Fermentation parameter intelligent decision-making method, device and equipment based on deep learning and medium thereof

Abstract

The invention relates to a fermentation parameter intelligent decision-making method and device based on deep learning, equipment and a medium. The method comprises the following steps: collecting time series data of two modes in a microbial fermentation process and performing multi-dimensional feature analysis to obtain a morphological feature vector and a metabolic feature matrix; a fusion feature matrix is obtained through time dimension matching, and a dynamic correlation intensity curve is generated through nonlinear coupling modeling; metabolic fluctuation is marked abnormally, a cross-dimension anomaly recognition and multi-mode collaborative prediction model is constructed, data are fused and then input into the prediction model, and a regulation and control strategy is generated through long and short-term memory network optimization derivation. By adopting the method, multi-modal data fusion and cross-scale correlation analysis can be realized, the fermentation abnormity identification accuracy and parameter regulation and control scientificity are improved, and the product yield, purity and batch stability control capability are enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of bioengineering, and particularly relates to a fermentation parameter intelligent decision-making method, device and equipment based on deep learning and a medium thereof. BACKGROUND

[0002] In the field of bioengineering, as the bio-manufacturing industry continues to improve the yield, purity and batch stability of fermentation products, fermentation parameter regulation technology has gradually developed. In the early stage, it mainly relies on manual experience combined with basic detection, and in the later stage, simple automatic equipment is introduced, and the core is to regulate single-dimensional data to maintain the fermentation process.

[0003] In the traditional technology, fermentation parameter regulation is mainly processed in two ways: one is manual regulation, technicians regularly take samples to detect the concentration of metabolic products, observe the morphology of microorganisms with a microscope, and adjust the process parameters according to historical experience to judge the fermentation state; the other is single-mode automatic regulation, which collects single data through sensing equipment and automatically adjusts the parameters according to the preset threshold relying on a simple linear model.

[0004] However, these traditional methods have significant problems: they cannot capture the dynamic changes of microbial morphology and metabolic products in real time, and the lack of internal correlation between the two leads to incomplete feature representation; there is a lack of cross-scale correlation modeling mechanism, and the parameter adjustment lacks scientific basis; abnormal identification relies on single index threshold, which is difficult to identify multi-dimensional coupled abnormalities and cannot accurately mark abnormal time; the model is a static fixed mode and cannot be iteratively optimized according to the morphological evolution data and metabolic response data after regulation, resulting in low regulation accuracy and large product batch differences, which cannot meet the high-quality and high-efficiency industrial production needs of the bio-manufacturing industry. SUMMARY

[0005] Therefore, it is necessary to provide a fermentation parameter intelligent decision-making method, device, equipment and medium based on deep learning, which can perform multi-modal fusion, cross-scale modeling, intelligent abnormal identification and accurate parameter regulation, to realize fusion analysis of multi-modal time series data and analysis of cross-scale correlation features, and to improve the accuracy of fermentation state prediction and regulation strategy generation through nonlinear coupling modeling and long short-term memory network optimization, thereby enhancing the control ability of product yield, purity and batch stability.

[0006] In a first aspect, the application provides a fermentation parameter intelligent decision-making method based on deep learning, comprising:

[0007] Collecting time series data of two modalities in the microbial fermentation process, performing multi-dimensional feature analysis on the time series data to obtain a morphology feature vector representing the morphology features of the microorganisms and a metabolic feature matrix representing the distribution law of the metabolic products;

[0008] a fusion feature matrix is obtained based on a time dimension matching rule of the morphological feature vector and the metabolic feature matrix;

[0009] Based on the fusion feature matrix, a dynamic correlation strength curve is generated by modeling the cross-scale correlation between the morphological dynamic change of microorganisms and the concentration fluctuation of metabolic products through a nonlinear coupling analysis technique.

[0010] Abnormal labeling is performed on metabolic fluctuations that exceed the historical statistical range during the fermentation process, and abnormal labeling data containing abnormal features and time stamps are formed.

[0011] Based on the dynamic correlation strength curve and the abnormal labeling data, a cross-dimensional anomaly recognition model and a multi-modal collaborative prediction model are constructed.

[0012] The dynamic correlation strength curve and the abnormal labeling data are input into the cross-dimensional anomaly recognition model, and the abnormal state label and the abnormal confidence are output. The abnormal state label and the abnormal confidence are fused with the fusion feature matrix to obtain multi-modal collaborative input data.

[0013] The multi-modal collaborative input data is input into the multi-modal collaborative prediction model, and the multi-modal collaborative input data is optimized and deduced through a long short-term memory network algorithm to generate a regulation strategy.

[0014] In one embodiment, multi-dimensional feature analysis is performed on time series data to obtain a morphological feature vector representing the morphological features of microorganisms and a metabolic feature matrix representing the distribution rule of metabolic products, including:

[0015] The time series data includes microbial image data and metabolomics data.

[0016] The microbial image data is subjected to image feature extraction processing to generate a morphological feature vector.

[0017] The metabolomics data is subjected to metabolic feature dimension reduction processing to obtain a metabolic feature matrix.

[0018] In one embodiment, a fusion feature matrix is obtained based on a time dimension matching rule of the morphological feature vector and the metabolic feature matrix, including:

[0019] The morphological feature vector is subjected to time series sliding window decomposition to obtain a morphological feature subspace sequence.

[0020] The metabolic feature matrix is subjected to metabolic feature subspace projection to obtain a metabolic feature subspace sequence.

[0021] The optimal matching path of the morphological feature subspace sequence and the metabolic feature subspace sequence is calculated by a dynamic time warping algorithm, and the time offset is determined based on the time displacement parameter of the optimal matching path.

[0022] The morphological feature subspace sequence and the metabolic feature subspace sequence are time axis aligned and compensated according to the time offset to generate a synchronized feature subspace sequence.

[0023] The synchronized feature subspace sequence is subjected to multi-scale nonlinear coupling analysis to obtain a fusion feature matrix.

[0024] In one embodiment, the fusion feature matrix is obtained based on a time dimension matching rule of the morphological feature vector and the metabolic feature matrix, and includes:

[0025] The morphological feature vector is subjected to time sequence sliding window decomposition to obtain a morphological feature subspace sequence.

[0026] The metabolic feature matrix is subjected to metabolic feature subspace projection to obtain a metabolic feature subspace sequence.

[0027] The optimal matching path of the morphological feature subspace sequence and the metabolic feature subspace sequence is calculated by a dynamic time warping algorithm, and the time offset is determined based on the time displacement parameter of the optimal matching path.

[0028] The morphological feature subspace sequence and the metabolic feature subspace sequence are time axis aligned and compensated according to the time offset to generate a synchronized feature subspace sequence.

[0029] The synchronized feature subspace sequence is subjected to multi-scale nonlinear coupling analysis to obtain a fusion feature matrix.

[0030] In one embodiment, based on the fusion feature matrix, the cross-scale correlation relationship between the morphological dynamic change and the metabolic product concentration fluctuation of the microorganism is modeled by a nonlinear coupling analysis technique to generate a dynamic correlation strength curve, including:

[0031] The fusion feature matrix is subjected to separation processing to obtain a morphological feature subspace and a metabolic feature subspace.

