A control system for microbial fermentation strain selection and breeding based on neural network model

Abstract

The present invention discloses a microbial fermentation strain selection and control system based on a neural network model, which relates to the field of digital fermentation technology. The system includes a bacterial data acquisition module, a biometric identification module, a prediction and classification module, a selection and decision-making module, an optimization and training module, an early warning and monitoring module, and a sorting and control module. The bacterial data acquisition module preprocesses strain data, the biometric identification module builds a strain feature library, the prediction and classification module predicts strain fitness and classifies strains, the selection and decision-making module selects strains and generates cultivation parameters, and through hyperparameter tuning, the early warning and monitoring module eliminates strains, and the sorting and control module generates a three-dimensional scatter plot and genetic characteristic report. This invention significantly improves the accuracy and efficiency of microbial fermentation strain selection and breeding, realizes the intelligence and automation of the selection and breeding process, reduces labor costs, enhances the stability and controllability of the fermentation process, and provides strong technical support for the innovative development of the microbial fermentation industry.

Description

Technical Field

[0001] The present invention relates to the field of digital fermentation technology, and in particular to a microbial fermentation strain breeding control system based on a neural network model. Background Art

[0002] The selection and breeding of microbial fermentation strains refers to the process of screening and optimizing microbial strains suitable for specific fermentation processes through a series of scientific methods and means. This process aims to improve the production efficiency of microorganisms, the quality and stability of products, and adapt to various conditions of industrial production. In the selection and breeding process, it is first necessary to obtain a strain library with different genetic characteristics through natural separation or artificial mutagenesis. Then, screening techniques such as growth rate, metabolite production, tolerance and other indicators are used to screen out excellent strains that meet production needs. In addition, genetic engineering technology can be used to carry out targeted transformation of strains to further improve their performance. The selected excellent strains can produce target products more efficiently during the fermentation process, while reducing the generation of by-products and improving overall production efficiency. This is of great significance to enhancing the competitiveness of the microbial fermentation industry and promoting the development of related industries. Therefore, the selection and breeding of microbial fermentation strains is a key link in microbial fermentation technology and an important driving force for the advancement of biotechnology.

[0003] In order to solve the problems of strong dependence on manual experience and low screening efficiency in the selection and breeding of microbial fermentation strains, the existing technology mainly relies on traditional screening methods and manual judgment to handle the problem. However, this method is highly dependent on the experience and intuition of researchers. It is not only time-consuming and labor-intensive, but also prone to unstable and inconsistent screening results. At the same time, due to the low screening efficiency, it is difficult to meet the needs of large-scale industrial production, which in turn leads to increased production costs and limits the widespread application of microbial fermentation technology. In view of this, a microbial fermentation strain selection and breeding control system based on a neural network model is proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide a microbial fermentation strain selection and control system based on a neural network model to solve the problems raised in the above background technology.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a microbial fermentation strain selection and breeding control system based on a neural network model, including a bacterial data acquisition module, a biometric recognition module, a prediction and classification module, a selection and breeding decision module, an optimization training module, an early warning monitoring module and a sorting and control module;

[0006] The bacterial cell data collection module collects and pre-processes microbial fermentation strain data;

[0007] The biometric identification module performs biometric identification on the pre-processed microbial fermentation strain data to construct a strain feature library;

[0008] The prediction and classification module is used to combine the bacterial strain feature library to build a fitness prediction model and a bacterial community metabolic network topology map, predict the fitness of the strain, and classify the strains according to their yield;

[0009] The breeding decision module performs a weighted combination of strain fitness and strain classification results to construct a strain breeding decision model and output the optimal strain number and culture parameters;

[0010] The optimization training module optimizes the hyperparameters of the fitness prediction model, the bacterial community metabolic network topology map, and the bacterial strain breeding decision model;

[0011] The early warning monitoring module sets the lower limit of bacterial metabolic rate warning and the threshold of abnormal fluctuation of bacterial density, thereby triggering the strain elimination warning;

[0012] The sorting and regulation module visualizes the main component distribution and genetic characteristics report of the strain through a three-dimensional scatter plot, triggers the automatic sorter to perform strain optimization and synchronously regulates the culture parameters.

[0013] A further improvement of the technical solution of the present invention is that: in the bacterial cell data acquisition module, the collection and preprocessing process of microbial fermentation strain data includes:

[0014] Acquire microbial fermentation strain data using a voltammetric biosensor probe and a near-infrared spectrometer, wherein the microbial fermentation strain data includes metabolite concentration data and bacterial density data in the microbial fermentation liquid;

[0015] A voltammetric biosensor probe is symmetrically arranged on the inner wall of the fermentation tank. A three-electrode system is used. A step voltage is applied through a potentiostat to collect redox current signals and detect the concentration data of metabolites in the microbial fermentation broth in real time. The spacing between the voltammetric biosensor probes is less than 30 cm, and the sampling frequency is set to 1 Hz. The three-electrode system includes a working electrode, a reference electrode, and a counter electrode. The metabolites include lactic acid and ethanol.

[0016] A near-infrared spectrometer was embedded in the fermentation broth circulation pipeline, with the light source and detector installed at a 45° angle. The absorbance of the bacteria was calculated based on the Lambert-Beer law, and a near-infrared spectral absorbance sequence was constructed to invert the bacterial density. The optical path was fixed at 10 mm, and a calibration white plate was placed opposite the detection window. Baseline calibration was performed daily before starting the machine.

[0017] Wavelet threshold denoising technology was used to perform a three-layer wavelet decomposition on the original redox current signal. A soft threshold function was used to set the noise threshold to the noise standard deviation multiplied by the square root logarithm of the signal length to eliminate high-frequency noise. Based on the local outlier factor algorithm, the neighborhood density ratio of the data point was calculated, and data that deviated from the density distribution was marked as abnormal. The abnormal data was then supplemented by sliding window interpolation.

[0018] The near-infrared spectral absorbance sequence was smoothed by a sliding window filter to eliminate the jitter noise of the near-infrared spectral analyzer. The data sampling interval was dynamically adjusted according to the fermentation liquid flow rate. When the flow rate was greater than 1.5 m / s, the inertia compensation algorithm was enabled. The current cell density value was obtained by adding the product of the compensation coefficient and the flow rate change to the previous cell density value.

