time point in the random disturbance sequence, such as how a small disturbance in the early stage affects the stability of the system in the later stage, to obtain the fourth expert prediction result.

[0070] S3. Construct a fermentation state vector and input it into the gated network to obtain the weights corresponding to the four experts, and then perform a weighted average of the expert prediction results based on the weights to obtain the multi-step prediction result of the disturbance term.

[0071] The fermentation state vector is:

[0072]

[0073] where t represents the current fermentation time, represents the one-hot encoding of the current fermentation phase determined according to the slope of the growth trend component, 、 、 、 represents the measurement value of each key variable at the current time, represents the recent change rate of the key variable.

[0074] The gating network (a small multi-layer perceptron) receives and outputs four expert corresponding weights , , , , and For example, in the logarithmic phase of rapid growth, the gating network may give higher weights to GRU and BCN; while in the stable phase with feeding operation, it may focus more on FCN and Attention network.

[0075] Finally, the multi-step prediction result of the disturbance term is the weighted average of all expert prediction results The method for obtaining the multi-step prediction result of the disturbance term includes:

[0076]

[0077] wherein, represents the multi-step prediction result of the disturbance term, represents the four expert corresponding weights, represents the four expert prediction results, h represents the index of a single prediction step (the "specific position" of the future time), and H represents the total number of multi-step predictions (the "length" of the prediction time window). Since h represents the "future time step", different h corresponds to different future time, therefore the prediction result of the disturbance term will be dynamically adjusted with the change of h (each h corresponds to a prediction value at a time).

[0078] S4. Based on the multi-step prediction result of the disturbance term, the trend term and the rhythm term extrapolated are combined to obtain the final prediction result.

[0079] In this embodiment, the predicted disturbance term is added to the extrapolated trend term and rhythm term to obtain the final prediction result:

[0080]

[0081] wherein, represents the final prediction result, represents the trend term obtained by extrapolating the learned parameter model for h steps, represents the rhythm term obtained by extrapolating the learned parameter model for h steps.

[0082] The training of the model uses an innovative composite loss function for end-to-end optimization:

[0083]

[0084] wherein, and The standard mean square error and mean absolute error, respectively, are used to focus on the overall prediction accuracy. The key turning point weighted loss is represented; for key process nodes such as the end of the logarithmic phase, the peak of yield, etc., a greater penalty weight is given to force the model to capture these decisive turning points more accurately. The multivariate consistency loss is represented; for example, the increase in biomass must be accompanied by the consumption of substrate, and this loss term penalizes those predictions that violate the basic material balance relationship (such as the simultaneous increase in biomass and substrate), prompting the model to learn the intrinsic coupling relationship between variables.

[0085] Example Two

[0086] In this example, the process of producing penicillin by fermenting Penicillium chrysogenum is taken as the background, aiming to predict key state variables such as biomass (Biomass, g / L) and penicillin concentration (Penicillin, g / L) in the future period (e.g. the next 12 hours).

[0087] Implementation Step 1: Data Preparation and Preprocessing

[0088] 1. Data Collection:

[0089] Collect 200 consecutive batches of historical fermentation data.

[0090] The collected variables are divided into three categories: (1) Key state variables (to be predicted): biomass, penicillin concentration. These data are usually measured offline at low frequency (e.g. every 4-6 hours). (2) Online process variables: pH, dissolved oxygen (DO, %), temperature (°C), stirring rate (rpm). These data are measured online at high frequency (e.g. every minute). (3) Control / operation variables: feed rate (L / h, including carbon and nitrogen sources), acid-base addition amount (mL / min).

[0091] 2. Data Preprocessing:

[0092] Time alignment and resampling: all high-frequency online data are down-sampled by average or median value, aligned with low-frequency offline data, and unified time resolution is 30 minutes. For the time points between offline data points, cubic spline interpolation method is used for filling to obtain smooth and continuous time series as the "true value" for model training.

[0093] Missing value processing: for a small amount of missing values caused by equipment failure or measurement error, linear interpolation or KNN (K-Nearest Neighbor) algorithm based on the previous and next time points is used for filling.