[0032] The time offset of the morphological feature subspace and the metabolic feature subspace is calculated by a dynamic time warping algorithm, and a synchronized feature subspace sequence is generated based on the time offset.

[0033] The dynamic evolution pattern of the morphological feature subspace and the concentration fluctuation pattern of the metabolic feature subspace are subjected to cross-scale nonlinear correlation analysis by a cross-scale correlation relationship algorithm to generate a dynamic correlation strength sequence.

[0034] Based on the time sequence distribution characteristics of the dynamic correlation strength sequence, a dynamic correlation strength curve is generated by a weighted integral algorithm.

[0035] In one of the embodiments, the multi-modal collaborative input data is input into a multi-modal collaborative prediction model, and the multi-modal collaborative input data is optimized and deduced through a long short-term memory network algorithm to generate a regulation strategy, including:

[0036] The multi-modal collaborative input data is input into a multi-modal collaborative prediction model, and multi-modal time series feature collaborative prediction processing is performed through a long short-term memory network algorithm to generate prediction trend data of the fermentation state trend;

[0037] Based on the direction and amplitude of the change in the concentration of the metabolite in the prediction trend data, multi-objective parameter optimization and deduction are performed on the temperature adjustment amount, the pH correction amount, and the feeding strategy to generate a regulation strategy of the fermentation process parameters.

[0038] In one of the embodiments, after constructing a cross-dimension anomaly recognition model and a multi-modal collaborative prediction model based on a dynamic correlation strength curve and abnormal marker data, and determining the regulation strategy, the method further includes:

[0039] Collecting microbial morphological change data and metabolic response data after regulation;

[0040] Performing feature dimension alignment processing on the microbial morphological change data and the metabolic response data to generate a feedback feature set;

[0041] According to the feedback feature set, a loss function is constructed, which is used to update the network parameters of the multi-modal collaborative prediction model through a gradient descent algorithm to obtain an updated multi-modal collaborative prediction model and optimized network parameters;

[0042] According to the optimized network parameters, the contribution weights of the morphological feature vectors and the metabolic feature matrices in the fused feature matrix are re-evaluated to obtain optimized feature fusion weights;

[0043] The initial parameter set for the next intelligent regulation is obtained by combining the optimized fused feature matrix, the updated multi-modal collaborative prediction model, and the optimized feature fusion weights.

[0044] In one of the embodiments, the multi-modal collaborative prediction model includes an attention mechanism decoder and a metabolic dynamics encoder, wherein:

[0045] The attention mechanism decoder is used to dynamically weight the intermediate features to generate a first candidate regulation correlation path;

[0046] The metabolic dynamics encoder is used to generate a second candidate regulation correlation path based on a metabolomics feature correlation graph and a fermentation process dynamics model;

[0047] The first candidate regulation correlation path and the second candidate regulation correlation path are fused through a confusion network algorithm to generate an optimal regulation strategy.

[0048] In a second aspect, the present application also provides a deep learning-based intelligent decision-making device for fermentation parameters, comprising:

[0049] a multi-modal time-series acquisition module, configured to acquire time-series data of two modalities in a microbial fermentation process, perform multi-dimensional feature analysis on the time-series data, and obtain a morphology feature vector representing microbial morphology features and a metabolism feature matrix representing metabolism product distribution rules;

[0050] a fusion feature matrix generation module, configured to obtain a fusion feature matrix based on a time dimension matching rule of the morphology feature vector and the metabolism feature matrix;

[0051] a dynamic correlation modeling module, configured to model a cross-scale correlation between microbial morphology dynamic changes and metabolism product concentration fluctuations based on the fusion feature matrix by using a nonlinear coupling analysis technique, and generate a dynamic correlation strength curve;

[0052] an abnormal fluctuation detection and labeling module, configured to implement abnormal labeling on metabolism fluctuations that are beyond a historical statistical range in the fermentation process, and form abnormal labeling data containing abnormal features and time stamps;

[0053] a model construction module, configured to construct a cross-dimensional abnormality recognition model and a multi-modal collaborative prediction model based on the dynamic correlation strength curve and the abnormal labeling data;

[0054] a data fusion module, configured to input the dynamic correlation strength curve and the abnormal labeling data into the cross-dimensional abnormality recognition model, output abnormal state labels and abnormal confidence, and fuse the abnormal state labels and the abnormal confidence with the fusion feature matrix to obtain multi-modal collaborative input data;

[0055] an intelligent decision-making generation module, configured to input the multi-modal collaborative input data into the multi-modal collaborative prediction model, optimize and deduce the multi-modal collaborative input data by using a long short-term memory network algorithm, and generate a regulation strategy.

[0056] In a third aspect, the present application also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any method in the first aspect of the present application when executing the computer program.

[0057] In a fourth aspect, the present application also provides a computer-readable storage medium, which stores a computer program, and the computer program implements the steps of any method in the first aspect of the present application when executed by a processor.

[0058] The application provides a fermentation parameter intelligent decision-making method based on deep learning, which realizes the feature cascade and time axis alignment of microbial image data and metabolomics data by constructing a multi-modal data fusion analysis system, and solves the problem of morphological and metabolic feature fragmentation in traditional single-mode regulation. Based on the dynamic time warping algorithm, the synchronization processing of morphological feature subspace and metabolic feature subspace is completed, a quantitative correlation model of morphological dynamic evolution and metabolic concentration fluctuation is established through cross-scale nonlinear correlation analysis, and a dynamic correlation strength curve is generated. An innovative anomaly detection mechanism is used to implement accurate timestamp labeling for metabolic fluctuations exceeding the historical statistical threshold. A cross-dimensional anomaly recognition model and a multi-modal collaborative prediction model are constructed, and the anomaly state label and the correlation strength feature are fused to form collaborative input data. The long short-term memory network is used to dynamically evolve and predict the multi-modal time series features, and the multi-objective optimization parameters of temperature, pH and feeding strategy are deduced based on the concentration change direction of metabolic products. A closed-loop feedback mechanism is established, the model parameters and feature fusion weights are updated through the regulated data, the initial parameter set of the next regulation is continuously optimized, and finally the multi-parameter collaborative precise regulation of the fermentation process is realized, and the product yield, purity and batch stability control ability are improved. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiment or related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0060] Figure 1 The implementation environment schematic diagram provided for an embodiment of the present application is shown in the figure.

[0061] Figure 2 The flow chart of the fermentation parameter intelligent decision-making method based on deep learning in an embodiment of the present application is shown in the figure.

[0062] Figure 3 The structure schematic diagram of the fermentation parameter intelligent decision-making device based on deep learning in an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0064] Firstly, the terms involved in the embodiments of the present application are briefly introduced.

[0065] Time series data: two types of modal data collected continuously over time during fermentation, which are the raw materials for subsequent feature analysis, fusion analysis and modeling.