[0019] A further improvement of the technical solution of the present invention is that: in the biometric identification module, the process of constructing the bacterial species feature library includes:

[0020] For the pre-processed metabolite concentration data, according to the fixed time window Calculate the metabolic rate V of lactate and ethanol respectively r and V y , dynamic time warping technology was used to align the metabolite concentration curves of different batches of strains and calculate the minimum path cumulative distance between the curves , get the growth curve similarity , the calculation process is as follows:

[0021] ;

[0022] ;

[0023] in, is the metabolic rate, is the concentration of metabolites at the current moment, and are the lengths of curve A and curve B, respectively. If the growth curve similarity is greater than 85%, the strain is marked as a stable growth strain. If the growth curve similarity is less than 85%, the strain is marked as a metabolically abnormal strain.

[0024] The principal component analysis of the near-infrared spectral absorbance sequence was performed to extract the first three principal components PC1, PC2 and PC3. The calculation process is as follows:

[0025] ;

[0026] Where k = 1, 2 and 3, is the weight coefficient of the kth principal component at wavelength i, is the absorbance value at the corresponding wavelength, The number of spectral wavelengths in the range of 900-1700nm is clustered according to the spatial distribution density of the principal components, and the fermentation stage is divided into the initial stage, logarithmic growth stage and decay stage, and the corresponding principal component threshold range is matched for different growth stages;

[0027] The lactate and ethanol metabolic rates, growth curve similarity, PC1, PC2, and PC3 are encoded in the format of strain number-timestamp-feature vector and integrated into the strain feature library. When new strain data is added, the Euclidean distance between it and the existing strains in the strain feature library is calculated. If the Euclidean distance is greater than 2, a new feature entry is created. After the historical strain data is retrained, the calculation is recalculated every 7 days. , update the principal component eigenvalues in the feature library.

[0028] A further improvement of the technical solution of the present invention is that: in the prediction and classification module, a fitness prediction model is constructed, and the process of predicting the fitness of the strain includes:

[0029] The lactate and ethanol metabolic rates were aligned with the growth curve similarity by time step to construct a 3D input vector covering the time series data of the entire fermentation process from 0 to 72 hours. The strain's historical fitness label was generated based on the growth stability and product yield of the strain under specific culture conditions.

[0030] A bidirectional gated recurrent unit network architecture was adopted, and the dimensions of the forward hidden state and the backward hidden state in the hidden layer were set to 64. A fitness prediction model was constructed. The input layer of the fitness prediction model received a three-dimensional feature vector, and the hidden layer controlled the information flow through the reset gate and the update gate. The output layer concatenated the bidirectional hidden states and mapped them to the strain fitness D through the fully connected layer. The loss function used the mean square error to calculate the difference between the predicted strain fitness and the strain's historical fitness label.

[0031] A further improvement of the technical solution of the present invention is that: in the prediction and classification module, the process of constructing a bacterial metabolic network topology map and classifying according to strain yield includes:

[0032] Based on historical data, classification labels were set for strains, including high-yield, medium-yield, low-yield, and metabolic abnormalities. Using a graph convolutional network architecture, the first three principal components PC1, PC2, and PC3 of each strain were input as node feature vectors into the input layer of the graph convolutional network architecture to construct a bacterial community metabolic network topology. An adjacency matrix was constructed based on the principal component spatial distance. If the Euclidean distance between the principal components of two strains was less than 1.5, an edge was defined between the nodes corresponding to the two strains.

[0033] The hidden layer of the bacterial community metabolic network topology map uses a two-layer graph convolution operation, with each layer outputting 16 dimensions. Node feature propagation is achieved by adding a self-connected adjacency matrix and degree matrix normalization. The output layer aggregates node features through global maximum pooling to generate a 3D probability vector. ;

[0034] Set the upper limit of the strain yield probability threshold to 0.7 and the lower limit of the strain yield threshold to 0.5. If the output 3D probability vector If it is greater than 0.7, the strain is marked as high-yielding. and If the three are the highest, the strain is marked as medium-yielding. and If it is the highest of the three, the strain is marked as low-yielding;

[0035] like If the difference between the two probabilities is less than 0.1, the category is determined according to the proximity of the principal component space distance to the historical high-yield cluster. If the three-category probability of the strain is less than 0.5, it is marked as a candidate strain for metabolic abnormality. If the Euclidean distance of the principal component feature of the strain is greater than 3, it is temporarily stored as an unclassified strain, and the upper and lower limits of the strain yield probability threshold are dynamically adjusted based on the proportion of high-yield strains in historical data.

[0036] A further improvement of the technical solution of the present invention is that: in the breeding decision module, the process of constructing a strain breeding decision model and outputting the optimal strain number and culture parameters includes:

[0037] According to the strain classification results, the weight coefficients of fitness and classification results are dynamically adjusted to construct a strain selection decision model. If the strain is classified as high-yielding, the strain classification results are given priority. If the strain is classified as medium-yielding or low-yielding, fitness is emphasized. Based on the weight coefficients of fitness and classification results, the classification gain Z and comprehensive score F are calculated. The calculation process is as follows:

[0038] ;

[0039] ;

[0040] in, is the weight coefficient of fitness, is the weight coefficient of strain classification results;

[0041] The upper and lower thresholds of the comprehensive score are set at 0.8 and 0.6 respectively. If the comprehensive score is ≥0.8, the strain enters the preferred queue. If the comprehensive score is 0.6≤comprehensive score<0.8, the strain enters the observation queue. If the comprehensive score is <0.6, the strain is marked as an elimination candidate. If the strain is marked as a metabolic abnormality candidate strain, it is pushed to the early warning monitoring module.

[0042] Introduce and record the pH and feed rate of the strain under optimal culture conditions The historical parameter mapping table is used to match the culture parameters of the strain with the same classification result and the highest fitness from the historical parameter mapping table, and the pH and feeding rate are updated. The calculation process is as follows:

[0043] ;

[0044] ;

[0045] in, and Represents the baseline pH and baseline rate, which are their historical optimal values respectively.

