Pomegranate leaf fermentation process optimization and intelligent control system based on digital twinning

Abstract

The invention relates to the technical field of biological process intelligent control, and discloses a pomegranate leaf fermentation process optimization and intelligent control system based on digital twinning. According to the system, the Internet of Things and a visual sensor are integrated through a multi-source data sensing module, so that synchronous acquisition and fusion of temperature, pH, dissolved oxygen, flora density and image data are realized; the intelligent data preprocessing module extracts high-quality features by adopting a self-adaptive filtering technology; the digital twin simulation engine module drives the virtual model to dynamically update based on real-time data; the deep prediction module accurately predicts the fermentation trend through a deep learning algorithm; the intelligent diagnosis module realizes abnormal root cause analysis by using a knowledge graph; the reinforcement learning optimization module autonomously generates an optimal control strategy based on the diagnosis result; and the self-adaptive execution module accurately regulates and controls fermentation parameters. According to the system, multi-dimensional perception, high-precision prediction and adaptive optimization control of the fermentation process are realized, and the accuracy and stability of process control are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of biological processes, in particular to a pomegranate leaf fermentation process optimization and intelligent control system based on digital twinning. BACKGROUND

[0002] The existing technology of pomegranate leaf fermentation mainly uses discrete sensors to independently monitor a single parameter, and adjusts the fermentation conditions according to pre-set fixed parameters or simple threshold rules through programmable logic controllers or conventional PID control algorithms. Visual monitoring means is usually independent of the physical parameter control system, resulting in data isolation. Process modeling mostly uses static mechanism models or offline statistical models based on ideal conditions for rough prediction of fermentation trends. Abnormal diagnosis generally relies on single variable overrun alarm, and optimization control highly depends on the experience of operators.

[0003] These technical solutions have limitations: there is a lack of effective fusion mechanism among multi-source heterogeneous data, resulting in fragmented cognition of the fermentation state, making it difficult to grasp the dynamic correlation of microbial growth and metabolism as a whole; static mathematical models cannot accurately describe the nonlinear and time-varying characteristics of the fermentation process, resulting in significant deviation between simulation prediction and actual process, and model updating lag. When the process is disturbed or abnormal, the traditional control strategy has weak self-adaptability, slow response, and is difficult to conduct in-depth root cause analysis and forward-looking optimization, which easily leads to unstable fermentation process and large fluctuations in product yield.

[0004] The present application needs to solve the problem of how to break through the limitations of traditional single data source and static model, realize the deep fusion of multi-modal data and real-time driven high-fidelity process simulation, and accurately capture the micro dynamic evolution of the fermentation process. Further need to solve how to realize the autonomous and intelligent optimization of control strategy on the basis of accurate perception and prediction, to dynamically adapt to the nonlinear changes of the process, replacing the control mode relying on fixed rules and artificial experience. SUMMARY

[0005] The present application aims to provide a pomegranate leaf fermentation process optimization and intelligent control system based on digital twinning to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides a pomegranate leaf fermentation process optimization and intelligent control system based on digital twinning, which comprises: A multi-source data perception module is used to collect multi-modal data of the fermentation tank through Internet of Things sensors and visual sensors, including temperature, pH, dissolved oxygen, bacterial population density and image data, and generate original multi-modal data stream after data fusion; An intelligent data preprocessing module is used to receive the original multi-modal data stream, perform adaptive filtering and feature extraction, and generate a preprocessed feature data set; The digital twin simulation engine module is used to receive the preprocessed feature dataset, interact with the digital twin model in real time, update the virtual fermentation tank status, and generate simulation status data. The deep prediction module is used to receive the simulation state data, apply deep learning methods to predict fermentation kinetic parameters, and generate predicted kinetic data. The intelligent diagnostic module is used to receive the predicted dynamics data and the preprocessed feature dataset, perform anomaly root cause analysis based on the knowledge graph, and generate diagnostic result signals. The reinforcement learning optimization module is used to receive the diagnostic result signal, use a reinforcement learning agent to optimize the control parameters, and generate an optimized control strategy. The adaptive execution module receives the optimized control strategy, adjusts the fermentation conditions through the actuator, and generates a sequence of control actions.

[0007] Preferably, the original multimodal data stream includes temperature data sequence, pH data sequence, dissolved oxygen data sequence, bacterial community density data sequence, and image data sequence; the preprocessed feature dataset includes time-domain features, frequency-domain features, and spatial features; the simulation state data includes virtual temperature, virtual pH, virtual dissolved oxygen, and virtual bacterial community density; the predicted kinetic data includes growth rate prediction and product concentration prediction; the diagnostic result signal includes anomaly type identifier and root cause probability; the optimized control strategy includes action value and strategy gradient; and the control action sequence includes actuator command and timestamp.

[0008] Preferably, the multi-source data sensing module includes: The sensor adapter submodule is used to connect heterogeneous sensor networks, perform protocol conversion and data verification, and generate standardized sensor data. The multimodal fusion submodule is used to receive the standardized sensor data, apply the Kalman filter algorithm to perform data fusion, and generate fused multimodal data; The data stream management submodule is used to receive the fused multimodal data, cache and segment it according to time windows, and generate the original multimodal data stream.

[0009] Preferably, the intelligent data preprocessing module includes: The noise removal submodule is used to receive the original multimodal data stream, apply wavelet transform algorithm to remove noise, and generate clean data; The feature extraction submodule is used to receive the clean data, extract time-domain and frequency-domain features using a convolutional neural network, and generate a feature vector set; The data normalization submodule is used to receive the feature vector set, perform min-max normalization processing, and generate a preprocessed feature dataset.

[0010] Preferably, the digital twin simulation engine module includes: The model loading submodule is used to load digital twin models, which include physical equations and machine learning components; The state update submodule is used to receive the preprocessed feature dataset as input to drive the state update of the digital twin model and generate the current virtual state; The data output submodule is used to extract state variables from the digital twin model and generate simulation state data.

[0011] Preferably, the depth prediction module includes: The sequence modeling submodule is used to receive the simulation state data, model the time series dependencies using a long short-term memory network, and generate a hidden state sequence. The prediction output submodule is used to receive the hidden state sequence, output predicted values ​​through a fully connected neural network layer, and generate prediction dynamics data.

[0012] Preferably, the intelligent diagnostic module includes: The knowledge graph query submodule is used to access the domain knowledge graph and query historical patterns of fermentation anomalies; The causal reasoning submodule is used to receive the prediction dynamics data and the preprocessed feature dataset, combine them with the knowledge graph to perform causal path analysis, and generate a root cause hypothesis set. The diagnostic decision submodule is used to evaluate the posterior probability of the root cause hypothesis set based on Bayesian networks and generate diagnostic result signals.