[0066] Nonlinear coupling analysis technology: for the time-aligned morphological features and metabolic features in fermentation, a nonlinear mathematical model is used to capture the cross-scale correlation rules of morphological dynamic changes and metabolic concentration fluctuations, generate dynamic correlation strength curves that reflect the correlation strength changes over time, solve the problems of ignoring the correlation between morphology and metabolism in traditional fermentation and inaccurate correlation analysis, and provide quantitative basis for subsequent anomaly identification and control strategy.

[0067] Cross-dimensional anomaly identification model: integrating microbial morphological features and metabolic features in fermentation process, combining dynamic correlation strength curves and anomaly marker data, identifying multi-dimensional coupled fermentation anomaly states, outputting anomaly state labels and confidence, solving the problem of traditional single-index threshold detection difficult to find complex correlation anomalies.

[0068] Multi-modal collaborative prediction model: integrating time series data of microbial morphological features and metabolic features in fermentation process, combining anomaly state information, using long short-term memory network to analyze the time dependence of multi-modal data, deriving temperature, pH, and feed control strategies, solving the problem of traditional single-modal prediction difficult to optimize multiple parameters collaboratively.

[0069] Long short-term memory network algorithm (LSTM): processing long-term dependence of multi-modal time series data in fermentation, combining metabolic change trend to derive collaborative control strategies of temperature, pH and feed, solving the problem of traditional model difficult to accurately predict multi-parameter coupled changes.

[0070] LC-MS (English full name: Liquid Chromatography-Mass Spectrometry, commonly used abbreviation: LC-MS) is a combined analysis technology that combines the separation ability of liquid chromatography (LC) with the detection ability of mass spectrometry (MS). Its core is to separate the target components in complex mixtures by liquid chromatography module, and then introduce the separated components into mass spectrometry module one by one. The mass spectrometry module can accurately identify the molecular mass and structure characteristics of the components and accurately determine the content of the components. This technology combines the high separation efficiency of liquid chromatography with the high sensitivity and high specificity of mass spectrometry, which can effectively solve the analysis problem of trace and trace target components in complex matrix, and is widely used in the fields of drug component analysis and metabolism research in biological and medical fields, pollutant detection in environmental monitoring field, harmful substance screening in food detection field, and other scenes that require accurate component analysis of complex samples.

[0071] U-Net (U-Net Convolutional Neural Network): A convolutional neural network architecture designed specifically for image segmentation tasks, characterized by a "U" shaped network structure composed of an encoder and a decoder. The encoder performs down-sampling through alternating convolutional layers and pooling layers, gradually extracting features from low-order to high-order, while reducing the spatial resolution of the feature maps to expand the receptive field. The decoder restores the spatial resolution of the feature maps through transpose convolutional layers, and at each up-sampling stage, it fuses the corresponding layer features from the encoder through a skip connection mechanism to supplement the details lost during up-sampling. U-Net effectively balances feature extraction depth and detail preservation accuracy, significantly improving the segmentation and positioning accuracy of complex targets.

[0072] ResNet-50 (Residual Neural Network-50): A deep residual convolutional neural network with 50 trainable convolutional layers. Its core is to establish a connection path between shallow and deep features through a "residual connection" mechanism to alleviate the "gradient vanishing" problem in deep network training. It also uses a "1x1 convolution-3x3 convolution-1x1 convolution" bottleneck residual block to balance computational efficiency and feature extraction capability. The network consists of an input layer, 4 stacked residual block groups, a global average pooling layer, and a fully connected layer. It can effectively extract low-order to high-order features by increasing the number of feature channels and reducing the spatial resolution at each level, significantly improving model generalization and training stability. It is widely used in computer vision tasks such as image classification, object detection, and image segmentation, which require deep feature extraction.

[0073] Dynamic Time Warping (DTW): An algorithm for measuring the similarity of time series. Its core is to construct a distance matrix between two time series through dynamic programming, and find the optimal alignment path in the matrix that minimizes the cumulative distance, to solve the problem of traditional Euclidean distance matching failure caused by different sequence lengths or time rhythm differences. This algorithm can effectively capture the intrinsic similarity of sequences under non-linear time stretching, and is widely used in speech recognition, gesture recognition, sensor time series data comparison, and other scenarios that require similarity analysis of time series.

[0074] Z-Score (Standard Score, commonly abbreviated as "Z-Score"): a statistical analysis tool for data standardization, which is based on the mean and standard deviation of the data set to convert the original data into dimensionless standardized values. This standardization process can eliminate the dimensional differences between different data, achieve cross-data set, cross-index data uniform scale comparison, and identify outliers in the data set by setting threshold values. It is widely used in machine learning feature preprocessing, industrial quality control, financial risk assessment, and other scenarios that require unified scale analysis or anomaly detection of data. Liquid chromatography-mass spectrometry (LC-MS): a combined analysis technique that combines the high-efficiency separation capability of liquid chromatography (LC) with the high-sensitivity detection capability of mass spectrometry (MS). Its core is to use the mobile phase to drive the differential distribution of complex mixtures on the stationary phase through the liquid chromatography module, achieving high-efficiency separation of different components in the mixture. Then, the separated components are introduced into the mass spectrometry module one by one, and the molecular mass and structural characteristics of the components are accurately determined through ionization and mass analysis. At the same time, the component content is accurately quantified. This technique combines the high separation efficiency of liquid chromatography with the high specificity and high sensitivity of mass spectrometry, effectively breaking through the analysis bottleneck of trace and trace target components in complex matrices, and is widely used in drug component analysis and metabolism research in the biomedicine field, pollutant screening in the environmental monitoring field, harmful substance detection in the food detection field, and other scenarios that require accurate component analysis of complex samples.

[0075] Principal Component Analysis (PCA, commonly abbreviated as "PCA"): a statistical method for data dimensionality reduction and feature extraction, which is based on orthogonal transformation to map high-dimensional original data to low-dimensional principal component space. Each principal component is a linear combination of original features and is orthogonal to each other, sorted by the contribution of variance from large to small. By selecting the first k principal components, the data dimension can be significantly reduced while preserving the main information, effectively removing feature redundancy, simplifying data structure, and reducing subsequent computational complexity. It is widely used in machine learning feature preprocessing, pattern recognition, data visualization, signal processing, and other scenarios that require simplified analysis of high-dimensional data.

[0076] Multi-Stage Convolutional Neural Network: It is an improved deep learning model based on convolutional neural network architecture, characterized by dividing the feature extraction and task processing process of the neural network into multiple stages with clear functional boundaries. Each stage implements feature processing through hierarchical progression or parallel collaboration: The pre-sequencing stage mainly extracts low-order features of the image through convolution, pooling, and other operations, and the subsequent stage further extracts and optimizes high-order features based on the output of the pre-sequencing stage. Each stage can achieve information interaction through feature fusion mechanism. Through this multi-stage division and cooperation mode, the model's ability to capture multi-level features in complex scenarios can be improved, thereby improving the processing accuracy and robustness of computer vision tasks such as image classification and target detection.