[0046] A further improvement of the technical solution of the present invention is that in the optimization training module, the process of optimizing the hyperparameters of the fitness prediction model, the bacterial community metabolic network topology map, and the bacterial strain breeding decision model includes:

[0047] The fitness prediction model dynamically adjusts its learning rate through the cosine annealing strategy. The initial learning rate is set to 0.001, the learning rate threshold is set to 0.0001, and the learning rate decays according to the cosine function as the training rounds increase. An early stopping mechanism is set. If the validation set loss of the fitness prediction model does not decrease for 10 consecutive rounds, the training is terminated to prevent overfitting.

[0048] The principal component features of the new strain are added to the topology graph, and the adjacency matrix is recalculated. If the maximum Euclidean distance between the new node and the existing node exceeds 2, the network structure is expanded. When the new data reaches 100 in each batch, the parameters of the graph convolutional network are updated using an incremental learning algorithm, and the cross-entropy loss function is used to optimize the topology of the bacterial community metabolic network.

[0049] When the mean square error between the actual yield of each batch of preferred strains and the comprehensive score predicted by the strain selection decision model exceeds 0.1, the fitness weight coefficient is updated according to the gradient descent method, the actual culture parameters and yield data of the preferred strains are added to the historical parameter mapping table, and the benchmark pH value and benchmark rate value are updated every 7 days.

[0050] A further improvement of the technical solution of the present invention is that: in the early warning monitoring module, the process of setting the early warning lower limit of the bacterial metabolic rate and the abnormal fluctuation threshold of the bacterial density, thereby triggering the strain elimination early warning, includes:

[0051] Based on the historical metabolic rate distribution of high-yield strains, the 5th percentile was selected as the lower limit of the bacterial metabolic rate warning. After each batch of high-yield strain data was added, the lower limit of the bacterial metabolic rate warning was recalculated according to the sliding window. If the current value deviated from the original threshold by more than 10%, it was updated to the bacterial metabolic rate threshold.

[0052] The percentage of the standard deviation of the cell density within the time window to the mean was calculated to obtain the cell density variation coefficient. The cell density abnormal fluctuation thresholds were set at 25%, 20%, and 10% according to the initial fermentation stage, logarithmic growth stage, and decay stage, respectively. If PC1>0.7 and PC2<0.3, the bacterial community was judged to be in the initial fermentation stage. If PC2>0.5 and PC3<0.2, the bacterial community was judged to be in the logarithmic growth stage. If PC3>0.6 and PC1<0.4, the bacterial community was judged to be in the decay stage. When the cell density variation coefficient exceeded the cell density abnormal fluctuation threshold of the corresponding fermentation stage, it was considered an abnormal fluctuation.

[0053] For strains in the early stages of fermentation, if the initial metabolic rate is lower than the current bacterial community metabolic rate warning lower limit and the bacterial density variation coefficient exceeds 25%, the strain will be marked as a growth-delayed strain and pushed to the manual review queue. If the strain does not enter the logarithmic growth phase for more than 72 hours in the early stages of fermentation, a metabolic stagnation warning will be triggered, and fermentation will be terminated and samples will be collected for analysis.

[0054] Abnormal strains are marked according to the lower warning limit of bacterial metabolic rate and the abnormal fluctuation threshold of bacterial density, and the principal component space deviation is verified twice. If the strain is confirmed to be abnormal, the culture is terminated and the culture parameters are adjusted, and the bacterial species feature library is updated synchronously.

[0055] A further improvement of the technical solution of the present invention is that: in the sorting and control module, the process of visualizing the main component distribution of the strains and marking the high-yield bacterial clusters through a three-dimensional scatter plot and generating a genetic characteristic comparison report includes:

[0056] The PC1, PC2, and PC3 of the strains were used as three-dimensional coordinate axes to construct a scatter plot. Each data point represented a strain. High-yield bacterial clusters were identified using a density clustering algorithm. The coordinates of the convex hull vertices were calculated and cluster boundaries were generated. Different strain classification labels were distinguished by different color codes, and the spatial distribution of high-yield clusters was dynamically annotated. Red indicates high-yield strains, yellow indicates medium-yield strains, blue indicates low-yield strains, and gray indicates strains with abnormal metabolism.

[0057] The differences in metabolic pathway gene expression between high-yield bacterial clusters and strains in the entire library were statistically analyzed. The metabolic pathway enrichment was calculated using the hypergeometric test. Significant pathways with metabolic pathway enrichment lower than 0.05 were screened. The mean and standard deviation of key enzyme activities were statistically analyzed. Highly expressed enzymes with activities greater than 2 times the standard deviation of the mean were marked. A comprehensive report containing a three-dimensional scatter plot, a pathway enrichment table, and an enzyme activity comparison bar chart was generated.

[0058] A further improvement of the technical solution of the present invention is that: in the sorting control module, the process of driving the automatic sorter and the fermentation parameter control system to perform strain selection and pH and feed rate adjustment includes:

[0059] The strains in the preferred queue are sorted in descending order of comprehensive scores, with the top 10% of strains being prioritized. If multiple strains exist in the same scoring interval, they are sorted based on the proximity of the principal component spatial distance to the historically high-yield cluster. The sorter locates the target strain based on the strain number using microfluidic chip sorting technology. After sorting, the strain is automatically inoculated into a new fermenter, and the sorting timestamp and culture parameters are simultaneously recorded.

[0060] If the strain is classified as high-yielding, the pH is adjusted normally. If the strain is in the logarithmic growth phase, an additional pH compensation of 0.05 is added. The basic feed rate is matched according to the historical parameter mapping table. If the key enzyme activity is detected to be greater than 2 standard deviations of the mean, the feed rate is increased by an additional 10%;

[0061] Terminate the cultivation of candidate strains with metabolic abnormalities in the fermentation tank, turn off the feed pump and start the tank cleaning program, record the main component characteristics and operation logs of the abnormal strains, push them to the strain feature library and mark them as eliminated data, and after each batch is sorted, update the fitness weight coefficient according to the gradient descent method based on the difference between the actual yield and the predicted score, and remove invalid parameter entries with actual yields 20% lower than the same category average from the historical parameter mapping table every 7 days.