[0013] Preferably, the reinforcement learning optimization module includes: The state representation submodule is used to receive the diagnostic result signal and compress the high-dimensional state information into a low-dimensional state representation vector through an autoencoder network. The multi-agent collaboration submodule is used to deploy multiple dedicated agents, which independently learn strategies for ventilation rate, stirring rate and temperature parameters, and generate an initial action suggestion set. The strategy fusion submodule is used to receive the initial action suggestion set, calculate the weight coefficient of each agent's suggestion through the attention mechanism, and perform weighted fusion to generate a collaborative control strategy; The strategy optimization submodule is used to receive the cooperative control strategy, perform policy gradient updates based on the near-end policy optimization algorithm, and generate an optimized control strategy.

[0014] Preferably, the adaptive execution module includes: The control translation submodule is used to receive the optimized control strategy and convert it into an actuator-specific command language; The execution driver submodule is used to send commands to the actuators to drive the fermenter equipment and generate control action sequences.

[0015] Preferably, the system further includes a model update module, used to retrain the digital twin model and the neural network model in the deep prediction module based on the real-time acquired data stream, and update the model parameters through an incremental learning algorithm.

[0016] Compared with the prior art, the beneficial effects of the present invention are: The multi-source data sensing module integrates IoT and visual sensors to simultaneously collect and fuse physicochemical parameters and microbial community image data, generating a unified multimodal data stream. The intelligent data preprocessing module employs adaptive filtering technology to dynamically reduce noise and extract features based on the data characteristics of each fermentation stage, outputting a high-quality feature dataset. This dataset is input in real-time into the digital twin simulation engine module, driving the virtual fermenter model to synchronously update its state and generate high-fidelity simulation state data. The deep prediction module applies deep learning algorithms to process this time-series simulation data, accurately capturing the complex nonlinear mapping relationships between fermentation kinetic parameters. This continuous data stream and model interaction, from multi-source sensing to high-precision prediction, transforms the description of the fermentation process from a static and local perspective to a dynamic and systemic perspective, enhancing the depth of perception of process state changes and the accuracy of trend prediction.

[0017] The intelligent diagnostic module, based on a constructed knowledge graph, performs multi-dimensional correlation analysis between predicted data and real-time features to trace and locate the root cause of anomalies. The reinforcement learning optimization module receives diagnostic signals and, supported by a digital twin environment, explores the optimal control action space through agent interaction and dynamically outputs optimized control strategies. The adaptive execution module transforms this strategy into a precise sequence of control actions, driving the actuators to adjust physical fermentation conditions. This closed-loop decision-making mechanism, from intelligent diagnosis to autonomous execution, shifts the basis of control decision-making from preset rules and human experience to a dynamic optimization process based on deep causal analysis and online self-learning, enabling the system to proactively adapt and learn to cope with process uncertainties and nonlinear changes. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the pomegranate leaf fermentation process optimization and intelligent control system based on digital twins as described in this invention. Figure 2 A flowchart illustrating the operation of the multi-source data sensing module; Figure 3 A flowchart illustrating how the digital twin simulation engine module works; Figure 4 A diagram for digital twin status monitoring and predictive analysis; Figure 5 This is a diagram for intelligent diagnosis and optimized control analysis. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides an optimization and intelligent control system for pomegranate leaf fermentation based on digital twins. The system includes: a multi-source data sensing module connecting IoT sensors and visual sensors to collect multimodal data such as temperature, pH, dissolved oxygen, bacterial density, and images within the fermenter. This raw data undergoes data fusion processing to generate a continuous multimodal data stream. An intelligent data preprocessing module receives this data stream and performs adaptive filtering and feature extraction to remove noise and generate a high-quality feature dataset. A digital twin simulation engine module uses the preprocessed data to drive a digital twin model, updating the state variables of the virtual fermenter in real time and outputting simulation state data to reflect the dynamic changes of the physical process. A deep prediction module applies deep learning algorithms based on the simulation state data to predict fermentation kinetic parameters such as growth rate and product concentration, generating predicted data for analysis. An intelligent diagnosis module combines the predicted data and feature dataset, performs anomaly root cause analysis using a knowledge graph, and outputs diagnostic result signals identifying the anomaly type and probability. A reinforcement learning optimization module uses reinforcement learning agents to optimize control parameters based on the diagnostic signals, generating optimization strategies for aeration rate, stirring rate, etc. The adaptive execution module converts the optimization strategy into actuator commands, adjusts the fermentation conditions, and generates a sequence of control actions to complete closed-loop control.

[0021] Example 1: In specific implementation, the original multimodal data stream is generated by a multi-source data sensing module, including temperature data sequence, pH data sequence, dissolved oxygen data sequence, microbial density data sequence, and image data sequence. The temperature data sequence is collected by temperature sensor nodes distributed in multiple locations inside the fermenter. Each temperature sensor node uses a digital thermocouple, with a measurement range covering 0 to 100 degrees Celsius. The sampling frequency is set to once per second, and the data is transmitted in digital signal form via an RS-485 bus to reduce errors caused by analog interference. The pH data sequence is acquired using a composite glass electrode sensor equipped with an automatic temperature compensation function to cope with temperature fluctuations in the fermentation broth, achieving a measurement accuracy of 0.01 pH units. The dissolved oxygen data sequence is measured by a dissolved oxygen sensor based on the optical fluorescence quenching principle. The sensor has a built-in pressure compensation algorithm to adapt to changes in pressure inside the tank. The microbial density data sequence is non-invasively collected by a near-infrared spectral sensor, and the biomass concentration is calculated using spectral absorption characteristics. The image data sequence is captured by a high-resolution industrial camera equipped with a polarizing filter to eliminate reflections, recording the surface morphology of the fermentation material at a rate of five frames per second. In practical implementation, the data sequences of the original multimodal data stream are aligned through a time synchronization module. A high-precision clock source is used for timestamps to ensure that temperature, pH, dissolved oxygen, microbial density, and image data sequences are integrated under the same time reference. The data sequences are buffered in a circular memory queue, with the queue depth configured according to the fermentation process duration to prevent data overflow. In some embodiments, sensor identifiers are appended to the data points of the temperature data sequence to trace the data source; the pH data sequence undergoes preliminary filtering to remove spike noise; the dissolved oxygen data sequence is calibrated to a standard dissolved oxygen solution; the microbial density data sequence is converted to optical density units; and the image data sequence is compressed into JPEG2000 format to reduce storage space. It can be understood that the generation of the original multimodal data stream is a continuous process. The data stream management submodule segments the data stream according to fixed time windows, with each window length set to five minutes, corresponding to the typical time constant of fermentation kinetics. The segmented data stream is packaged into data packets and transmitted to the intelligent data preprocessing module via Ethernet.