[0077] The deep learning-based fermentation parameter intelligent method provided by the embodiments of the present application can be applied to an application environment as shown in Figure 1 The terminal 100 is connected to the detection instrument 101 through a network. The terminal 100 can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices, wherein the portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The detection instrument can be, but is not limited to, a microscope, a sensor, a gas chromatograph-mass spectrometer, and a liquid chromatograph-mass spectrometer.

[0078] As shown in Figure 2 The present application provides a deep learning-based fermentation parameter intelligent decision-making method. Taking the application of the method to the terminal 100 as an example, the method comprises:

[0079] S101: Collecting time series data of two modalities in the microbial fermentation process, and the terminal 100 performs multi-dimensional feature analysis on the time series data to obtain a morphological feature vector representing the morphological characteristics of the microorganism and a metabolic feature matrix representing the distribution rule of the metabolites.

[0080] Exemplarily, the time series data covers two types of modalities, i.e., microbial microscopic image data and metabolomics detection data. For image data, spatial features such as bacterial morphology contour and colony spatial structure are extracted through image analysis technology, and a morphological feature vector is generated after feature integration; for metabolomics data, professional metabolomics detection instruments such as liquid chromatograph-mass spectrometer are used to quantitatively and qualitatively analyze metabolites in the fermentation system, thereby collecting metabolomics data.

[0081] S102: Obtaining a fused feature matrix based on the time dimension matching rule of the morphological feature vector and the metabolic feature matrix.

[0082] Exemplarily, to achieve the precise time alignment of morphological and metabolic characteristics, the terminal 100 performs a time window decomposition technique on the morphological characteristic vector to obtain a morphological characteristic subspace sequence; simultaneously, a characteristic projection technique is performed on the metabolic characteristic matrix to generate a metabolic characteristic subspace sequence; an optimal matching path of the two types of subspace sequences is calculated through a time axis alignment algorithm, a time offset is determined based on a path time displacement parameter, and time axis alignment compensation processing is performed on the sequence to generate a synchronized characteristic subspace sequence. Finally, a multi-scale correlation analysis technique is performed on the synchronized sequence to integrate the time sequence correlation information of morphological and metabolic characteristics, and a fusion characteristic matrix is obtained.

[0083] S103: Based on the fusion characteristic matrix, a cross-scale correlation relationship between the dynamic changes of microbial morphology and the concentration fluctuations of metabolic products is modeled through a nonlinear coupling analysis technique to generate a dynamic correlation strength curve.

[0084] Exemplarily, the terminal 100 performs feature separation processing on the fusion characteristic matrix to disassemble independent morphological characteristic subspaces and metabolic characteristic subspaces, uses a dynamic time warping algorithm to verify and ensure the time synchronization of the two types of subspaces again to generate a synchronized characteristic subspace sequence; uses a cross-scale correlation relationship algorithm combined with a nonlinear modeling technique to perform correlation analysis on the dynamic evolution mode of the morphological characteristic subspace and the concentration fluctuation mode of the metabolic characteristic subspace, captures the correlation strength changes across time steps through a neural network algorithm with time sequence dependence capturing capability, and generates a dynamic correlation strength sequence; based on the time sequence distribution characteristics of the dynamic correlation strength sequence, a sliding weighted integral algorithm is used to perform smoothing integration processing on the sequence, and finally a dynamic correlation strength curve is generated, wherein the neural network algorithm can be selected from types such as a gated recurrent neural network.

[0085] S104: Abnormal markers are implemented for metabolic fluctuations that exceed the historical statistical range in the fermentation process to form abnormal marker data containing abnormal characteristics and time stamps.

[0086] Exemplarily, based on the metabolomics data accumulated in multiple batches of normal fermentation processes, a historical statistical range of the concentration of each metabolic product is constructed, which needs to cover the metabolic fluctuation interval in the normal fermentation scenario; in the current fermentation process, the change of the metabolomics data is monitored in real time, when the concentration of a certain metabolic product exceeds the preset historical statistical range, an abnormal marker mechanism is triggered, and the metabolic characteristic information corresponding to the abnormality and the accurate time stamp of the abnormality occurrence are recorded; the characteristic information and time stamp of all abnormal events are integrated to form structured abnormal marker data.

[0087] S105: Based on the dynamic correlation strength curve and the abnormal marker data, a cross-dimensional abnormality recognition model and a multi-modal collaborative prediction model are constructed.

[0088] Exemplarily, the peak distribution of the dynamic correlation strength curve, the abnormal features in the abnormal marker data, the timestamp information, and the like are taken as model training samples, wherein the samples need to cover data of both normal fermentation and abnormal fermentation; a nonlinear classification algorithm is used to construct a cross-dimensional anomaly recognition model for distinguishing between normal and abnormal states in the fermentation process; meanwhile, a deep learning algorithm with time series prediction capability is used to construct a multi-modal collaborative prediction model with the fusion feature matrix, the dynamic correlation strength curve, and the abnormal marker data as multi-modal input features, for predicting the future fermentation state trend; the nonlinear classification algorithm can be a support vector machine, and the time series prediction algorithm can be a long short-term memory network; the process is constructed by means of double models, and functions of fermentation anomaly recognition and state prediction are realized respectively, so as to cover the core requirements of fermentation process control, solve the problem of incomplete control response caused by the single function of the traditional model, and improve the adaptation capability of the model to complex fermentation scenarios.

[0089] S106: The dynamic correlation strength curve and the abnormal marker data are input into the cross-dimensional anomaly recognition model, and an abnormal state label and an abnormal confidence are output, and the abnormal state label and the abnormal confidence are fused with the fusion feature matrix to obtain multi-modal collaborative input data.

[0090] The dynamic correlation strength curve and the abnormal marker data are input into the cross-dimensional anomaly recognition model, and an abnormal state label and an abnormal confidence are output, and the abnormal state label and the abnormal confidence are fused with the fusion feature matrix to obtain multi-modal collaborative input data. Specifically, the dynamic correlation strength curve and the abnormal marker data are input into the cross-dimensional anomaly recognition model, and the model outputs an abnormal state label and an abnormal confidence; the abnormal state label and the confidence are spliced with the fusion feature matrix to form multi-modal collaborative input data. The fusion of the abnormal state information and the fusion feature matrix provides input containing abnormal semantics for multi-modal collaborative prediction, and enhances the understanding ability of the prediction model for the fermentation state.

[0091] S107: The multi-modal collaborative input data are input into the multi-modal collaborative prediction model, and the multi-modal collaborative input data are optimized and deduced by the long short-term memory network algorithm to generate a control strategy.

[0092] The multi-modal collaborative input data are input into the multi-modal collaborative prediction model, and the multi-modal collaborative input data are optimized and deduced by the long short-term memory network algorithm to generate a control strategy. Specifically, the multi-modal collaborative input data are input into the multi-modal collaborative prediction model, and the long short-term memory network algorithm is used to capture the time series dependence of the fermentation state to predict the future change trend of the metabolic product concentration; based on the direction and amplitude of the concentration change in the predicted trend, multi-objective parameter optimization is performed on the temperature adjustment amount, the pH correction amount, and the feeding strategy to generate a control strategy for the fermentation process parameters.