[0094] Data Standardization: To eliminate the influence of different variable dimensions and make the data fall within the range suitable for neural network training, Z-score standardization method is used for all variable sequences (except fermentation stage encoding):

[0095]

[0096] where, and represent the mean and standard deviation of the variable in all training batch data, respectively. These two values are saved for reverse standardization when predicting output to restore their original physical meaning.

[0097] Sliding Window Sample Construction: The continuous time series data of each batch is constructed into a supervised learning sample pair by sliding window method. Set the historical window length L = 48 (representing the past 24 hours of data), and the prediction window length H = 24 (representing the future 12 hours of data). Each sample contains an input tensor and an output tensor.

[0098] Input Tensor: Dimension (L, num_variables), containing the standardized sequence of all related variables in the past 24 hours.

[0099] Output Tensor: Dimension (H, num_predict_variables), containing the true value sequence of the future 12 hours of biomass and penicillin concentration.

[0100] Implementation Step 2: Model Architecture and Parameter Configuration

[0101] In this implementation, the architecture of the entire prediction model is based on Python 3.8 and PyTorch 1.12 deep learning framework.

[0102] 1. GMS Decoupling Module:

[0103] Growth Trend Component : Parameters of the generalized Logistic function are generated by a small MLP (Multi-Layer Perceptron). The MLP contains 2 hidden layers with 64 neurons each, and the activation function is ReLU. Its input is the initial conditions of the current batch (such as initial substrate concentration, inoculum) and the biomass sequence of the past 12 hours.

[0104] Metabolic Rhythm Component : The rhythm model is set to contain learnable Gaussian-like kernels. The parameter set of each kernel is directly trainable as a model, optimized through backpropagation.

[0105] 2. Heterogeneous mixture-of-experts network (MoE) module:

[0106] The input to all experts is the stochastic disturbance component S(t) within the history window L = 48 and other standardized process / control variables.

[0107] FCN expert: stacked by 3 one-dimensional convolutional layers. First layer: input channel `num_variables`, output channel 128, kernel size 8. Second layer: input channel 128, output channel 256, kernel size 5. Third layer: input channel 256, output channel 128, kernel size 3. Each layer is followed by a BatchNorm1d layer and a ReLU activation function. Finally, a global average pooling layer and a fully connected layer output the prediction result with dimension H.

[0108] GRU expert: contains a 2-layer stacked GRU network with hidden layer dimension 256. The hidden state of the last time step of the GRU is fed into a fully connected layer to decode the predicted sequence with length H.

[0109] BCN expert: embedded with a simplified Luedeking-Piret-like equation to constrain the prediction of the penicillin disturbance term:

[0110]

[0111] where a and b (representing the rate coefficients of product synthesis and degradation) are dynamically predicted by a small neural network according to the current state, represents the stochastic disturbance component of penicillin concentration, represents the stochastic disturbance component of the time series of bacterial biomass, represents the stochastic disturbance component of penicillin concentration. Then use Euler method or Runge-Kutta method for numerical integration to get the prediction of the future H steps. This network is mainly used to predict the disturbance component of penicillin.

[0112] Attention expert: uses a standard Transformer encoder structure, containing 2 attention heads (heads), model dimension 128, and feedforward network dimension 512. The input sequence is sent to the attention layer after position encoding, and finally the prediction result is output.

[0113] 3. Adaptive gating network:

[0114] Construction of fermentation state vector FCV: According to the biomass trend slope at the current time t, the fermentation process is divided into 4 stages (lag phase, logarithmic phase, stationary phase, decline phase), and one-hot encoding is performed. This encoding is spliced with the standardized values of the current time's cell biomass, glucose concentration, pH, DO, and the change rate of biomass (use approximation) to form FCV.

[0115] Gating network: an MLP containing 1 hidden layer (32 neurons, ReLU activation). The input is FCV, and the output layer uses the Softmax activation function to generate the weights corresponding to the 4 experts.