[0022] The preprocessed feature dataset is generated by the intelligent data preprocessing module after processing the original multimodal data stream. It includes temporal, frequency, and spatial features. Temporal features are extracted from temperature, pH, dissolved oxygen, and bacterial density data sequences, calculating metrics such as moving average, standard deviation, and zero-crossing rate. The moving average is calculated using a 30-second sliding window to reflect short-term trends, the standard deviation characterizes volatility, and the zero-crossing rate indicates signal oscillation frequency. Frequency domain features are converted from temporal data to the frequency domain using a Fast Fourier Transform (FFT) algorithm, extracting key frequency components and power spectral density. Spatial features are extracted from image data sequences using a convolutional neural network model. The network structure includes convolutional and pooling layers, outputting texture features such as gray-level co-occurrence matrix statistics. In practice, the temporal feature extraction process integrates a real-time computing engine to avoid data latency. Frequency domain feature analysis employs piecewise Fourier transform to handle non-stationary signals. Spatial feature extraction involves preprocessing image data, including grayscale conversion and size normalization. Optionally, time-domain features are supplemented with autocorrelation function calculations, frequency-domain features include wavelet packet decomposition, and spatial features incorporate shape descriptors. In some embodiments, the preprocessed feature dataset has a multi-dimensional vector structure. The vector dimension is reduced through principal component analysis to decrease computational load, and the feature values ​​are normalized to zero mean and unit variance to enhance model training stability. It can be understood that the preprocessed feature dataset serves as input to the digital twin simulation engine module, driving virtual model updates.

[0023] The simulation state data is output by the digital twin simulation engine module, including virtual temperature, virtual pH, virtual dissolved oxygen, and virtual microbial density. Virtual temperature is obtained by solving the partial differential equation for heat conduction, based on the fermenter's geometry and material properties. After discretization, the equation is calculated iteratively using the finite difference method. Virtual pH is based on an electrochemical equilibrium model, considering hydrogen ion activity and buffer concentration. Virtual dissolved oxygen simulates the gas-liquid two-phase mass transfer process, applying the two-film theory to calculate the oxygen transfer rate. Virtual microbial density is based on the Monod growth kinetics equation, integrating substrate consumption and product inhibition factors. In practice, the computational domain for virtual temperature is divided into grid cells, each assigned a thermal conductivity parameter. The model parameters for virtual pH are calibrated online using real-time pH data sequence feedback. The mass transfer coefficient of virtual dissolved oxygen is correlated with the stirring rate, and the specific growth rate of virtual microbial density is updated as a state variable. Optionally, the output frequency of the simulation state data matches the sensor sampling rate, the virtual temperature field is visualized to display the heat distribution, the virtual pH value is output as a scalar, the virtual dissolved oxygen concentration is mapped to three-dimensional space, and the virtual microbial density considers population dynamics. It is understandable that simulation state data is used as input to the depth prediction module to provide the virtual environment state.

[0024] The predicted kinetic data is generated by a deep prediction module, including growth rate prediction and product concentration prediction. Growth rate prediction uses a long short-term memory (LSTM) network model to analyze the time series of simulated state data. The network input includes historical window data of virtual temperature, virtual pH, virtual dissolved oxygen, and virtual microbial density, and the output is an estimate of the growth rate at future time points. Product concentration prediction is achieved through a fully connected neural network. Input features include the predicted growth rate and substrate concentration, and the output is the concentration of products such as enzymes or metabolites. In specific implementations, the LTM network for growth rate prediction contains three hidden layers, each with 128 units, and training uses time-series cross-validation. The network structure for product concentration prediction includes a batch normalization layer to prevent overfitting. In some embodiments, confidence intervals are appended to the predicted kinetic data, and uncertainty is estimated using the Monte Carlo Dropout method. The unit for growth rate prediction is per hour, and the unit for product concentration prediction is converted to mass concentration. It can be understood that the predicted kinetic data supports the decision-making process of the intelligent diagnostic module.

[0025] The diagnostic result signal is generated by the intelligent diagnostic module, including anomaly type identification and root cause probability. The anomaly type identification is based on a rule engine matching prediction kinetic data and a preprocessed feature dataset. Identification types include temperature anomalies, pH shifts, or pollution events. The root cause probability is calculated using a Bayesian network, where network nodes represent possible causes such as sensor malfunction, operational errors, or environmental interference. The probability values ​​are obtained through training on historical data. In practice, the anomaly type identification output is a classification label, and the root cause probability is normalized to the range of 0 to 1, summing to 1. It can be understood that the diagnostic result signal triggers action optimization in the reinforcement learning optimization module.

[0026] The optimized control strategy is generated by a reinforcement learning optimization module, including action values ​​and policy gradients. Action values ​​are continuous values ​​representing actuator settings such as vent valve opening, stirrer motor speed, or heater power. The policy gradient indicates the optimization direction, guiding the adjustment of action values ​​to maximize the reward function. In practice, action values ​​are learned through a deep deterministic policy gradient algorithm, and the policy gradient is updated using a policy optimization method. Essentially, the optimized control strategy is translated into actuator commands. The control action sequence is output by the adaptive execution module, including actuator commands and timestamps. Actuator commands are specific control instructions such as analog voltage signals or digital pulses, and timestamps record the command issuance time to ensure timing accuracy. In practice, actuator commands are sent via a fieldbus, and timestamps are synchronized with the system clock.

[0027] Example 2: See Figure 2In practical implementation, the multi-source data sensing module includes a sensor adaptation submodule, a multimodal fusion submodule, and a data flow management submodule. The sensor adaptation submodule connects to heterogeneous sensor networks, performs protocol conversion and data verification, and generates standardized sensor data. Heterogeneous sensor networks contain various types of sensor devices, such as temperature sensors, pH sensors, dissolved oxygen sensors, microbial density sensors, and vision sensors. These sensors may use different communication protocols. The sensor adaptation submodule implements protocol conversion through a hardware interface board and software driver, converting raw sensor data into unified JSON format data packets. Each data packet contains a sensor identifier, numerical reading, timestamp, and checksum. Data verification employs a cyclic redundancy check algorithm to ensure the integrity and accuracy of data transmission and prevent data errors caused by noise or interference. In practical implementation, the hardware interface board of the sensor adaptation submodule integrates multiple analog-to-digital converters and digital input / output ports, supporting simultaneous connection of multiple sensors. The software driver implements protocol parsing logic, such as parsing register addresses and data values ​​in Modbus frames and mapping them to standard fields. The standardized sensor data is then sent to the multimodal fusion submodule for further processing.