[0093] The technical scheme provided in the application comprises the following technical effects: the patent application constructs a multi-modal data fusion analysis system, realizes feature cascade and time axis alignment of microbial image data and metabolomics data, and solves the problem of split of morphological and metabolic characteristics in traditional single-modal regulation. Synchronization processing of morphological feature subspace and metabolic feature subspace is completed based on a dynamic time warping algorithm, a quantitative correlation model of morphological dynamic evolution and metabolic concentration fluctuation is established through cross-scale nonlinear correlation analysis, and a dynamic correlation strength curve is generated. An innovative anomaly detection mechanism is used to implement accurate timestamp labeling on metabolic fluctuations exceeding historical statistical thresholds. A cross-dimension anomaly identification model and a multi-modal collaborative prediction model are constructed, and an anomaly state label and a correlation strength feature are fused to form collaborative input data. A long short-term memory network is used to dynamically evolve and predict multi-modal time series features, and multi-objective optimization parameters of temperature, pH and feeding strategy are deduced based on the concentration change direction of metabolic products. A closed-loop feedback mechanism is established, model parameters and feature fusion weights are updated through post-regulation data, the initial parameter set of the next regulation is continuously optimized, and finally multi-parameter collaborative precise regulation of the fermentation process is realized, and the control ability of product yield, purity and batch stability is improved.

[0094] On the basis of the above-mentioned embodiments, multi-dimensional feature analysis is performed on the time series data to obtain a morphological feature vector representing the morphological characteristics of the microorganism and a metabolic feature matrix representing the distribution rule of the metabolic products, comprising:

[0095] Step 201: The time series data comprises microbial image data and metabolomics data.

[0096] The microbial image data is collected by an industrial-grade microscopic imaging system, covering bright field / fluorescent microscopic images, live cell dynamic tracking images, etc., with a time resolution of 1 minute / frame and a spatial resolution of microns, accurately capturing the morphological dynamic changes of the microorganism in the fermentation process; the metabolomics data is collected by a liquid chromatography-mass spectrometry platform, covering the concentration, type and metabolic pathway activity information of thousands of metabolic products in the fermentation system, with a collection frequency of once per hour, comprehensively reflecting the material basis of microbial metabolic activity.

[0097] Step 202: Image feature extraction processing is performed on the microbial image data to generate a morphological feature vector.

[0098] The multi-stage convolutional neural network architecture is adopted to realize feature extraction. The adaptive median filtering technology is used to eliminate optical noise in the image, and then the contrast limited histogram equalization is used to enhance the visual distinction between the bacteria and the background. Then, the U-Net-based semantic segmentation model is used to accurately segment the boundary of a single bacterium or a bacterial colony, and the morphological region of the microorganism is output. The pre-trained ResNet-50 network is used to extract the shallow features and deep semantic features in the image. After global average pooling and feature compression operation, the morphological feature vector containing 30-50 dimensional key morphological indicators is generated, which covers three types of information including geometric morphology, spatial structure and dynamic evolution trend. Finally, the morphological feature vectors at multiple time points are time-series spliced and Z-Score normalized to ensure the consistency of the feature dimension and the time sequence continuity, and finally the morphological feature vector representing the dynamic evolution of the microorganism morphology is obtained, thereby improving the accuracy of the fermentation state representation.

[0099] Step 203: Metabolic feature dimension reduction processing is performed on the metabolomics data to obtain a metabolic feature matrix.

[0100] For the metabolomics raw data collected by the LC-MS combined technology platform, the peak alignment is performed to eliminate the peak displacement deviation caused by the instrument drift, the low signal-to-noise ratio peaks are filtered out, and the total ion flow is normalized to complete the preprocessing. Then, the PCA is used to compress the high-dimensional metabolic data to obtain a low-dimensional feature matrix representing the key metabolic rules at each time point and each table. Finally, the metabolic feature matrix is obtained after combining the metabolic pathway knowledge and labeling the biological significance of the features. The key information of the fermentation is retained to provide a clear basis for subsequent fermentation state judgment and regulation.

[0101] In an embodiment of the present application, based on the time dimension matching rule of the morphological feature vector and the metabolic feature matrix, a fusion feature matrix is obtained, including:

[0102] Step 301: Time-series sliding window decomposition is performed on the morphological feature vector to obtain a morphological feature subspace sequence.

[0103] For the morphological feature vector collected in the continuous fermentation process, the time-series sliding window technology is used for processing: a fixed time window is set to divide the continuous fermentation time into several equal intervals. After feature recombination and dimension reduction, the morphological feature vector in each window generates a morphological feature subspace sequence corresponding to the time interval. This step converts the continuous time sequence of morphological features into a discrete subspace sequence through window decomposition, providing a basic unit for subsequent time alignment with metabolic features.

[0104] Step 302: Metabolic feature subspace projection is performed on the metabolic feature matrix to obtain a metabolic feature subspace sequence.

[0105] For the metabolic feature matrix, the metabolic feature subspace projection technology is adopted: the high-dimensional metabolic feature matrix is mapped to a low-dimensional feature subspace by using linear projection or nonlinear projection method, to generate a metabolic feature subspace sequence corresponding to each time point. This step reduces the feature dimension while retaining the key information of the metabolism through subspace projection, providing a lightweight input for the subsequent time sequence matching of morphological features.

[0106] Step 303: Calculate the optimal matching path of the morphological feature subspace sequence and the metabolic feature subspace sequence by the dynamic time warping algorithm, and determine the time offset based on the time displacement parameter of the optimal matching path.

[0107] The morphological feature subspace sequence and the metabolic feature subspace sequence are input into the DTW algorithm, and the optimal matching path of the two types of sequences is solved by dynamic programming: the algorithm calculates the similarity of subspace features at different time points, iteratively finds the matching path that maximizes the cumulative similarity or minimizes the distance, and determines the time offset of the two types of data based on the time displacement parameter of the optimal matching path. This step uses the elastic matching capability of the DTW algorithm to accurately capture the asynchrony of morphological and metabolic features in the time dimension, providing a quantitative basis for subsequent time alignment.

[0108] Step 304: Time axis alignment compensation is performed on the morphological feature subspace sequence and the metabolic feature subspace sequence according to the time offset to which they belong, to generate a synchronized feature subspace sequence.

[0109] Time axis alignment compensation is performed on the morphological feature subspace sequence and the metabolic feature subspace sequence: if the metabolic feature sequence lags behind, the metabolic sequence is time-shifted forward for interpolation, so that the two types of sequences are strictly aligned in the time axis; a synchronized feature subspace sequence is generated after alignment. This step eliminates the time asynchrony between image acquisition and metabolic detection through time offset compensation, achieving precise time synchronization of morphological and metabolic features.

[0110] Step 305: Perform multi-scale nonlinear coupling analysis on the synchronized feature subspace sequence to obtain a fusion feature matrix.

[0111] The synchronized feature subspace sequence is processed using multi-scale nonlinear coupling analysis technology: first, the sequence is decomposed into different time scales by wavelet transform, and the morphological-metabolic correlation features at each scale are extracted; then, the attention mechanism is introduced to dynamically weight the correlation features at different scales; finally, the weighted multi-scale features are fused to generate a fusion feature matrix. This step deeply excavates the internal correlation between morphological dynamic changes and metabolic fluctuations through multi-scale analysis and nonlinear coupling, providing a fusion feature input for precise characterization of fermentation state.