[0062] Due to the adoption of the above technical solution, the present invention has the following technical advancements compared to the prior art:

[0063] 1. The present invention provides a microbial fermentation strain selection and control system based on a neural network model, which significantly improves the accuracy and efficiency of strain selection and breeding. Through intelligent processing and analysis, it greatly shortens the selection and breeding cycle and reduces labor costs.

[0064] 2. The present invention provides a microbial fermentation strain breeding control system based on a neural network model, which realizes the comprehensive automation and real-time monitoring of the breeding process, effectively avoids human errors, and enhances the stability and reliability of the breeding results.

[0065] 3. The present invention provides a microbial fermentation strain selection and control system based on a neural network model. By optimizing the training module and continuously iterating and upgrading the model parameters, it ensures that the system always remains in the optimal state, providing a solid foundation for continuous innovation and technological upgrading in the microbial fermentation industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0067] Figure 1 A block diagram of the present invention. DETAILED DESCRIPTION

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0069] Examples, such as Figure 1 As shown, the present invention provides a microbial fermentation strain selection and breeding control system based on a neural network model, including a bacterial data acquisition module, a biological feature recognition module, a prediction and classification module, a selection and breeding decision module, an optimization training module, an early warning monitoring module, and a sorting and control module;

[0070] The bacterial data acquisition module collects and preprocesses microbial fermentation strain data, and obtains microbial fermentation strain data through voltammetry biosensor probes and near-infrared spectrometer. The microbial fermentation strain data includes metabolite concentration data and bacterial density data in the microbial fermentation broth. The voltammetry biosensor probes are symmetrically arranged on the inner wall of the fermentation tank. A three-electrode system is used. A step voltage is applied through a constant potential meter to collect redox current signals and detect the metabolite concentration data in the microbial fermentation broth in real time. The spacing between the voltammetry biosensor probes is less than 30 cm, and the sampling frequency is set to 1 Hz. The three-electrode system includes a working electrode, a reference electrode, and a counter electrode. Metabolites include lactic acid and ethanol. A near-infrared spectrometer is embedded in the fermentation broth circulation pipeline. The light source and the detector are installed at a 45° angle. The bacterial absorbance value is calculated based on the Lambert-Beer law. A near-infrared spectral absorbance sequence was constructed to invert bacterial density. The optical path was fixed at 10 mm, a calibration white plate was placed opposite the detection window, and baseline calibration was performed daily before starting the machine. Wavelet threshold denoising technology was used to perform a three-layer wavelet decomposition of the original redox current signal. A soft threshold function was used to set the noise threshold to the noise standard deviation multiplied by the square root logarithm of the signal length to eliminate high-frequency noise. Based on the local outlier factor algorithm, the neighborhood density ratio of the data point was calculated, and data that deviated from the density distribution was marked as abnormal. Sliding window interpolation was used to supplement abnormal data. Sliding window smoothing filtering was performed on the near-infrared spectral absorbance sequence to eliminate jitter noise of the near-infrared spectrometer analyzer. The data sampling interval was dynamically adjusted according to the fermentation liquid flow rate. When the flow rate was greater than 1.5 m / s, the inertia compensation algorithm was enabled. The current bacterial density value was the product of the previous bacterial density value plus the compensation coefficient and the flow rate change.

[0071] The biometric identification module performs biometric identification on the pre-processed microbial fermentation strain data, builds a strain feature library, and performs biometric identification on the pre-processed metabolite concentration data according to a fixed time window. Calculate the metabolic rate V of lactate and ethanol respectively r and V y , dynamic time warping technology was used to align the metabolite concentration curves of different batches of strains and calculate the minimum path cumulative distance between the curves , get the growth curve similarity , the calculation process is as follows:

[0072] ;

[0073] ;

[0074] in, is the metabolic rate, is the concentration of metabolites at the current moment, and are the lengths of curve A and curve B, respectively. If the growth curve similarity is greater than 85%, it is marked as a stable growth strain. If the growth curve similarity is less than 85%, it is marked as a metabolic abnormal strain. Principal component analysis is performed on the near-infrared spectral absorbance sequence to extract the first three principal components PC1, PC2, and PC3. The calculation process is as follows:

[0075] ;

[0076] Where k = 1, 2 and 3, is the weight coefficient of the kth principal component at wavelength i, is the absorbance value at the corresponding wavelength, The number of spectral wavelengths in the range of 900-1700nm is clustered according to the spatial distribution density of the principal components, and the fermentation stage is divided into the initial stage, logarithmic growth stage and decay stage. The corresponding principal component threshold range is matched for different growth stages. The lactic acid and ethanol metabolic rates, growth curve similarity, PC1, PC2 and PC3 are encoded in the format of strain number-timestamp-feature vector and integrated into the strain feature library. When new strain data is added, its Euclidean distance with the existing strains in the strain feature library is calculated. If the Euclidean distance is greater than 2, a new feature entry is created. After the historical strain data is retrained, the Euclidean distance is recalculated every 7 days. , update the principal component eigenvalues in the feature library;