[0028] The multimodal fusion submodule receives standardized sensor data from the sensor adaptation submodule and applies a Kalman filter algorithm to fuse the data, generating fused multimodal data. The Kalman filter algorithm is based on a state-space model and includes prediction and update steps. The prediction step uses the system model to estimate the current state, while the update step combines sensor observations to correct the estimate and reduce uncertainty. In the multimodal fusion submodule, state variables may include the actual values ​​of temperature, pH, dissolved oxygen, and bacterial density. Observation variables come from standardized sensor data. The parameters of the Kalman filter algorithm, such as the process noise covariance matrix and the observation noise covariance matrix, are pre-calibrated according to sensor characteristics, for example, adjusted for the high accuracy of the temperature sensor and the delay characteristics of the pH sensor. In practice, the multimodal fusion submodule is implemented as a software module running on an embedded processor. The Kalman filter algorithm executes iteratively, updating the state each time new standardized sensor data is received. The fused multimodal data output is a comprehensive data structure containing fused temperature, pH, dissolved oxygen, bacterial density, and image feature summaries. These data are time-aligned, and interpolation methods are used to process sensor data with different sampling rates. In some embodiments, the multimodal fusion submodule is extended to use an unscented Kalman filter algorithm to handle nonlinear system dynamics and improve fusion accuracy. It can be understood that the multimodal fusion submodule improves measurement reliability and reduces the impact of individual sensor errors through data fusion.

[0029] The data stream management submodule receives fused multimodal data from the multimodal fusion submodule, caches and segments it according to time windows, and generates the original multimodal data stream. The size of the time window is configurable, typically set to 5 to 10 minutes to accommodate the slow dynamic characteristics of the fermentation process. The caching mechanism uses a circular buffer data structure, with the buffer size calculated based on the data rate and window length to prevent memory overflow. Segmentation is triggered based on a timer or data volume threshold. For example, when the buffer data volume reaches a preset size, the current window data is packaged into data segments, and segment identifiers and start and end timestamps are added to form the original multimodal data stream. In specific implementations, the data stream management submodule is implemented as a multithreaded application. One thread is responsible for data reception and caching, while another thread handles segmentation and transmission. The original multimodal data stream is sent to the intelligent data preprocessing module via TCP / IP protocol to ensure real-time performance. In some embodiments, the data stream management submodule supports dynamically adjusting the time window size, adaptively changing based on the data stream rate, for example, shrinking the window to reduce latency at high data rates. It can be understood that the role of the data stream management submodule is to manage the timing and storage of the data stream, providing structured input for subsequent modules.

[0030] The intelligent data preprocessing module includes a noise removal submodule, a feature extraction submodule, and a data normalization submodule. The noise removal submodule receives the raw multimodal data stream from the data stream management submodule, applies a wavelet transform algorithm to remove noise, and generates clean data. The wavelet transform algorithm selects appropriate wavelet basis functions, performs multi-resolution analysis, decomposes the signal into approximation coefficients and detail coefficients, suppresses noise by thresholding the detail coefficients, and then reconstructs the signal. The noise removal submodule applies specific parameters for different types of data; for example, it uses a soft threshold for temperature data sequences and a hard threshold for image data sequences. The clean data retains the main features of the signal while eliminating high-frequency noise. In specific implementations, the noise removal submodule is implemented by calling a digital signal processing library. The wavelet transform algorithm processes the streaming data using a sliding window approach, with the window size coordinated with the time window of the data stream management submodule to ensure real-time processing. Optionally, the noise removal submodule integrates alternative denoising algorithms, such as median filtering or Kalman filtering, for users to choose according to their scenario. In essence, the noise removal submodule improves data quality, preparing for feature extraction.

[0031] The feature extraction submodule receives cleaned data from the noise removal submodule and uses a convolutional neural network (CNN) to extract temporal and frequency domain features, generating a feature vector set. The CNN structure includes an input layer, convolutional layers, pooling layers, and fully connected layers. The convolutional layers use multiple kernels to scan the input data, extracting local patterns, such as trend changes from temperature data sequences or edge textures from image data sequences. Pooling layers reduce data dimensionality, and fully connected layers output feature vectors. Temporal features include mean, variance, and autocorrelation function. Frequency domain features are converted from time-domain data to the frequency domain using a fast Fourier transform (FFT), calculating spectral features such as center frequency and bandwidth. In practice, the CNN in the feature extraction submodule is pre-trained on a historical fermentation dataset. Network parameters such as kernel size and stride are optimized to capture specific fermentation patterns. The feature vector set is organized as a matrix, with each row corresponding to a time point and each column corresponding to a feature dimension. Optionally, the feature extraction submodule supports other feature extraction methods, such as principal component analysis or autoencoders, for dimensionality reduction. Essentially, the feature extraction submodule transforms raw data into meaningful features, facilitating model processing.

[0032] The data normalization submodule receives the feature vector set from the feature extraction submodule, performs min-max normalization, and generates a preprocessed feature dataset. Min-max normalization linearly scales each feature value to the [0,1] interval. The scaling formula is based on the minimum and maximum values ​​of the training data. For example, for each feature dimension, the global minimum and maximum values ​​are calculated, and then the normalization formula is applied. The preprocessed feature dataset ensures that different feature scales are consistent, avoiding biases caused by differences in numerical ranges during model training. In practice, the data normalization submodule calculates the normalization parameters in real time, uses a sliding window to statistically analyze the minimum and maximum values ​​to adapt to changes in data distribution, and outputs the preprocessed feature dataset as a standardized tensor, which is directly input into the digital twin simulation engine module.

[0033] Example 3: See Figure 3In its implementation, the digital twin simulation engine module includes a model loading submodule, a state update submodule, and a data output submodule. The model loading submodule loads the digital twin model, which contains physical equations and machine learning components. The physical equations are constructed based on the laws of conservation of mass and energy, describing the heat and mass transfer processes within the fermenter. For example, the heat balance equation considers heat input from stirring and heating, and heat output through heat dissipation via the tank wall. The mass balance equation tracks the substrate consumption and product generation rates. The machine learning components use a feedforward neural network structure with multiple hidden layers and the sigmoid function as the activation function. This is used to approximate nonlinear relationships in the physical equations that are difficult to model analytically, such as the complex mapping between microbial growth rate and environmental parameters. The model loading submodule reads model parameters from the local file system or a remote database. These parameters include coefficients of the physical equations, such as the heat transfer coefficient and reaction rate constant, as well as the weight matrix and bias vector of the neural network. When initializing the simulation environment, the module allocates memory space and sets initial state variable values, such as initial temperature and pH. In practical implementation, the model loading submodule is implemented as a dynamic link library, supporting model version management and allowing switching between different versions of the digital twin model at runtime. The digital twin model is stored in XML or JSON format, containing the model structure and parameter definitions. The initialization process verifies the integrity of parameters to prevent invalid values ​​from causing simulation errors. In some embodiments, the model loading submodule supports loading models from cloud storage for centralized management. The physical equations of the digital twin model are described using a system of ordinary differential equations, such as the fermenter temperature change equation combined with energy conservation. The neural network of the machine learning component is trained using historical fermentation data to compensate for model errors.