[0116] Step 3: Model training

[0117] 1. Loss function:

[0118] Composite loss function: ;

[0119] Key inflection point loss : By analyzing the training data, determine the point where the biomass growth rate begins to decrease significantly (end point of the logarithmic phase) and the point where the penicillin concentration reaches its peak. When calculating the loss, the prediction error of these time points and their adjacent areas (e.g. 3 hours before and after) is weighted 5 times.

[0120] Multivariate consistency loss : Calculate a penalty term where is the predicted change. This term penalizes the physically inconsistent situation where biomass and glucose (substrate) increase simultaneously.

[0121] 2. Training process:

[0122] Optimizer: Use AdamW optimizer, initial learning rate is 1e-4, weight decay is 1e-5. Learning rate scheduling: Use "cosine annealing" learning rate scheduler to dynamically adjust the learning rate during training. Training strategy: Train for a total of 300 epochs. Use "teacher forcing" strategy to train the GMS decoupling module and MoE module. Use 80% of the data (160 batches) as the training set and 20% of the data (40 batches) as the test set. Further divide 20% of the training set as the validation set for model selection and early stopping. Set the patience value of early stopping to 20, that is, if the validation set loss does not improve for 20 consecutive epochs, stop training. Hardware environment: Train on a server equipped with NVIDIA RTX 3090 24GB graphics card.

[0123] Implementation Step 4: Prediction and Application

[0124] Online Prediction: The trained model is deployed to the production server. Every 30 minutes, the system automatically fetches the latest process data from PCS and LIMS, and standardizes and sliding window constructs according to the preprocessing step. The latest historical window data is input into the model, and the model outputs the predicted values of cell biomass and penicillin concentration in the next 12 hours (24 time points) (after inverse standardization).

[0125] Example Three

[0126] In this embodiment, the penicillium cultivation process prediction system based on metabolic kinetics decoupling and mixed expert network includes a decoupling framework modeling module, a mixed network prediction module, a weighting module, and a result integration module.

[0127] The decoupling framework modeling module is used to construct a growth trend-metabolic rhythm-random disturbance decoupling framework, and obtain a random disturbance sequence based on the decoupling framework.

[0128] The mixed network prediction module is used to construct a mixed expert network of four parallel expert networks, and based on the mixed expert network, respectively, multi-step prediction is performed on the random disturbance sequence to obtain four expert prediction results.

[0129] The weighting module is used to construct a fermentation state vector and input it into a gating network to obtain the weights corresponding to the four experts. Based on the weights, the expert prediction results are weighted and averaged to obtain the multi-step prediction results of the disturbance term.

[0130] The result integration module obtains the final prediction results based on the multi-step prediction results of the disturbance term, combined with the extrapolated trend term and rhythm term.

[0131] The above-described embodiments only describe the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for prediction of Penicillium cultivation process based on metabolic kinetic decoupling and hybrid expert networks, characterized by, The method comprises the following steps: a growth trend-metabolic rhythm-random disturbance decoupling framework is constructed, and a random disturbance sequence is obtained based on the decoupling framework; a hybrid expert network of four parallel expert networks is constructed, and multi-step prediction is performed on the random disturbance sequence based on the hybrid expert network to obtain four expert prediction results; a fermentation state vector is constructed and input into a gating network to obtain weights corresponding to the four experts, and the expert prediction results are weighted and averaged based on the weights to obtain a disturbance term multi-step prediction result; a final prediction result is obtained based on the disturbance term multi-step prediction result in combination with an extrapolated trend term and a rhythm term.

2. The method of claim 1, wherein the method is based on a metabolic kinetic decoupling and hybrid expert network-based prediction of a Penicillium cultivation process, characterized in that, The decoupling framework is: wherein, a time series representing a key process variable, a growth trend component, a metabolic rhythm component, a random disturbance sequence.

