Anaerobic fermentation collaborative management system and method based on digital twinning and hybrid intelligence

Abstract

This invention proposes a collaborative control system and method for anaerobic fermentation based on digital twins and hybrid intelligence. The system comprises a physical layer, a digital twin layer, a hybrid intelligent decision-making layer, and a collaborative execution layer. The physical layer is the source of data generation and the final execution target of control commands. The digital twin layer constructs a complete, dynamic, and high-fidelity digital copy of the physical plant in virtual space. The hybrid intelligent decision-making layer uses intelligent algorithms to find the optimal control strategy within the digital twin environment. The collaborative execution layer receives the optimal control strategy, decomposes it into specific, time-sequential control commands for each actuator, and distributes them to the corresponding actuators in the physical layer for execution. Simultaneously, the execution results are fed back to the digital twin layer, forming a closed-loop control system. This invention achieves an intelligent leap from stable operation of a single tank to maximizing the economic benefits of the entire plant, significantly improving the overall operational efficiency and intelligence level of the biomass energy station.

Description

Technical Field

[0001] This invention relates to the technical field of industrial internet and intelligent manufacturing, and in particular to an anaerobic fermentation collaborative control system and method based on digital twin and hybrid intelligence. Background Technology

[0002] Anaerobic fermentation is a core biotechnology for treating organic waste (such as kitchen waste, livestock and poultry manure, and municipal sludge) and producing biogas (mainly methane). Modern anaerobic fermentation plants typically consist of clusters of multiple large fermenters to achieve large-scale, continuous material handling and energy production. However, existing control systems face significant technical bottlenecks when dealing with this complex multi-tank, multi-material coupling system.

[0003] Current control systems are mostly based on distributed control systems (DCS) or programmable logic controllers (PLCs), combined with upper-level monitoring and data acquisition (SCADA) software. Their operating logic typically involves setting up independent control loops for individual fermenters. For example, a PID (proportional-integral-derivative) control loop is formed using temperature sensors and heaters to maintain a constant temperature inside the tank; another PID loop is formed using pH meters and acid / alkali dosing devices to stabilize pH levels. This control method plays a role in ensuring stable operation of a single reactor near preset parameters, but it is essentially a "passive," error-feedback-based single-variable control. For modern biogas plants or biomass energy stations with multiple fermenters and processing various wastes, achieving collaborative optimization across devices, materials, and time and space presents a higher-level challenge. Existing control systems generally have the following limitations:

[0004] 1. Independent operation of individual tanks, lacking global coordination: The control systems of each fermentation tank operate independently, like information silos. Control decisions are based solely on the internal parameters of the tank itself, completely ignoring the interrelationships between tanks and the overall operational goals of the entire plant. For example, the system cannot intelligently determine which tank to prioritize for a newly arrived batch of high-calorific-value materials, based on the plant's overall material inventory and the "health status" of each tank, in order to maximize the plant's total gas production. Material allocation, biogas storage, and thermal energy scheduling lack a "smart brain" for coordinated control from a global perspective of maximizing plant-wide material balance, energy scheduling, and economic efficiency.

[0005] 2. Rigid Control Strategies and Poor Adaptability: The control logic is mostly based on fixed rules preset by engineers (such as the permission-based logic control mentioned in some published patents CN118109284A) or single-objective PID control. Once these strategies are set, they rarely change and cannot adapt to the highly nonlinear and time-varying characteristics of anaerobic fermentation processes, as well as the dynamic changes in external conditions such as feed components and ambient temperature. For example, when the organic load of the feed fluctuates, a fixed stirring intensity may lead to insufficient stirring or excessive energy consumption. The system lacks multi-objective dynamic optimization capabilities and cannot find the optimal balance point among multiple mutually constraining objectives such as "maximum gas production," "most stable operation," and "lowest energy consumption" based on real-time operating conditions.