[0034] The state update submodule receives the preprocessed feature dataset from the intelligent data preprocessing module as input to drive the state update of the digital twin model, generating the current virtual state. The input data is mapped to the boundary conditions and external inputs of the digital twin model. For example, temperature feature values ​​in the preprocessed feature dataset are used as heat source terms, and pH feature values ​​are used as acidity / alkalinity inputs. The state update solves the differential equations in the digital twin model using numerical integration methods. The numerical integration method adopts the fourth-order Runge-Kutta method, which calculates the increment of state variables through multiple intermediate steps to improve accuracy. The current virtual state includes temperature field distribution, pH spatial variation, virtual dissolved oxygen concentration, and virtual bacterial community density field. The temperature field calculates the temperature value of each cell through a discrete grid. The pH distribution considers mixing inhomogeneity. The virtual dissolved oxygen concentration simulates oxygen diffusion, and the virtual bacterial community density is based on growth kinetics. In specific implementation, the state update submodule is implemented as a real-time computing thread with adaptively adjusted iteration step size based on system stability and real-time requirements. For example, the step size is reduced when the state changes drastically. The input data is aligned with the model time step through interpolation processing. The current virtual state is stored as a multi-dimensional array to support fast access. In some embodiments, the state update submodule integrates multiple numerical methods to suit different accuracy requirements, and the visualization of the current virtual state is used to monitor the simulation process. It can be understood that the core of the state update submodule is the numerical solution of the dynamic system, ensuring that the virtual state is synchronized with the physical process.

[0035] The data output submodule extracts state variables from the digital twin model to generate simulation state data. State variables include virtual temperature, virtual pH, virtual dissolved oxygen, and virtual bacterial density. These variables are read from the current virtual state. Virtual temperature is a scalar average or grid value, virtual pH is a time-series value, virtual dissolved oxygen is a concentration scalar, and virtual bacterial density is a biomass value. The simulation state data is formatted as a time-series structure, including timestamps and variable values. The timestamps are synchronized with sensor data, and variable values ​​are converted to engineering units, such as degrees Celsius for temperature and unitless for pH. In specific implementations, the data output submodule is implemented as a data extractor, periodically reading state variables from the digital twin model's memory. The sampling frequency is configurable, typically matching the sensor sampling rate. The simulation state data is published via a message queue for use by the deep prediction module. In some embodiments, the data output submodule supports data compression to reduce bandwidth, and metadata such as model version and confidence level is added to the simulation state data. It can be understood that the data output submodule bridges the simulation model and downstream modules, providing a consistent interface.

[0036] The deep prediction module comprises a sequence modeling submodule and a prediction output submodule. The sequence modeling submodule receives simulation state data from the digital twin simulation engine module, models time-series dependencies using a Long Short-Term Memory (LSTM) network, and generates a hidden state sequence. The LTM network structure includes an input gate, a forget gate, an output gate, and cell states. The input gate controls the inflow of new information, the forget gate determines the discarding of old information, the output gate regulates the output of hidden states, and cell states maintain long-term memory. The hidden state sequence encodes time-dynamic information, with each hidden state vector corresponding to an abstract representation of a time point. In implementation, the LTM network of the sequence modeling submodule is pre-trained on a historical fermentation dataset. Network parameters such as weights and biases are optimized using a time backpropagation algorithm. The hidden state sequence dimension is set to 128 dimensions to capture long-term trends. The update process of the sequence modeling submodule involves a gating mechanism, where all weights, biases, and input data are normalized dimensionless values. For example, the forget gate calculation is based on the current input and the previous hidden state, expressed by the formula:

[0037] in: It is the dimensionless output of the forget gate at time step t. It is the Sigmoid activation function. It is the dimensionless weight matrix of the forget gate. It is the dimensionless hidden state vector at time step t-1. It is the dimensionless input vector at time step t. It is the dimensionless bias vector of the forget gate. This represents a vector concatenation operation. Other gating mechanisms employ similar dimensionless computation methods, with the hidden state sequence generated through iterative computation. It can be understood that the sequence modeling submodule utilizes a long short-term memory network to process temporal data, overcoming the gradient vanishing problem of traditional recurrent neural networks.

[0038] The prediction output submodule receives the hidden state sequence from the sequence modeling submodule, outputs predicted values ​​through a fully connected neural network layer, and generates prediction dynamics data. This fully connected neural network layer contains multiple fully connected layers, using the ReLU activation function to avoid gradient saturation. The predicted values ​​include growth rate predictions and product concentration predictions. The growth rate prediction represents the specific growth rate of the microorganism, in hours, and the product concentration prediction represents the concentration of the target metabolite, in grams per liter. Uncertainty estimation is added to the prediction dynamics data using Dropout or Bayesian methods. In specific implementations, the input dimension of the fully connected neural network structure in the prediction output submodule matches the dimension of the hidden state sequence, the output layer is a linear layer, generating continuous values, and the prediction dynamics data is formatted as a structured data packet containing the predicted value, time range, and confidence interval. Optionally, the prediction output submodule integrates ensemble learning methods, such as multiple neural network voting, to improve robustness. In essence, the prediction output submodule maps the hidden state to specific predictions, supporting the decision-making process. The digital twin simulation engine module and the deep prediction module work together to achieve a closed loop from data to prediction. The digital twin simulation engine module provides the virtual environment state, and the deep prediction module makes forward-looking predictions based on the state.

[0039] See Figure 4 This study demonstrates the multi-parameter monitoring and prediction capabilities of a digital twin system during pomegranate leaf fermentation. The charts and graphs use multiple curves to illustrate the dynamic changes in virtual temperature, pH, dissolved oxygen, microbial density, predicted growth rate, and predicted product concentration. These data reflect the complex biochemical reactions within the fermenter; temperature fluctuations indicate heating control strategies; the pH curve reflects acid-base balance regulation; decreased dissolved oxygen reflects microbial oxygen consumption; increased microbial density demonstrates microbial proliferation dynamics; and the prediction curves provide forward-looking guidance for optimized control. The entire chart showcases the real-time monitoring and prediction capabilities of digital twin technology during fermentation.

[0040] Example 4: In specific implementation, the intelligent diagnosis module includes a knowledge graph query submodule, a causal reasoning submodule, and a diagnosis decision submodule. The knowledge graph query submodule is used to access the domain knowledge graph and query historical patterns of fermentation anomalies. The domain knowledge graph is built on a graph database and contains entity nodes and relation edges. Entity node types include environmental parameter entities, microbial strain entities, equipment component entities, and anomaly type entities. Relationship edge types include causal relationship edges, membership relationship edges, correlation relationship edges, and temporal relationship edges. The knowledge graph query submodule sends query requests to the graph database through the SPARQL query language. The query conditions are based on the feature values ​​in the currently received predictive kinetic data and the preprocessed feature dataset. For example, when the predicted growth rate value in the predictive kinetic data is lower than a certain threshold, the module queries historical anomaly patterns related to "growth inhibition." The query results return a set of matching historical cases, each case containing information on the anomaly manifestation, root cause, and treatment measures. In practice, the knowledge graph query submodule is implemented as an independent microservice. It interacts with the graph database via a RESTful API, encapsulates query requests in JSON format, including query conditions and return field constraints, and caches query results in an in-memory database to improve the response speed of subsequent similar queries. Essentially, the role of the knowledge graph query submodule is to quickly locate potential anomaly patterns using structured knowledge, providing background information for causal analysis.