[0077] The prediction and classification module is used to combine the strain feature library, build a fitness prediction model and a bacterial metabolic network topology map, predict the strain fitness, and classify it according to the strain yield. The lactic acid and ethanol metabolic rates are aligned with the growth curve similarity by time step, and a three-dimensional input vector is constructed to cover the time series data of the entire fermentation process from 0 to 72 hours. The strain historical fitness label is generated based on the growth stability and product yield of the strain under specific culture conditions. A bidirectional gated recurrent unit network architecture is used, and the forward hidden state and backward hidden state dimensions in the hidden layer are set to 64 to construct a fitness prediction model. The input layer of the fitness prediction model receives a three-dimensional feature vector, and the hidden layer controls the information flow by resetting the gate and updating the gate. The output layer splices the bidirectional hidden state and maps it to the strain fitness D through the fully connected layer. The loss function uses the mean square error. The difference between the predicted strain fitness and the historical fitness label of the strain is calculated, and a classification label is set for the strain based on the historical data. The strain classification label includes high yield, medium yield, low yield and metabolic abnormalities. The graph convolutional network architecture is used to input the first three principal components PC1, PC2 and PC3 of each strain as the node feature vector into the input layer of the graph convolutional network architecture to construct a microbial metabolic network topology map. The adjacency matrix is constructed based on the principal component space distance. If the Euclidean distance of the principal components of the two strains is less than 1.5, it is defined that there is an edge between the nodes corresponding to the two strains. The hidden layer of the microbial metabolic network topology map adopts a two-layer graph convolution operation, and each layer outputs 16 dimensions. The node feature propagation is realized by adding a self-connected adjacency matrix and degree matrix normalization. The output layer aggregates the node features through global maximum pooling to generate a 3D probability vector , set the strain yield probability threshold upper limit to 0.7, the strain yield threshold lower limit to 0.5, if the output 3D probability vector If it is greater than 0.7, the strain is marked as high-yielding. and If the three are the highest, the strain is marked as medium-yielding. and If the three are the highest, the strain is marked as low-yielding. If the probability difference between the two is less than 0.1, the category is determined based on the proximity of the principal component space distance to the historical high-yield cluster. If the three-category probability of the strain is less than 0.5, it is marked as a candidate strain for metabolic abnormality. If the Euclidean distance of the principal component feature of the strain is greater than 3, it is temporarily stored as an unclassified strain, and the upper and lower limits of the strain yield probability threshold are dynamically adjusted based on the proportion of high-yield strains in historical data.

[0078] The breeding decision module performs a weighted combination of strain fitness and strain classification results to construct a strain breeding decision model, output the optimal strain number and culture parameters, and dynamically adjust the weight coefficients of fitness and classification results according to the strain classification results to build a strain breeding decision model. If the strain is classified as high-yield, the strain classification results are given priority. If the strain is classified as medium-yield or low-yield, fitness is emphasized. Based on the weight coefficients of fitness and classification results, the classification gain Z and comprehensive score F are calculated. The calculation process is as follows:

[0079] ;

[0080] ;

[0081] in, is the weight coefficient of fitness, The weight coefficient of the strain classification result is set, and the upper and lower limits of the comprehensive score threshold are set to 0.8 and 0.6 respectively. If the comprehensive score is ≥0.8, the strain enters the preferred queue. If 0.6≤comprehensive score<0.8, the strain enters the observation queue. If the comprehensive score is <0.6, the strain is marked as an elimination candidate. If the strain is marked as a metabolic abnormality candidate strain, it is pushed to the early warning monitoring module, and the pH and feed rate of the strain under the optimal culture conditions are recorded. The historical parameter mapping table is used to match the culture parameters of the strain with the same classification result and the highest fitness from the historical parameter mapping table, and the pH and feeding rate are updated. The calculation process is as follows:

[0082] ;

[0083] ;

[0084] in, and represents the baseline pH and baseline rate, which are their historical optimal values respectively;

[0085] The training module was optimized, including the hyperparameters of the fitness prediction model, the microbial metabolic network topology, and the strain selection decision model. The fitness prediction model dynamically adjusted its learning rate using a cosine annealing strategy, with an initial learning rate of 0.001 and a learning rate threshold of 0.0001. The learning rate decayed according to a cosine function with increasing training rounds. An early stopping mechanism was set. If the validation set loss of the fitness prediction model did not decrease for 10 consecutive rounds, training was terminated to prevent overfitting. The principal component features of the new strain were added to the topology, and the adjacency matrix was recalculated. If the maximum Euclidean distance between the new node and the existing node exceeded 2, the network structure was expanded. When the new data reached 100 per batch, the graph convolutional network parameters were updated using an incremental learning algorithm. The cross-entropy loss function was used to optimize the microbial metabolic network topology. When the mean square error between the actual yield of each batch of selected strains and the comprehensive score predicted by the strain selection decision model exceeded 0.1, the fitness weight coefficient was updated using the gradient descent method. The actual culture parameters and yield data of the selected strains were added to the historical parameter mapping table, and the baseline pH value and baseline rate values were updated every 7 days.

[0086] The early warning monitoring module sets a lower limit for bacterial metabolic rate and a threshold for abnormal cell density fluctuations, thereby triggering a strain elimination warning. Based on the metabolic rate distribution of historical high-yield strains, the 5th percentile is selected as the lower limit for bacterial metabolic rate warning. After each batch of high-yield strain data is added, the lower limit for bacterial metabolic rate warning is recalculated using a sliding window. If the current value deviates from the original threshold by more than 10%, it is updated to the bacterial metabolic rate threshold. The percentage of the standard deviation of cell density within the time window to the mean is calculated to obtain the cell density variation coefficient. The abnormal cell density fluctuation thresholds are set at 25%, 20%, and 10% based on the initial fermentation stage, logarithmic growth stage, and decay stage, respectively. If PC1 is greater than 0.7 and PC2 is less than 0.3, the bacterial population is considered to be in the initial fermentation stage. If PC2 is greater than 0.5 and PC3 is less than 0. 2, the bacterial community is judged to be in the logarithmic growth phase. If PC3>0.6 and PC1<0.4, the bacterial community is judged to be in the decline phase. When the bacterial density variation coefficient exceeds the bacterial density abnormal fluctuation threshold of the corresponding fermentation stage, it is regarded as an abnormal fluctuation. For the strain in the early stage of fermentation, if the initial metabolic rate is lower than the current bacterial community metabolic rate warning lower limit and the bacterial density variation coefficient exceeds 25%, it is marked as a growth-delayed strain and pushed to the manual review queue. If the strain does not enter the logarithmic growth phase for more than 72 hours in the early stage of fermentation, a metabolic stagnation warning is triggered, the fermentation is terminated, and sampling and analysis are carried out. The abnormal strain is marked according to the bacterial community metabolic rate warning lower limit and the bacterial density abnormal fluctuation threshold, and the principal component space deviation is verified twice. If the strain is confirmed to be abnormal, the culture is terminated and the culture parameters are adjusted, and the strain feature library is updated synchronously.