[0006] 3. Lack of forward-looking planning and delayed decision-making: Existing systems primarily rely on real-time feedback control, i.e., "post-event control." They cannot utilize foreseeable future information to plan production in advance. For example, the system cannot proactively schedule production and implement flexible production based on 24-hour weather forecasts (which significantly impact tank insulation energy consumption) or time-of-use electricity pricing (which directly affects biogas power generation revenue). If a sharp drop in temperature could be predicted for tomorrow night, the system could have preemptively "stored" more heat in the tanks during the lower-cost daytime hours; if the peak electricity price could be predicted for tomorrow afternoon, the system could have adjusted its feeding and operating strategies to strive for peak gas production during that period. This lack of capability causes the plant to miss numerous opportunities to optimize operating costs and improve economic returns.

[0007] 4. Knowledge cannot be transferred across tanks, and experience is difficult to reuse: During the long-term operation of a fermenter, an efficient operating mode for a specific material may be discovered through manual adjustments or by chance. However, this valuable "experienced knowledge" often exists in an unstructured form in the operator's mind or is limited to the tank's historical data. Existing systems lack effective mechanisms to mine and solidify this knowledge, and to safely and quickly replicate or transfer it to other tanks processing similar materials. This leads to widespread "repeated trial and error" and "repeated learning," severely hindering the overall improvement of the plant's intelligence level and operational efficiency.

[0008] Therefore, there is an urgent need for a new control technology that can break through the limitations of single-tank control, take a holistic view, integrate future information, and achieve adaptive, collaborative, and economically optimal operation of multi-tank anaerobic fermentation facilities. Summary of the Invention

[0009] To address the technical problems of isolated control systems, rigid strategies, delayed decision-making, and lack of knowledge sharing in existing technologies, this invention proposes a collaborative control system and method for anaerobic fermentation based on digital twins and hybrid intelligence. The aim is to construct a closed-loop intelligent control system covering a "physical fermentation tank group - digital virtual factory." This system uses high-fidelity digital twin technology to accurately simulate and extrapolate the entire anaerobic fermentation process, and combines hybrid intelligent algorithms (especially multi-objective reinforcement learning) for global and forward-looking optimization decisions. Ultimately, this achieves adaptive collaboration among multiple anaerobic fermentation facilities, multi-objective dynamic optimization, and maximized economic benefits.

[0010] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0011] An anaerobic fermentation collaborative control system based on digital twin and hybrid intelligence includes: a physical layer, a digital twin layer, a hybrid intelligence decision-making layer, and a collaborative execution layer;

[0012] The physical layer includes multiple anaerobic fermentation reactors, each equipped with a monitoring and diagnostic system for monitoring internal state parameters and an independently adjustable actuator for executing control commands.

[0013] The digital twin layer interacts with the physical layer in real time. It contains multiple high-fidelity virtual tank models that correspond one-to-one with multiple anaerobic fermentation reactors, as well as a virtual factory scheduling model that characterizes the material and energy flow between multiple anaerobic fermentation reactors. Each high-fidelity virtual tank model deeply integrates a mechanism model based on the anaerobic digestion model ADM1, a three-dimensional state visualization model for presenting the multidimensional physical field distribution inside the tank, and a radiomics diagnostic model based on non-destructive tomography technology.

[0014] The hybrid intelligent decision-making layer optimizes decisions based on the virtual environment provided by the digital twin layer. It includes a real-time adaptive optimizer and a group knowledge transfer module. The real-time adaptive optimizer employs a multi-objective reinforcement learning algorithm to simulate and extrapolate within the virtual factory scheduling model, generating a collaborative control strategy that optimizes global benefits. The group knowledge transfer module is used to safely transfer optimization strategies learned from a single high-fidelity virtual tank model to other high-fidelity virtual tank models after verification.

[0015] The collaborative execution layer is used to receive the collaborative control strategy generated by the hybrid intelligent decision-making layer, decompose it into specific and time-sequential control instructions for each execution mechanism, and send them to the execution mechanism corresponding to the physical layer for execution. At the same time, the execution results are fed back to the digital twin layer to form a closed-loop control.