[0041] The causal reasoning submodule receives prediction kinetic data and a preprocessed feature dataset. It then combines this data with historical fermentation anomaly patterns obtained from the knowledge graph query submodule to perform causal path analysis, generating a root cause hypothesis set. This causal path analysis employs a graph traversal algorithm, starting from anomaly nodes in the knowledge graph and tracing back along causal edges to identify potential root cause nodes, forming one or more causal paths. Each causal path corresponds to a root cause hypothesis. The root cause hypothesis set is organized in list form, with each hypothesis including a hypothesis description, causal path length, and path confidence attribute. In practice, the graph traversal algorithm in the causal reasoning submodule uses a depth-first search strategy, setting a maximum search depth to prevent infinite loops. Path confidence is calculated based on the weights of relationships in the knowledge graph, derived from historical data statistics or expert experience. The analysis process of the causal reasoning submodule can consider temporal constraints; for example, a causal path is considered valid only if the event time corresponding to the cause node is earlier than the result node. Optionally, the causal reasoning submodule integrates a probabilistic graphical model to assign prior probabilities to different causal paths. The root cause hypothesis set serves as input to the diagnostic decision submodule.

[0042] The diagnostic decision-making submodule evaluates the posterior probability of the root cause hypothesis set based on a Bayesian network, generating a diagnostic result signal. A Bayesian network is a probabilistic graphical model where nodes represent random variables, including observed variables such as sensor readings and hidden variables such as root causes, and edges represent conditional dependencies between variables. The diagnostic decision-making submodule maps each hypothesis in the root cause hypothesis set to a hidden variable state in the Bayesian network, using the feature values ​​from the current prediction dynamics data and the preprocessed feature dataset as observational evidence. It calculates the posterior probability of each root cause hypothesis using a Bayesian inference algorithm. The diagnostic result signal includes an anomaly type identifier and a root cause probability. The anomaly type identifier is directly derived from the description of the root cause hypothesis, and the root cause probability is the calculated posterior probability value. In specific implementations, the Bayesian network structure of the diagnostic decision-making submodule is predefined by domain experts or learned from historical data. The Bayesian inference process uses exact inference algorithms such as connection tree algorithms or approximate inference algorithms such as Monte Carlo methods. The diagnostic result signal is output in a specific data format; for example, the anomaly type identifier is a string enumeration type, and the root cause probability is a floating-point number type. The implementation of the diagnostic decision submodule ensures robust diagnosis even under uncertain evidence.

[0043] The reinforcement learning optimization module includes a state representation submodule, a multi-agent collaboration submodule, a policy fusion submodule, and a policy optimization submodule. The state representation submodule receives the diagnostic result signal and compresses the high-dimensional state information into a low-dimensional state representation vector through an autoencoder network. The autoencoder network consists of an encoder and a decoder. The encoder compresses the high-dimensional input into a low-dimensional latent space representation, and the decoder reconstructs the original input from the latent space representation. The training objective is to minimize the reconstruction error. The low-dimensional state representation vector captures the most discriminative features from the original high-dimensional state information. In practice, the autoencoder network of the state representation submodule uses a fully connected layer structure. The encoder compresses the diagnostic result signal and other relevant state information into a vector with significantly reduced dimensionality. This low-dimensional state representation vector serves as the observation input for the reinforcement learning agent. The state representation submodule is trained using unsupervised learning, offline, based on a large amount of historical state data. It can be understood that the state representation submodule solves the curse of dimensionality problem in reinforcement learning through dimensionality reduction, thus improving learning efficiency.

[0044] The multi-agent collaborative submodule deploys multiple dedicated agents, each independently learning policies for parameters such as aeration rate, stirring rate, and temperature to generate an initial action suggestion set. Each dedicated agent employs a deep Q-network architecture, with each network consisting of an input layer, hidden layers, and an output layer. The input layer receives a low-dimensional state representation vector, and the output layer corresponds to the Q-value of its respective action space. The initial action suggestion set is the set of actions selected by each dedicated agent based on its current state. For example, the aeration rate agent suggests increasing or decreasing the aeration valve opening, the stirring rate agent suggests increasing or decreasing the stirring motor speed, and the temperature agent suggests increasing or decreasing the heater power. In implementation, the three dedicated agents in the multi-agent collaborative submodule operate in parallel within their respective environments, collecting experience data and updating their deep Q-network parameters through interaction with the environment. The initial action suggestion set is stored as a vector. The implementation of the multi-agent collaborative submodule requires consideration of communication mechanisms between agents to avoid action conflicts. Table 1 illustrates the basic configuration of the dedicated agents in the multi-agent collaborative submodule.

[0045] Table 1: Dedicated Agents in the Multi-Agent Collaboration Submodule

[0046] The policy fusion submodule receives an initial set of action proposals from the multi-agent collaboration submodule. It calculates the weight coefficients of each agent's proposals using an attention mechanism, and performs weighted fusion to generate a collaborative control policy. The attention mechanism calculates the importance score of each agent's proposal based on the current low-dimensional state representation vector. The weight coefficients are obtained by softmax normalization of the importance scores. The weighted fusion process multiplies each agent's initial action proposal by its corresponding weight coefficient and sums the results to obtain the collaborative control policy. In specific implementations, the attention mechanism of the policy fusion submodule is implemented as a small feedforward neural network. This network takes the low-dimensional state representation vector as input and outputs the attention weights for each agent's proposal. The collaborative control policy is a comprehensive action vector containing adjustment suggestions for all controlled parameters. The policy fusion submodule ensures that the action proposals of multiple agents are coordinated and consistent, jointly optimizing the global objective. Optionally, the policy fusion submodule can employ other fusion strategies, such as average fusion or rule-based selection.

[0047] The policy optimization submodule receives the cooperative control policy, performs policy gradient updates based on the proximal policy optimization algorithm, and generates an optimized control policy. The proximal policy optimization algorithm is a policy gradient algorithm that ensures training stability by limiting the step size of policy updates. The policy optimization submodule maintains a policy network that maps states to the probability distribution of actions. The optimized control policy is the updated policy network parameters. In practice, the proximal policy optimization algorithm in the policy optimization submodule uses a pruned objective function to prevent excessively large policy updates. The policy gradient is estimated by sampling empirical trajectories. The optimized control policy is finally saved and deployed in the form of network parameters. The iterative optimization process of the policy optimization submodule enables the system to continuously improve its control performance.