[0087] The sorting and regulation module visualizes the principal component distribution and genetic characteristics report of the strain through a three-dimensional scatter plot, triggers the automatic sorter to perform strain optimization and synchronously regulates the culture parameters, uses the PC1, PC2 and PC3 of the strain as the three-dimensional coordinate axis, constructs a scatter plot, and each data point represents a strain. The high-yield bacterial cluster is identified by the density clustering algorithm, the convex hull vertex coordinates are calculated and the cluster boundary is generated. Different strain classification labels are distinguished by different color codes, and the spatial distribution of high-yield clusters is dynamically annotated. Among them, red represents high-yield strains, yellow represents medium-yield strains, blue represents low-yield strains, and gray represents metabolic abnormalities. The difference in metabolic pathway gene expression between high-yield bacterial clusters and the entire library strains is statistically analyzed, and the hypergeometric test is used to calculate the metabolic pathway enrichment. Significant pathways with metabolic pathway enrichment less than 0.05 are screened, the mean and standard deviation of key enzyme activities are statistically analyzed, and highly expressed enzymes with activities greater than 2 times the standard deviation of the mean are marked. A comprehensive report containing a three-dimensional scatter plot, a pathway enrichment table, and an enzyme activity comparison bar chart is generated. The preferred cohort strains are ranked in descending order according to the comprehensive score. First, the top 10% of strains were sorted. If multiple strains were found in the same scoring interval, they were sorted based on their proximity to historically high-yield clusters using principal component spatial distance. The sorter located the target strain based on the strain number using microfluidic chip sorting technology. After sorting, the strain was automatically inoculated into a new fermenter, and the sorting timestamp and culture parameters were simultaneously recorded. If the strain was classified as high-yielding, the pH was adjusted normally. If the strain was in the logarithmic growth phase, an additional pH compensation of 0.05 was added. The basal feed rate was matched according to the historical parameter mapping table. If the key enzyme activity was detected to be greater than 2 standard deviations above the mean, the feed rate was increased by an additional 10%. Cultivation of candidate strains with metabolic abnormalities in the fermenter was terminated, the feed pump was shut down, and the tank cleaning procedure was initiated. The principal component characteristics and operation logs of the abnormal strains were recorded, pushed to the strain feature library, and marked as eliminated data. After each batch of sorting, the fitness weight coefficient was updated using the gradient descent method based on the difference between the actual yield and the predicted score. Invalid parameter entries with actual yields less than 20% of the category mean were removed from the historical parameter mapping table every 7 days.

[0088] First, through the bacterial data acquisition module, the system will automatically collect and pre-process the bacterial strain data in the microbial fermentation process to ensure the accuracy and availability of the data. Subsequently, the biometric identification module will conduct an in-depth analysis of these pre-processed data, identify the biological characteristics of the bacterial strain, and build a comprehensive bacterial strain feature library. Next, the prediction and classification module will use the bacterial strain feature library to build a fitness prediction model and a bacterial community metabolic network topology map to predict the fitness and yield of the strain, and classify the strain according to the yield. The breeding decision module will comprehensively consider the fitness and classification results of the strain, and through a weighted combination method, construct a bacterial strain breeding decision model, and finally output the number of the optimal strain and its corresponding culture parameters. During use, the optimization training module will continuously optimize the hyperparameters of the fitness prediction model, the microbial metabolic network topology and the strain breeding decision model to improve the system's prediction accuracy and breeding efficiency. At the same time, the early warning monitoring module will monitor the metabolic rate and bacterial density of the microbial community in real time. Once the set warning lower limit is exceeded or abnormal fluctuations occur, the system will immediately trigger the strain elimination warning to ensure the stability and safety of the breeding process. Finally, the sorting and control module will visualize the principal component distribution and genetic characteristics report of the strain through a three-dimensional scatter plot to help users intuitively understand the genetic background and principal component distribution of the strain, thereby triggering the automatic sorter to execute strain optimization and synchronously adjust the culture parameters to achieve automation and intelligence of strain breeding.

[0089] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A microbial fermentation strain selection and control system based on a neural network model, characterized by: It includes bacterial data collection module, biological feature recognition module, prediction and classification module, breeding decision module, optimization training module, early warning monitoring module and sorting and control module; The bacterial cell data collection module collects and pre-processes microbial fermentation strain data; The biometric identification module performs biometric identification on the pre-processed microbial fermentation strain data to construct a strain feature library; The prediction and classification module is used to combine the bacterial strain feature library to build a fitness prediction model and a bacterial community metabolic network topology map, predict the fitness of the strain, and classify the strains according to their yield; The breeding decision module performs a weighted combination of strain fitness and strain classification results to construct a strain breeding decision model and output the optimal strain number and culture parameters; The optimization training module optimizes the hyperparameters of the fitness prediction model, the bacterial community metabolic network topology map, and the bacterial strain breeding decision model; The early warning monitoring module sets the lower limit of bacterial metabolic rate warning and the threshold of abnormal fluctuation of bacterial density, thereby triggering the strain elimination warning; The sorting and regulation module visualizes the main component distribution and genetic characteristics report of the strain through a three-dimensional scatter plot, triggers the automatic sorter to perform strain optimization and synchronously regulates the culture parameters.

2. The microbial fermentation strain selection and breeding control system based on a neural network model according to claim 1, characterized in that: In the bacterial cell data collection module, the collection and preprocessing process of microbial fermentation strain data includes: Acquire microbial fermentation strain data using a voltammetric biosensor probe and a near-infrared spectrometer, wherein the microbial fermentation strain data includes metabolite concentration data and bacterial density data in the microbial fermentation liquid; A voltammetric biosensor probe is symmetrically placed on the inner wall of the fermentation tank. Using a three-electrode system, a step voltage is applied through a potentiostat to collect redox current signals and detect the concentration of metabolites in the microbial fermentation broth in real time. The metabolites include lactic acid and ethanol. A near-infrared spectrometer was embedded in the fermentation liquid circulation pipeline. The light source and detector were installed at a 45-degree angle. The absorbance value of the bacteria was calculated based on the Lambert-Beer law, and a near-infrared spectral absorbance sequence was constructed to invert the bacterial density. Wavelet threshold denoising technology was used to perform a three-layer wavelet decomposition on the original redox current signal, and a sliding window smoothing filter was performed on the near-infrared spectral absorbance sequence. The data sampling interval was dynamically adjusted according to the fermentation liquid flow rate. When the flow rate was greater than 1.5 m / s, the inertia compensation algorithm was enabled, and the current cell density value was obtained by adding the product of the compensation coefficient and the flow rate change to the previous cell density value.