[0016] Preferably, the monitoring and diagnostic system includes a multi-type sensor array, a non-destructive tomography unit, and a data acquisition and fault diagnosis module;

[0017] The multi-type sensor array includes general-purpose sensors for measuring temperature, pH value, pressure, liquid level, and flow rate, as well as specialized chemical sensors for online detection of volatile fatty acids, biogas components, and redox potential.

[0018] The non-destructive tomography unit uses electrical resistive tomography (ERT) or ultrasonic tomography (UT) technology to acquire media distribution image data of the internal cross section of the anaerobic fermentation reactor.

[0019] The data acquisition and fault diagnosis module is used to acquire and fuse data from the multi-type sensor array and non-destructive tomography unit in real time, and perform signal filtering and normalization processing; then, based on the data-driven method, the monitoring data is analyzed to achieve early warning and diagnosis of abnormal operating conditions; among which, abnormal operating conditions include acidification, blockage, and foaming.

[0020] Preferably, the implementation method of the radiomics diagnostic model in the high-fidelity virtual tank model is as follows: extracting radiomics features characterizing the mixing uniformity, bubble distribution, and solid-liquid stratification of the material inside the tank from the cross-sectional image obtained by the non-destructive tomography unit. The radiomics features include texture features, shape features, and grayscale distribution statistical features. Then, the radiomics features are input into a pre-trained machine learning classification or regression model to achieve quantitative diagnosis and evaluation of complex spatial anomalies such as crusting, short-circuiting, and stirring dead zones.

[0021] Preferably, the real-time adaptive optimizer employs a reinforcement learning algorithm that defines the entire anaerobic fermentation plant as a Markov decision process (MDP), wherein:

[0022] The state space is defined as a joint vector containing all current operating parameters of the pressurized oxygen fermentation reactor, health status index output by the radiomics diagnostic model, material inventory information, and weather forecast and time-of-use electricity price information for a future period.

[0023] The action space is defined as a joint action vector that includes the regulation of all pressurized oxygen fermentation reactors, including the feed rate, feed type ratio, stirring power and start-stop strategy, and heating power of each reactor.

[0024] The reward function is defined as a multi-objective composite function, which is a weighted sum of the factory's total economic benefits, operational stability, and total energy consumption costs over a preset time period. The specific mathematical expression is as follows:

[0025] ;

[0026] in, State at time t Next action Instant rewards received The item is based on biogas production and real-time electricity prices, calculating the revenue from electricity generation. This is a penalty term that is generated when a critical operating parameter deviates from the optimal range. The item is the energy cost of stirring and heating operations. These are adjustable weighting coefficients used to balance different optimization objectives.

[0027] Preferably, the group knowledge transfer module employs a policy distillation algorithm to achieve knowledge transfer, and its implementation method includes:

[0028] The reinforcement learning policy network that has been trained on the source virtual tank model is used as the "teacher network".

[0029] Initialize a new policy network as the "student network" for the target virtual tank model;

[0030] In the simulation environment of the target virtual tank model, the output action distribution of the "teacher network" is used to guide the training of the "student network". Its loss function consists of two parts: one part is the reinforcement learning loss generated by the interaction between the student network and the environment, and the other part is the KL divergence loss that measures the difference between the output action distribution of the student network and the teacher network.

[0031] By minimizing this composite loss function, the student network can quickly learn and inherit the excellent strategies of the teacher network, thereby accelerating the learning process on the target tank.

[0032] Preferably, the collaborative execution layer adopts a model predictive control (MPC) architecture, using the long-cycle, high-level collaborative control strategy generated by the real-time adaptive optimizer as a dynamic setpoint or optimization target. Based on this dynamic setpoint and combined with the short-term predictive capability of the high-fidelity virtual tank model, the MPC controller performs rolling optimization within a relatively short prediction time domain to calculate a series of refined, time-sequential control command sequences for the future, and sends the first control command in the sequence to the physical layer's execution mechanism, repeating this process in the next control cycle.