[0048] See Figure 5 This chart demonstrates the collaborative working mechanism of the intelligent diagnostic and optimization control system. The stacked bar charts show the probability distribution of various abnormal root causes at different time points, reflecting the application of knowledge graphs and Bayesian networks in fault diagnosis. The probability changes reflect the dynamic evolution of risk factors during fermentation. Simultaneously, the line graphs on the secondary axis show the real-time optimization and adjustment of key control parameters by the reinforcement learning module. The coordinated changes in aeration rate, stirring rate, and temperature reflect the multi-agent collaborative control strategy, and the curve fluctuations reflect the system's adaptive response to environmental disturbances. The entire chart demonstrates the deep integration and synergistic effect of intelligent diagnostics and optimization control in the fermentation process.

[0049] Example 5: In a specific implementation, the adaptive execution module includes a control translation submodule and an execution drive submodule. The control translation submodule receives the optimized control strategy from the reinforcement learning optimization module and converts it into an actuator-specific command language. The optimized control strategy includes action values ​​and strategy gradient information. For example, the action value may be expressed as increasing the vent valve opening by 3%, decreasing the stirring rate by 50 revolutions per minute, and keeping the heater power unchanged. The control translation submodule maintains a command mapping table internally, mapping the abstract optimized control strategy action values ​​to low-level instructions that the specific actuator can recognize and execute. For vent valve opening adjustment, the command mapping table converts the percentage change into an analog current output value, ranging from 4 mA to 20 mA, corresponding to a valve opening of 0% to 100%. For stirring rate adjustment, the command mapping table converts the speed change into the frequency setpoint of the frequency converter, in Hertz. For temperature control, the command mapping table converts the power adjustment into a duty cycle control signal of a solid-state relay. In practical implementation, the control translation submodule is implemented as a configuration-driven translator. The command mapping table is stored in an XML or JSON configuration file, supporting online modification and updates. Action values ​​in the optimized control strategy undergo range checks and rationality verification, such as checking whether valve opening changes are within the safe range allowed by the equipment to prevent actuators from receiving illegal commands. The converted actuator-specific command language format depends on the actuator type; for example, for intelligent actuators supporting the Modbus RTU protocol, the command language is standard Modbus function codes and register address data frames. In essence, the control translation submodule acts as a bridge between high-level control algorithms and underlying hardware devices, translating policy decisions into executable actions in the physical world.

[0050] The execution drive submodule sends commands to actuators to drive the fermenter equipment and generate control action sequences. It establishes physical connections with each actuator via an industrial fieldbus network, such as using a PROFIBUS-DP bus to connect a positioner for a venting control valve, a PROFINET bus to connect a frequency converter for a stirring motor, and an analog output module to connect a power controller for a heater. The execution drive submodule encapsulates the actuator-specific command language generated by the control translation submodule into data frames conforming to the bus protocol specification, adds the target actuator address and checksum, and then sends it out through the bus interface card. The control action sequence records the detailed content of each command sent to the actuator, along with a precise timestamp derived from the system's high-precision clock source. In implementation, the execution drive submodule is implemented as a multi-threaded application. One thread manages the command queue and schedules the transmission, ensuring commands are issued in sequence and according to timing requirements. Another thread monitors the actuator's status feedback signals for closed-loop verification. The control action sequence includes not only the command itself but also the command sequence number, target device identifier, and expected device status changes. This information is written to the operation log database in real time for process traceability and fault diagnosis. For example, when the optimized control strategy decides to increase the ventilation rate from 50% to 53%, the control translation submodule converts this into an analog current signal of 16.24 mA sent to the ventilation valve positioner. The execution drive submodule then applies this current signal to the valve positioner through the analog output module and generates a record containing "Timestamp: 2024-05-21 10:15:30.123, Device: Ventilation Valve V-101, Command: Set Opening Degree 53%, Corresponding Current: 16.24mA" and stores it in the control action sequence. It can be understood that the execution drive submodule is the final execution link in the control loop, and its reliability and real-time performance directly determine whether the optimized control strategy can be accurately implemented during the physical fermentation process.

[0051] The system also includes a model update module, which is used to retrain the digital twin model and the neural network model in the deep prediction module based on the real-time data stream. The model parameters are updated through an incremental learning algorithm. The real-time data stream comes from the multimodal data of the fermentation tank continuously collected by the multi-source data perception module and the control action sequence effect data generated by the adaptive execution module. This data is cached in a circular data buffer. The model update module periodically checks the amount of data in the buffer. When the amount of data reaches a preset threshold or exceeds a certain time interval since the last model update, the retraining process is triggered. The retraining process targets the machine learning component in the digital twin model and the long short-term memory network model in the deep prediction module. The incremental learning algorithm uses online sequence learning or mini-batch gradient descent methods to fine-tune the existing model parameters using newly collected data, rather than training from scratch. In practice, the model update module runs as a background service. It reads the latest batch of data from the data buffer, performs preprocessing and feature engineering on the data consistent with the offline training phase, and then loads the parameters of the currently deployed digital twin model and the neural network model in the deep prediction module. It applies incremental learning algorithms to update the parameters; for example, for the Long Short-Term Memory network model in the deep prediction module, the model update module might use a stochastic gradient descent algorithm with momentum, iterating several times on the new batch of data with a small learning rate to avoid catastrophic forgetting of old knowledge. The updated model parameters are evaluated on a validation set. If the performance improvement meets predetermined standards, it is deployed to the production environment to replace the old model. In essence, the model update module enables the system to continuously learn and adaptively optimize, gradually adapting to the dynamic characteristics of the fermentation process and changes in equipment status.

[0052] In practical implementation, the triggering strategy of the model update module can be flexibly configured. One strategy is based on a fixed time interval, such as triggering a model update every eight hours. Another strategy is data-driven, automatically triggering an update when the model's prediction error consistently exceeds a threshold. The selection of incremental learning algorithms needs to consider the model type and computing resources. For relatively simple machine learning components in the digital twin model, lightweight algorithms such as recursive least squares may be used. For complex deep neural networks in the deep prediction module, algorithms with strong adaptability and effective control over forgetting, such as elastic weight consolidation algorithms, are adopted. When performing update operations, the model update module saves a backup of the current model and records detailed update logs, including the reason for the update trigger, the amount of data used, and a comparison of model performance before and after the update, so that it can quickly roll back to the previous stable version if the update effect is unsatisfactory. Understandably, this cautious update mechanism ensures the security of system operation.

[0053] Optionally, the model update module can support multiple update modes: a fully automatic mode requiring no manual intervention, a semi-automatic mode requiring operator confirmation before updating the model, and a manual mode allowing engineers to specify the dataset and parameters for updating. The model update module can also integrate a model performance monitoring dashboard, visually displaying the trends in model prediction accuracy and error distribution over time, providing support for decision-making. In some embodiments, the model update module can be designed with a distributed architecture, distributing computationally intensive model retraining tasks across multiple computing nodes for parallel execution, significantly reducing the update time. This is particularly important for the parallel monitoring and optimization of multiple tanks in large-scale fermentation plants.