3. The method of claim 2, wherein the method is based on metabolic kinetic decoupling and hybrid expert network for prediction of Penicillium cultivation process, characterized in that, The growth trend component is: wherein, denotes a first set of learnable parameters of the model, denotes a maximum variable value achievable by the fermentation process, denotes a coefficient related to the initial state, denotes a growth / variation rate, denotes a point in time at which half of the maximum rate is reached.

4. The method of claim 2, wherein the method is based on a metabolic kinetic decoupling and hybrid expert network-based prediction of a Penicillium cultivation process. The metabolic rhythm component is: wherein, a second set of learnable parameters representing the model, represents a preset number of rhythm components, represents an amplitude of the i-th rhythm, represents a center of a main time window of the i-th rhythm activity, represents a width of a main time window of the i-th rhythm activity, represents a frequency of the i-th rhythm, represents a phase of the i-th rhythm.

5. The method of claim 1, wherein the method is based on a metabolic kinetic decoupling and hybrid expert network-based prediction of a Penicillium cultivation process, characterized by, The hybrid expert network comprises a spatial convolution network, a gating recurrent unit network, a biochemical kinetics constraint network, and a self-attention mechanism network; The spatial convolution network uses a plurality of one-dimensional convolution kernels to extract local patterns and morphological features in the random disturbance sequence to obtain a first expert prediction result; The gating recurrent unit network is used to capture long-range temporal dependencies and memory effects in the random disturbance sequence to obtain a second expert prediction result; The biochemical kinetics constraint network inlines biochemical reaction rate equations to constrain the prediction of the random disturbance sequence, to obtain a third expert prediction result; wherein the biochemical reaction rate equation is , represents the predicted concentration of the key substrate, represents the real-time concentration of the key substrate, C represents other process variables, and U represents control variables; The self-attention mechanism network captures non-continuous and long-distance event correlations by calculating the correlation weights between each time point in the random disturbance sequence to obtain a fourth expert prediction result.

6. The method of claim 1, wherein the method is based on a metabolic kinetic decoupling and hybrid expert network-based prediction of a Penicillium cultivation process, characterized by, The fermentation state vector is: wherein, one-hot encoding of the current fermentation phase determined from the slope of the growth trend component, , , , denotes the measured value of the respective key variable at the current time instant, denotes the recent change rate of the key variable.

7. The method of claim 1, wherein the method is based on a metabolic kinetic decoupling and hybrid expert network-based prediction of a Penicillium cultivation process, characterized by, The method for obtaining the disturbance term multi-step prediction result comprises: wherein, denotes the multi-step prediction result of the disturbance term, denotes the weight corresponding to the four experts, denotes the prediction result of the four experts, h denotes the index of a single prediction step, and H denotes the total number of steps of the multi-step prediction.

8. The method of claim 7, wherein the method is based on a metabolic kinetic decoupling and hybrid expert network-based prediction of a Penicillium cultivation process, characterized by, The method for obtaining the final prediction result comprises: wherein, represents the final prediction result, represents a trend term obtained by forward-propagating the learned parametric model by h steps, represents a rhythm term obtained by forward-propagating the learned parametric model by h steps.

9. A Penicillium cultivation process prediction system based on metabolic kinetic decoupling with hybrid expert networks, said system applying the method according to any one of claims 1 to 8, characterized in that, comprises: a decoupling framework construction module, a hybrid network prediction module, a weighting module, and a result integration module; The decoupling framework construction module is used to construct a growth trend-metabolic rhythm-random disturbance decoupling framework, and a random disturbance sequence is obtained based on the decoupling framework; The hybrid network prediction module is used to construct a hybrid expert network of four parallel expert networks, and multi-step prediction is performed on the random disturbance sequence based on the hybrid expert network to obtain four expert prediction results; The weighting module is used to construct a fermentation state vector, input the fermentation state vector into a gating network to obtain weights corresponding to the four experts, and perform weighted averaging on the expert prediction results based on the weights to obtain a disturbance term multi-step prediction result; The result integration module obtains a final prediction result based on the disturbance term multi-step prediction result in combination with an extrapolated trend term and a rhythm term.

Images