3. The microbial fermentation strain selection and breeding control system based on a neural network model according to claim 2, characterized in that: In the biometric identification module, the process of constructing a bacterial strain feature library includes: For the pre-processed metabolite concentration data, according to the fixed time window Calculate the metabolic rates of lactate and ethanol separately r and V y , dynamic time warping technology was used to align the metabolite concentration curves of different batches of strains and calculate the minimum path cumulative distance between the curves , get the growth curve similarity If the growth curve similarity is greater than 85%, it is marked as a stable growth strain; if the growth curve similarity is less than 85%, it is marked as a metabolic abnormal strain; Principal component analysis was performed on the near-infrared spectral absorbance sequence to extract the first three principal components, PC1, PC2, and PC3. Clustering was performed based on the spatial distribution density of the principal components to divide the fermentation stage into the initial stage, logarithmic growth stage, and decay stage. The corresponding principal component threshold ranges were matched for different growth stages. The lactate and ethanol metabolic rates, growth curve similarity, PC1, PC2, and PC3 are encoded in the format of strain number-timestamp-feature vector and integrated into the strain feature library. When new strain data is added, the Euclidean distance between it and the existing strains in the strain feature library is calculated. If the Euclidean distance is greater than 2, a new feature entry is created. After the historical strain data is retrained, the calculation is recalculated every 7 days. , update the principal component eigenvalues in the feature library.

4. The microbial fermentation strain selection and breeding control system based on a neural network model according to claim 3, characterized in that: In the prediction and classification module, a fitness prediction model is constructed, and the process of predicting the fitness of the strain includes: The lactate and ethanol metabolic rates were aligned with the growth curve similarity by time step to construct a 3D input vector covering the time series data of the entire fermentation process from 0 to 72 hours. The strain's historical fitness label was generated based on the growth stability and product yield of the strain under specific culture conditions. A bidirectional gated recurrent unit network architecture was adopted, and the dimensions of the forward hidden state and the backward hidden state in the hidden layer were set to 64. A fitness prediction model was constructed. The input layer of the fitness prediction model received a three-dimensional feature vector, and the hidden layer controlled the information flow through the reset gate and the update gate. The output layer concatenated the bidirectional hidden states and mapped them to the strain fitness D through the fully connected layer. The loss function used the mean square error to calculate the difference between the predicted strain fitness and the strain's historical fitness label.

5. The microbial fermentation strain selection and breeding control system based on a neural network model according to claim 4, characterized in that: In the prediction and classification module, the process of constructing a bacterial metabolic network topology map and classifying the strains according to their yields includes: Based on historical data, classification labels were set for strains, including high-yield, medium-yield, low-yield, and metabolic abnormalities. Using a graph convolutional network architecture, the first three principal components PC1, PC2, and PC3 of each strain were input as node feature vectors into the input layer of the graph convolutional network architecture to construct a bacterial community metabolic network topology. An adjacency matrix was constructed based on the principal component spatial distance. If the Euclidean distance between the principal components of two strains was less than 1.5, an edge was defined between the nodes corresponding to the two strains. The hidden layer of the bacterial community metabolic network topology map uses a two-layer graph convolution operation, with each layer outputting 16 dimensions. Node feature propagation is achieved by adding a self-connected adjacency matrix and degree matrix normalization. The output layer aggregates node features through global maximum pooling to generate a 3D probability vector. ; Set the upper and lower limits of the strain yield probability threshold to 0.7 and 0.5 respectively. If the output 3D probability vector If it is greater than 0.7, the strain is marked as high-yielding. and If the three are the highest, the strain is marked as medium-yielding. and If it is the highest of the three, the strain is marked as low-yielding; like If the difference between the two probabilities is less than 0.1, the category is determined according to the proximity of the principal component space distance to the historical high-yield cluster. If the three-category probability of the strain is less than 0.5, it is marked as a candidate strain for metabolic abnormality. If the Euclidean distance of the principal component feature of the strain is greater than 3, it is temporarily stored as an unclassified strain, and the upper and lower limits of the strain yield probability threshold are dynamically adjusted based on the proportion of high-yield strains in historical data.

6. The microbial fermentation strain selection and control system based on a neural network model according to claim 5, characterized in that: In the breeding decision module, the process of constructing a strain breeding decision model and outputting the optimal strain number and culture parameters includes: The weight coefficients of fitness and classification results are dynamically adjusted according to the strain classification results to construct a strain selection decision model. If the strain is classified as high-yielding, the strain classification results are given priority. If the strain is classified as medium-yielding or low-yielding, fitness is emphasized. Based on the weight coefficients of fitness and classification results, the classification gain Z and comprehensive score F are calculated. The upper and lower thresholds of the comprehensive score are set at 0.8 and 0.6 respectively. If the comprehensive score is ≥0.8, the strain enters the preferred queue. If the comprehensive score is 0.6≤comprehensive score<0.8, the strain enters the observation queue. If the comprehensive score is <0.6, the strain is marked as an elimination candidate. If the strain is marked as a metabolic abnormality candidate strain, it is pushed to the early warning monitoring module. Introduce and record the pH and feed rate of the strain under optimal culture conditions The historical parameter mapping table is used to match the culture parameters of the strain with the same classification result and the highest fitness from the historical parameter mapping table, and the pH and feeding rate are updated.