[0054] Optionally, the command mapping table of the control translation submodule can have an adaptive adjustment function. For example, when it is detected that the response characteristics of an actuator have drifted due to mechanical wear, i.e., the actual opening degree corresponding to a given current signal deviates from the nominal value, the control translation submodule can automatically fine-tune the parameters in the command mapping table based on historical operating data and the actual effects of sensor feedback. This allows the issued commands to achieve the expected control effect more accurately. This adaptive adjustment further enhances the robustness of the entire system in the face of real-world problems such as equipment aging. The communication mechanism of the execution drive submodule also has a redundant design. For example, in addition to the mainstream Profibus-DP bus, a hard-wired backup control loop may be equipped for critical actuators. When a bus communication interruption is detected, the execution drive submodule can automatically switch to the backup loop to send simple switching commands, ensuring that basic safety functions are not affected. It can be understood that these optional enhancement features are a deep consideration of the complexity and reliability requirements of industrial environments.

[0055] Optionally, the data management strategy of the model update module includes a data filtering mechanism. Not all real-time collected data will be used for model updates. The model update module will combine diagnostic result signals to prioritize high-quality data that represent normal operating conditions or typical abnormal patterns for training. Invalid data or outliers that are obviously caused by sensor failures or extreme random events will be filtered out. This can improve the efficiency of incremental learning and the generalization performance of the new model. The model update module can also be linked with the knowledge graph query submodule. When new abnormal patterns or process knowledge are added to the knowledge graph, a targeted model update task can be triggered to integrate the new domain knowledge into the digital twin model and the prediction model.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital twin-based optimization and intelligent control system for pomegranate leaf fermentation process, characterized in that, The system achieves intelligent control through the following modules: The multi-source data sensing module is used to collect multimodal data from the fermenter through IoT sensors and vision sensors, including temperature, pH, dissolved oxygen, microbial density and image data, and generate raw multimodal data streams through data fusion; The intelligent data preprocessing module is used to receive the original multimodal data stream, perform adaptive filtering and feature extraction, and generate a preprocessed feature dataset; The digital twin simulation engine module is used to receive the preprocessed feature dataset, interact with the digital twin model in real time, update the virtual fermentation tank status, and generate simulation status data. The deep prediction module is used to receive the simulation state data, apply deep learning methods to predict fermentation kinetic parameters, and generate predicted kinetic data. The intelligent diagnostic module is used to receive the predicted dynamics data and the preprocessed feature dataset, perform anomaly root cause analysis based on the knowledge graph, and generate diagnostic result signals. The reinforcement learning optimization module is used to receive the diagnostic result signal, use a reinforcement learning agent to optimize the control parameters, and generate an optimized control strategy. The adaptive execution module receives the optimized control strategy, adjusts the fermentation conditions through the actuator, and generates a sequence of control actions.

2. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twin as described in claim 1, characterized in that, The original multimodal data stream includes temperature data sequences, pH data sequences, dissolved oxygen data sequences, microbial density data sequences, and image data sequences. The preprocessed feature dataset includes time-domain features, frequency-domain features, and spatial features. The simulation state data includes virtual temperature, virtual pH, virtual dissolved oxygen, and virtual microbial density. The predicted kinetic data includes growth rate prediction and product concentration prediction. The diagnostic result signal includes anomaly type identifier and root cause probability. The optimized control strategy includes action values ​​and strategy gradients. The control action sequence includes actuator commands and timestamps.

3. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twins according to claim 1, characterized in that, The multi-source data sensing module includes: The sensor adapter submodule is used to connect heterogeneous sensor networks, perform protocol conversion and data verification, and generate standardized sensor data. The multimodal fusion submodule is used to receive the standardized sensor data, apply the Kalman filter algorithm to perform data fusion, and generate fused multimodal data; The data stream management submodule is used to receive the fused multimodal data, cache and segment it according to time windows, and generate the original multimodal data stream.

4. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twin as described in claim 1, characterized in that, The intelligent data preprocessing module includes: The noise removal submodule is used to receive the original multimodal data stream, apply wavelet transform algorithm to remove noise, and generate clean data; The feature extraction submodule is used to receive the clean data, extract time-domain and frequency-domain features using a convolutional neural network, and generate a feature vector set; The data normalization submodule is used to receive the feature vector set, perform min-max normalization processing, and generate a preprocessed feature dataset.

5. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twin as described in claim 1, characterized in that, The digital twin simulation engine module includes: The model loading submodule is used to load digital twin models, which include physical equations and machine learning components; The state update submodule is used to receive the preprocessed feature dataset as input to drive the state update of the digital twin model and generate the current virtual state; The data output submodule is used to extract state variables from the digital twin model and generate simulation state data.

6. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twin as described in claim 1, characterized in that, The depth prediction module includes: The sequence modeling submodule is used to receive the simulation state data, model the time series dependencies using a long short-term memory network, and generate a hidden state sequence. The prediction output submodule is used to receive the hidden state sequence, output predicted values ​​through a fully connected neural network layer, and generate prediction dynamics data.

7. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twin as described in claim 1, characterized in that, The intelligent diagnostic module includes: The knowledge graph query submodule is used to access the domain knowledge graph and query historical patterns of fermentation anomalies; The causal reasoning submodule is used to receive the prediction dynamics data and the preprocessed feature dataset, combine them with the knowledge graph to perform causal path analysis, and generate a root cause hypothesis set. The diagnostic decision submodule is used to evaluate the posterior probability of the root cause hypothesis set based on Bayesian networks and generate diagnostic result signals.

8. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twin as described in claim 1, characterized in that, The reinforcement learning optimization module includes: The state representation submodule is used to receive the diagnostic result signal and compress the high-dimensional state information into a low-dimensional state representation vector through an autoencoder network. The multi-agent collaboration submodule is used to deploy multiple dedicated agents, which independently learn strategies for ventilation rate, stirring rate and temperature parameters, and generate an initial action suggestion set. The strategy fusion submodule is used to receive the initial action suggestion set, calculate the weight coefficient of each agent's suggestion through the attention mechanism, and perform weighted fusion to generate a collaborative control strategy; The strategy optimization submodule is used to receive the cooperative control strategy, perform policy gradient updates based on the near-end policy optimization algorithm, and generate an optimized control strategy.

9. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twin as described in claim 1, characterized in that, The adaptive execution module includes: The control translation submodule is used to receive the optimized control strategy and convert it into an actuator-specific command language; The execution driver submodule is used to send commands to the actuators to drive the fermenter equipment and generate control action sequences.

10. The pomegranate leaf fermentation process optimization and intelligent control system based on digital twin as described in claim 1, characterized in that, The system also includes a model update module, which is used to retrain the digital twin model and the neural network model in the deep prediction module based on the real-time data stream, and update the model parameters through an incremental learning algorithm.

Images