[0112] In an embodiment of the present application, the multi-modal collaborative input data is input into a multi-modal collaborative prediction model, and the multi-modal collaborative input data is optimized and deduced through a long short-term memory network algorithm to generate a regulation strategy, including:

[0113] Step 401: input the multi-modal collaborative input data into the multi-modal collaborative prediction model, and perform multi-modal time series feature collaborative prediction processing through a long short-term memory network algorithm to generate prediction trend data of the fermentation state trend.

[0114] The multi-modal collaborative input data of the fusion dynamic correlation strength curve, the abnormal state label, the abnormal confidence and the fusion feature matrix are acquired, and are input into a multi-modal collaborative prediction model; the model is based on a long short-term memory network architecture, and through a multi-modal feature embedding layer, the time series fluctuation mode of the dynamic correlation strength curve, the semantic information of the abnormal state label and the morphological-metabolic correlation features of the fusion feature matrix are vectorized and encoded to be unified into a tensor format that can be processed by the long short-term memory network; a bidirectional long short-term memory network layer is used to capture long-distance dependence in the time dimension of the fermentation process, and an attention mechanism layer is combined to dynamically weight the features of the key time steps, and finally generate fermentation state prediction trend data including the change trend of the metabolite concentration, the evolution trend of the microbial growth phase and the abnormal risk evolution trend.

[0115] Step 402: based on the direction and amplitude of the change of the metabolite concentration in the prediction trend data, multi-objective parameter optimization deduction is performed on the temperature regulation amount, the pH correction amount and the feeding strategy to generate a regulation strategy of the fermentation process parameters.

[0116] The direction and amplitude of the change of the metabolite concentration are extracted from the fermentation state prediction trend data, and a fermentation kinetics model describing the quantitative relationship between temperature, pH, feeding amount and microbial growth rate, metabolite synthesis amount is combined; a multi-objective optimization algorithm with constraints is used to iteratively optimize the process parameters such as temperature regulation amount, pH correction amount and feeding strategy; a target function is constructed with the core optimization objectives of maximizing the target product yield, minimizing the byproduct generation amount and maintaining the stability of the pH in the fermentation process, and the parameter combination that optimizes the multi-objective function is solved to generate a fermentation process parameter regulation strategy including the parameter type, the adjustment amount and the execution time.

[0117] In an embodiment of the present application, after the cross-dimension abnormality recognition model and the multi-modal collaborative prediction model are constructed based on the dynamic correlation strength curve and the abnormal label data, and the regulation strategy is determined, the method further includes:

[0118] Microbial morphological change data and metabolic response data of the biological fermentation after regulation are collected.

[0119] After the intelligent control strategy is executed in the fermentation process, morphological change time series data of microorganisms in the fermentation system are collected by an industrial-grade microscopic imaging system, covering information such as bacterial morphological profile, colony spatial structure, and dynamic evolution of subcellular organelles; meanwhile, metabolic response time series data are collected by a liquid chromatography-mass spectrometry platform, including concentration, type, and metabolic pathway activity change information of thousands of metabolites in the fermentation broth; the time collection frequency of the two types of data is matched with the fermentation stage.

[0120] The morphological change data and metabolic response data are processed for feature dimension alignment to generate a feedback feature set.

[0121] First, the time resolution is unified to the minute level by a time axis interpolation algorithm, and then feature mapping and dimension reduction techniques are used to realize dimension alignment: low-dimensional statistical features such as bacterial length-width ratio and colony density are extracted from morphological data, and high-dimensional metabolomics information is compressed through principal component analysis for metabolic data; finally, the feature tensors of the two types of data are spliced into a structured feedback feature set, ensuring the consistency of morphological and metabolic features in time and space dimensions.

[0122] According to the feedback feature set, a loss function is constructed, which is used to update the network parameters of the multi-modal collaborative prediction model through a gradient descent algorithm to obtain an updated multi-modal collaborative prediction model and optimized network parameters.

[0123] Based on the feedback feature set, a loss function containing three parts of sub-loss is constructed: the first part is the mean square error loss of the predicted value and the true value, which quantifies the prediction deviation of the model for the fermentation state trend; the second part is the abnormal confidence loss, which constrains the accuracy of the model's identification confidence for abnormal states; the third part is the regularization loss, which avoids model overfitting; the gradient descent algorithm is executed through the adaptive matrix estimator optimizer to iteratively update the weight matrix and bias parameters of the multi-modal collaborative prediction model along the negative gradient direction of the loss function, and the prediction model with optimized parameters and network parameters are obtained after multiple iterations converge.

[0124] According to the optimized network parameters, the contribution weights of the morphological feature vector and the metabolic feature matrix in the fused feature matrix are re-evaluated to obtain the optimized feature fusion weights.

[0125] The optimized network parameters are input into the feature weight evaluation module, which quantifies the contribution of the morphological feature vector and the metabolic feature matrix to the model prediction by calculating the feature importance score; the contribution scores of the two types of features are normalized to obtain the dynamic fusion weight coefficients of the morphological features and the metabolic features, realizing dynamic adjustment of the contribution weights of different modal features in the fused feature matrix.

[0126] The initial parameter set for the next intelligent control is obtained by combining the optimized fused feature matrix, the updated multi-modal collaborative prediction model, and the optimized feature fusion weights.

[0127] The optimized feature fusion weight is applied to a fusion feature matrix generation process, a multi-modal collaborative prediction model is combined after parameter updating, an initial parameter set containing temperature adjustment, pH correction and feeding strategy is generated by model forward propagation based on initial state data of the current fermentation period; the parameter set is used as the starting input of intelligent regulation and control for the next fermentation period, so that the regulation and control strategy can adapt to the dynamic evolution characteristics of the fermentation system.

[0128] In an embodiment of the present application, the multi-modal collaborative prediction model comprises an attention mechanism decoder and a metabolic dynamics encoder, wherein:

[0129] The attention mechanism decoder is used for dynamically weighting the intermediate features to generate a first candidate regulation and control correlation path.

[0130] The multi-modal collaborative prediction model adopts a double-branch architecture to integrate the attention mechanism decoder and the metabolic dynamics encoder; the attention mechanism decoder dynamically weights the intermediate features through multi-head self-attention mechanism to focus on time series patterns strongly related to the fermentation state; the metabolic dynamics encoder realizes dynamic constraint coding of metabolic response features based on a pre-trained metabolomics feature correlation graph and a fermentation process dynamics model.

[0131] The metabolic dynamics encoder is used for generating a second candidate regulation and control correlation path based on the metabolomics feature correlation graph and the fermentation process dynamics model.