7. The microbial fermentation strain selection and breeding control system based on a neural network model according to claim 6, characterized in that: In the optimization training module, the process of optimizing the hyperparameters of the fitness prediction model, the bacterial community metabolic network topology map, and the bacterial strain selection decision model includes: The fitness prediction model dynamically adjusts its learning rate through the cosine annealing strategy. The initial learning rate is set to 0.001, the learning rate threshold is set to 0.0001, and the learning rate decays according to the cosine function as the training rounds increase. An early stopping mechanism is set. If the validation set loss of the fitness prediction model does not decrease for 10 consecutive rounds, the training is terminated to prevent overfitting. The principal component features of the new strain are added to the topology graph, and the adjacency matrix is recalculated. If the maximum Euclidean distance between the new node and the existing node exceeds 2, the network structure is expanded. When the new data reaches 100 in each batch, the parameters of the graph convolutional network are updated using an incremental learning algorithm, and the cross-entropy loss function is used to optimize the topology of the bacterial community metabolic network. When the mean square error between the actual yield of each batch of preferred strains and the comprehensive score predicted by the strain selection decision model exceeds 0.1, the fitness weight coefficient is updated according to the gradient descent method, the actual culture parameters and yield data of the preferred strains are added to the historical parameter mapping table, and the benchmark pH value and benchmark rate value are updated every 7 days.

8. The microbial fermentation strain selection and breeding control system based on a neural network model according to claim 7, characterized in that: In the early warning monitoring module, the process of setting the early warning lower limit of bacterial metabolic rate and the threshold of abnormal fluctuation of bacterial density to trigger the strain elimination early warning includes: Based on the historical metabolic rate distribution of high-yield strains, the 5th percentile was selected as the lower limit of the bacterial metabolic rate warning. After each batch of high-yield strain data was added, the lower limit of the bacterial metabolic rate warning was recalculated according to the sliding window. If the current value deviated from the original threshold by more than 10%, it was updated to the bacterial metabolic rate threshold. The percentage of the standard deviation of the cell density within the time window to the mean was calculated to obtain the cell density variation coefficient. The cell density abnormal fluctuation thresholds were set at 25%, 20%, and 10% according to the initial fermentation stage, logarithmic growth stage, and decay stage, respectively. If PC1>0.7 and PC2<0.3, the bacterial community was judged to be in the initial fermentation stage. If PC2>0.5 and PC3<0.2, the bacterial community was judged to be in the logarithmic growth stage. If PC3>0.6 and PC1<0.4, the bacterial community was judged to be in the decay stage. When the cell density variation coefficient exceeded the cell density abnormal fluctuation threshold of the corresponding fermentation stage, it was considered an abnormal fluctuation. For strains in the early stages of fermentation, if the initial metabolic rate is lower than the current bacterial community metabolic rate warning lower limit and the bacterial density variation coefficient exceeds 25%, the strain will be marked as a growth-delayed strain and pushed to the manual review queue. If the strain does not enter the logarithmic growth phase for more than 72 hours in the early stages of fermentation, a metabolic stagnation warning will be triggered, and fermentation will be terminated and samples will be collected for analysis. Abnormal strains are marked according to the lower warning limit of bacterial metabolic rate and the abnormal fluctuation threshold of bacterial density, and the principal component space deviation is verified twice. If the strain is confirmed to be abnormal, the culture is terminated and the culture parameters are adjusted, and the bacterial species feature library is updated synchronously.

9. The microbial fermentation strain selection and breeding control system based on a neural network model according to claim 8, characterized in that: In the sorting and control module, the process of visualizing the main component distribution of strains and marking high-yield bacterial clusters through a three-dimensional scatter plot and generating a genetic characteristics comparison report includes: The PC1, PC2, and PC3 of the strains were used as three-dimensional coordinate axes to construct a scatter plot. Each data point represented a strain. High-yield bacterial clusters were identified using a density clustering algorithm. The coordinates of the convex hull vertices were calculated and cluster boundaries were generated. Different strain classification labels were distinguished by different color codes, and the spatial distribution of high-yield clusters was dynamically annotated. Red indicates high-yield strains, yellow indicates medium-yield strains, blue indicates low-yield strains, and gray indicates strains with abnormal metabolism. The differences in metabolic pathway gene expression between high-yield bacterial clusters and strains in the entire library were statistically analyzed. The metabolic pathway enrichment was calculated using the hypergeometric test. Significant pathways with metabolic pathway enrichment lower than 0.05 were screened. The mean and standard deviation of key enzyme activities were statistically analyzed. Highly expressed enzymes with activities greater than 2 times the standard deviation of the mean were marked. A comprehensive report containing a three-dimensional scatter plot, a pathway enrichment table, and an enzyme activity comparison bar chart was generated.

10. The microbial fermentation strain selection and breeding control system based on a neural network model according to claim 9, characterized in that: In the sorting and control module, the process of driving the automatic sorter and the fermentation parameter control system to perform strain optimization and pH and feed rate adjustment includes: The strains in the preferred queue are sorted in descending order of comprehensive scores, with the top 10% of strains being prioritized. If multiple strains exist in the same scoring interval, they are sorted based on the proximity of the principal component spatial distance to the historically high-yield cluster. The sorter locates the target strain based on the strain number using microfluidic chip sorting technology. After sorting, the strain is automatically inoculated into a new fermenter, and the sorting timestamp and culture parameters are simultaneously recorded. If the strain is classified as high-yielding, the pH is adjusted normally. If the strain is in the logarithmic growth phase, an additional pH compensation of 0.05 is added. The basic feed rate is matched according to the historical parameter mapping table. If the key enzyme activity is detected to be greater than 2 standard deviations of the mean, the feed rate is increased by an additional 10%; Terminate the cultivation of candidate strains with metabolic abnormalities in the fermentation tank, turn off the feed pump and start the tank cleaning program, record the main component characteristics and operation logs of the abnormal strains, push them to the strain feature library and mark them as eliminated data, and after each batch is sorted, update the fitness weight coefficient according to the gradient descent method based on the difference between the actual yield and the predicted score, and remove invalid parameter entries with actual yields 20% lower than the same category average from the historical parameter mapping table every 7 days.

Images