[0132] The attention mechanism decoder performs dynamic weighting on the intermediate features of the multi-modal collaborative input data, strengthens the weight of key time series features by calculating the attention scores between features, and generates a first candidate regulation and control correlation path based on the weighted feature sequence; the metabolic dynamics encoder converts the metabolomics feature correlation graph and the fermentation dynamics model into learnable parameters, transmits the dependency relationship of metabolic pathways through a graph neural network, and generates a second candidate regulation and control correlation path in combination with the dynamics equation constraint.

[0133] The first candidate regulation and control correlation path and the second candidate regulation and control correlation path are fused through a confusion network algorithm to generate an optimal regulation and control strategy.

[0134] The first candidate regulation and control correlation path and the second candidate regulation and control correlation path are input into the confusion network algorithm module, the module quantifies the regulation and control effect confidence of the two paths through a probability allocation mechanism, sorts and selects the candidate paths based on multi-objective optimization rules; a dynamic threshold decision strategy is adopted to adaptively adjust the fusion weight in different stages of fermentation, and finally an optimal regulation and control strategy containing temperature adjustment, pH correction and feeding strategy is output.

[0135] In an embodiment of the present application, as Figure 3The application also provides a deep learning-based fermentation parameter intelligent decision device 500, which comprises the following modules.

[0136] A multi-modal time sequence acquisition module 501 is configured to acquire time sequence data of two modes in a microbial fermentation process, perform multi-dimensional feature analysis on the time sequence data, and obtain a morphology feature vector representing microbial morphology features and a metabolism feature matrix representing metabolism product distribution rules.

[0137] A fusion feature matrix generation module 502 is configured to obtain a fusion feature matrix based on a time dimension matching rule of the morphology feature vector and the metabolism feature matrix.

[0138] A dynamic correlation modeling module 503 is configured to model a cross-scale correlation between microbial morphology dynamic changes and metabolism product concentration fluctuations based on the fusion feature matrix by using a nonlinear coupling analysis technique, and generate a dynamic correlation strength curve.

[0139] An abnormal fluctuation detection and labeling module 504 is configured to label abnormal metabolism fluctuations that are beyond a historical statistical range in the fermentation process, and form abnormal labeling data containing abnormal features and time stamps.

[0140] A model construction module 505 is configured to construct a cross-dimensional abnormality recognition model and a multi-modal collaborative prediction model based on the dynamic correlation strength curve and the abnormal labeling data.

[0141] A data fusion module 506 is configured to input the dynamic correlation strength curve and the abnormal labeling data into the cross-dimensional abnormality recognition model, output abnormal state labels and abnormal confidence, and fuse the abnormal state labels and the abnormal confidence with the fusion feature matrix to obtain multi-modal collaborative input data.

[0142] An intelligent decision generation module 507 is configured to input the multi-modal collaborative input data into the multi-modal collaborative prediction model, optimize and deduce the multi-modal collaborative input data by using a long short-term memory network algorithm, and generate a regulation strategy.

[0143] In an embodiment of the application, the multi-modal time sequence acquisition module 501 is further configured to:

[0144] perform image feature extraction processing on the microbial image data to generate the morphology feature vector;

[0145] perform metabolism feature dimension reduction processing on the metabolomics data to obtain the metabolism feature matrix.

[0146] In an embodiment of the application, the fusion feature matrix generation module 502 is further configured to:

[0147] perform time sequence sliding window decomposition on the morphology feature vector to obtain a morphology feature subspace sequence;

[0148] projecting the metabolic feature matrix to a metabolic feature subspace to obtain a metabolic feature subspace sequence;

[0149] calculating an optimal matching path of the morphological feature subspace sequence and the metabolic feature subspace sequence by using a dynamic time warping algorithm, and determining a time offset based on a time displacement parameter of the optimal matching path;

[0150] aligning and compensating the morphological feature subspace sequence and the metabolic feature subspace sequence on a time axis according to the time offset to generate a synchronized feature subspace sequence;

[0151] performing multi-scale nonlinear coupling analysis on the synchronized feature subspace sequence to obtain a fusion feature matrix.

[0152] In an embodiment of the present application, the dynamic correlation modeling module 503 is further configured to:

[0153] performing separation processing on the fusion feature matrix to obtain a morphological feature subspace and a metabolic feature subspace;

[0154] calculating a time offset of the morphological feature subspace and the metabolic feature subspace by using a dynamic time warping algorithm, and generating a synchronized feature subspace sequence based on the time offset;

[0155] performing cross-scale nonlinear correlation analysis on a dynamic evolution mode of the morphological feature subspace and a concentration fluctuation mode of the metabolic feature subspace by using a cross-scale correlation algorithm to generate a dynamic correlation strength sequence;

[0156] generating a dynamic correlation strength curve by using a weighted integral algorithm based on a time series distribution feature of the dynamic correlation strength sequence.

[0157] In an embodiment of the present application, the device further comprises:

[0158] a data acquisition module configured to acquire morphological change data and metabolic response data of a microorganism after regulation and control;

[0159] a feature dimension alignment module configured to perform feature dimension alignment processing on the morphological change data and the metabolic response data to generate a feedback feature set;

[0160] a model parameter updating module configured to construct a loss function based on the feedback feature set, and update network parameters of the multi-modal collaborative prediction model by using a gradient descent algorithm based on the loss function to obtain an updated multi-modal collaborative prediction model and optimized network parameters;

[0161] a feature weight optimization module configured to reevaluate contribution weights of the morphological feature vector and the metabolic feature matrix in the fusion feature matrix based on the optimized network parameters to obtain optimized feature fusion weights.

[0162] An initial parameter set generation module is configured to combine the optimized fusion feature matrix, the updated multi-modal collaborative prediction model and the optimized feature fusion weight to obtain an initial parameter set for next intelligent regulation and control.

[0163] In an embodiment of the present application, the multi-modal collaborative prediction model comprises an attention mechanism decoder and a metabolic dynamics encoder, wherein:

[0164] The attention mechanism decoder is configured to dynamically weight the intermediate features to generate a first candidate regulation and control correlation path.

[0165] The metabolic dynamics encoder is configured to generate a second candidate regulation and control correlation path based on a metabolomics feature correlation graph and a fermentation process dynamics model.

[0166] The first candidate regulation and control correlation path and the second candidate regulation and control correlation path are fused by a confusion network algorithm to generate an optimal regulation strategy.

[0167] It should be understood that, although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0168] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the power supply safety management method as described above when executing the computer program.

[0169] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the above method embodiments.

[0170] For the device embodiment, since it basically corresponds to the method embodiment, the relevant part can be seen from the part of the method embodiment. The device embodiment described above is only schematic, wherein the components shown as separate components can or can not be physically separate, and the components shown as a unit can or can not be a physical unit, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present disclosure. Those skilled in the art can understand and implement it without creative labor.

[0171] The above-described embodiments only express several implementation manners of the present application, which are described in detail, but cannot be understood as a limitation on the patent scope of the application. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application.

Claims

1. A deep learning-based intelligent decision-making method for fermentation parameters, characterized in that, The method comprises: Collecting time series data of two modes in a microbial fermentation process, performing multi-dimensional feature analysis on the time series data to obtain a morphology feature vector representing microbial morphology features and a metabolic feature matrix representing metabolic product distribution rules; Based on the time dimension matching rule of the morphology feature vector and the metabolic feature matrix, a fusion feature matrix is obtained; Based on the fusion feature matrix, a cross-scale correlation between microbial morphology dynamic change and metabolic product concentration fluctuation is modeled by a nonlinear coupling analysis technique to generate a dynamic correlation strength curve; Abnormal labeling is performed on metabolic fluctuations that exceed the historical statistical range during fermentation, forming abnormal labeling data containing abnormal features and time stamps; Based on the dynamic correlation strength curve and the abnormal labeling data, a cross-dimensional anomaly recognition model and a multi-modal collaborative prediction model are constructed; The dynamic correlation strength curve and the abnormal labeling data are input into the cross-dimensional anomaly recognition model to output an abnormal state label and an abnormal confidence, and the abnormal state label and the abnormal confidence are fused with the fusion feature matrix to obtain multi-modal collaborative input data; The multi-modal collaborative input data is input into a multi-modal collaborative prediction model, and the multi-modal collaborative input data is optimized and deduced by a long short-term memory network algorithm to generate a regulation strategy.

2. The method of claim 1, wherein, The multi-dimensional feature analysis on the time series data to obtain a morphology feature vector representing microbial morphology features and a metabolic feature matrix representing metabolic product distribution rules comprises: The time series data includes microbial image data and metabolomics data; The microbial image data is subjected to image feature extraction processing to generate the morphology feature vector; The metabolomics data is subjected to metabolic feature dimension reduction processing to obtain a metabolic feature matrix.

3. The method of claim 1, wherein, The fusion feature matrix is obtained based on the time dimension matching rule of the morphology feature vector and the metabolic feature matrix, comprising: The morphology feature vector is subjected to time series sliding window decomposition to obtain a morphology feature subspace sequence; The metabolic feature matrix is subjected to metabolic feature subspace projection to obtain a metabolic feature subspace sequence; The optimal matching path of the morphology feature subspace sequence and the metabolic feature subspace sequence is calculated by a dynamic time warping algorithm, and the time offset is determined based on the time displacement parameter of the optimal matching path; The morphology feature subspace sequence and the metabolic feature subspace sequence are subjected to time axis alignment compensation processing according to the time offset to generate a synchronized feature subspace sequence; The synchronized feature subspace sequence is subjected to multi-scale nonlinear coupling analysis to obtain a fusion feature matrix.

4. The method of claim 1, wherein, The cross-scale correlation between microbial morphology dynamic change and metabolic product concentration fluctuation is modeled by a nonlinear coupling analysis technique based on the fusion feature matrix to generate a dynamic correlation strength curve, comprising: The fusion feature matrix is subjected to separation processing to obtain a morphology feature subspace and a metabolic feature subspace; calculating time offsets of the morphological feature subspace and the metabolic feature subspace by a dynamic time warping algorithm, and generating a synchronized feature subspace sequence based on the time offsets; performing cross-scale nonlinear correlation analysis on the dynamic evolution mode of the morphological feature subspace and the concentration fluctuation mode of the metabolic feature subspace by a cross-scale correlation algorithm, and generating a dynamic correlation strength sequence; generating the dynamic correlation strength curve by a weighted integral algorithm based on the time sequence distribution characteristics of the dynamic correlation strength sequence.

5. The method of claim 1, wherein, The method comprises the following steps: inputting the multi-modal collaborative input data into a multi-modal collaborative prediction model, and optimizing and deducing the multi-modal collaborative input data by a long short-term memory network algorithm to generate a regulation strategy, which comprises: inputting the multi-modal collaborative input data into a multi-modal collaborative prediction model, and performing multi-modal time sequence feature collaborative prediction processing on the multi-modal collaborative input data by a long short-term memory network algorithm to generate prediction trend data of the fermentation state trend; 6. The method of claim 1, wherein, based on the direction and amplitude of the change of the metabolic product concentration in the prediction trend data, performing multi-objective parameter optimization and deduction on the temperature regulation amount, the pH correction amount and the feeding strategy to generate a regulation strategy of the fermentation process parameters. After the multi-modal collaborative prediction model and the cross-dimensional anomaly recognition model are constructed based on the dynamic correlation strength curve and the abnormal marker data, and the regulation strategy is determined, the method further comprises the following steps: collecting microbial morphological change data and metabolic response data of the biological fermentation after regulation; performing feature dimension alignment processing on the microbial morphological change data and the metabolic response data to generate a feedback feature set; constructing a loss function according to the feedback feature set, the loss function being used to update network parameters of the multi-modal collaborative prediction model by a gradient descent algorithm to obtain an updated multi-modal collaborative prediction model and optimized network parameters; re-evaluating the contribution weights of the morphological feature vector and the metabolic feature matrix in the fusion feature matrix according to the optimized network parameters to obtain optimized feature fusion weights; 7. The method of claim 1, wherein, combining the optimized fusion feature matrix, the updated multi-modal collaborative prediction model and the optimized feature fusion weights to obtain an initial parameter set for next intelligent regulation. The multi-modal collaborative prediction model comprises an attention mechanism decoder and a metabolic kinetics encoder, wherein: the attention mechanism decoder is used for dynamically weighting intermediate features to generate a first candidate regulation association path; the metabolic kinetics encoder is used for generating a second candidate regulation association path based on a metabolomics feature association graph and a fermentation process kinetics model; 8. A deep learning-based intelligent decision device for fermentation parameters, characterized in that, the first candidate regulation association path and the second candidate regulation association path are fused by a confusion network algorithm to generate an optimal regulation strategy. The device comprises: a multi-modal time sequence acquisition module, which is used for acquiring time sequence data of two modalities in a microbial fermentation process, performing multi-dimensional feature analysis on the time sequence data, and obtaining a morphological feature vector representing microbial morphological features and a metabolic feature matrix representing metabolic product distribution rules; a fusion feature matrix generation module configured to obtain a fusion feature matrix based on a time dimension matching rule between the morphological feature vector and the metabolic feature matrix; a dynamic correlation modeling module configured to model a cross-scale correlation between dynamic changes in microbial morphology and fluctuations in metabolic product concentration based on the fusion feature matrix by using a nonlinear coupling analysis technique, and generate a dynamic correlation strength curve; an abnormal fluctuation detection and labeling module configured to implement abnormal labeling on metabolic fluctuations that exceed a historical statistical range in the fermentation process, and form abnormal labeling data containing abnormal features and time stamps; a model construction module configured to construct a cross-dimensional abnormality recognition model and a multi-modal collaborative prediction model based on the dynamic correlation strength curve and the abnormal labeling data; a data fusion module configured to input the dynamic correlation strength curve and the abnormal labeling data into the cross-dimensional abnormality recognition model, output an abnormal state label and an abnormal confidence, and fuse the abnormal state label and the abnormal confidence with the fusion feature matrix to obtain multi-modal collaborative input data; an intelligent decision generation module configured to input the multi-modal collaborative input data into a multi-modal collaborative prediction model, optimize and deduce the multi-modal collaborative input data by using a long short-term memory network algorithm, and generate a regulation strategy. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images