[0033] A collaborative control method for anaerobic fermentation based on digital twins and hybrid intelligence, comprising the following steps:

[0034] Step 1: Digital Twin Construction and State Mapping: Through the physical layer monitoring and diagnostic system, real-time operation data and internal state images of multiple anaerobic fermentation reactors are collected and injected into the digital twin layer to drive the high-fidelity virtual tank model and the virtual factory scheduling model to be updated synchronously, thus completing the real-time state mapping from the physical factory to the virtual factory.

[0035] Step 2: Future Scenario and Strategy Generation: External environment prediction information is input into the hybrid intelligent decision layer. The real-time adaptive optimizer of the hybrid intelligent decision layer uses the virtual factory as a simulation sandbox. Based on the preset multi-objective reward function, it uses reinforcement learning algorithm to perform tens of thousands of rapid simulations, automatically explores and generates a set of collaborative control strategy sequences that can achieve global optimality.

[0036] Step 3: Virtual Verification and Security Assessment: The collaborative control strategy sequence generated in Step 2 is simulated and executed in a complete cycle in the virtual factory of the digital twin layer to assess its expected benefits, impact on system stability, and whether there is a risk of violating security constraints; if the verification fails, the optimizer parameters are adjusted and the process returns to Step 2 to regenerate the strategy.

[0037] Step 4: Command Decomposition and Closed-Loop Execution: The verified collaborative control strategy is sent to the collaborative execution layer, which decomposes it into time-sequential control commands for the actuators of each anaerobic fermentation reactor and sends them to the physical layer. After the physical layer executes the commands, its state changes are captured by the monitoring system and fed back to the digital twin layer for continuous calibration and correction of the twin model, thus forming an adaptive and self-optimizing closed-loop control process.

[0038] The beneficial effects of this invention are:

[0039] 1) Achieving global multi-objective optimization and improving economic efficiency: This invention fundamentally breaks through the mode of independent operation of a single tank. Through a plant-level digital twin model and multi-objective optimization algorithm, it can dynamically balance multiple mutually restrictive objectives such as gas production efficiency, operational stability, energy consumption, and material handling capacity from the perspective of the entire plant. The system of this invention no longer pursues the stability of a single parameter, but rather the maximization of the overall economic benefits of the entire plant, which can significantly improve the profitability of the anaerobic fermentation station.

[0040] 2) Enhanced operational flexibility through forward-looking scheduling capabilities: By integrating future information such as weather forecasts and time-of-use electricity pricing into the decision-making process, the system of this invention possesses the ability to "predict the future and plan ahead." It can formulate optimal production plans and resource scheduling schemes hours or even days in advance, achieving flexible production, proactively avoiding risks (such as responding to cold waves), and proactively seizing opportunities (such as maximizing power generation revenue during peak electricity price periods), transforming factory operations from passive response to proactive management.

[0041] 3) Achieving knowledge sharing and secure transfer, accelerating intelligent evolution: The group knowledge transfer module of this invention effectively breaks down the "data silos" and "experience barriers" between different fermenters. Through mechanisms such as strategic distillation, the "successful experiences" accumulated by one fermenter through learning can be safely and efficiently learned and reused by other fermenters, realizing "collective intelligence" across the entire plant. This not only reduces the cost of repeated trial and error but also significantly accelerates the iteration and evolution of the entire plant's level of intelligence.

[0042] 4) Provides advanced decision support and assists strategic planning: The high-fidelity digital twin layer itself is a powerful "what-if" scenario simulation platform. Factory managers can use this virtual factory to conduct various simulation experiments without affecting actual production. For example, they can simulate adding a new fermenter, changing the pretreatment process of a certain material, or adopting a new biogas purification technology. The system can quickly predict the long-term impact of these changes on the overall plant operation and efficiency, thereby providing scientific and quantitative data support for major strategic decisions such as capacity expansion planning, process modification, and investment evaluation. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the system of the present invention.

[0045] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0046] 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.

[0047] Examples, such as Figure 1 As shown, an anaerobic fermentation collaborative control system based on digital twin and hybrid intelligence includes: a physical layer, a digital twin layer, a hybrid intelligence decision-making layer, and a collaborative execution layer.

[0048] 1. The Physical Layer, serving as the physical foundation and execution terminal of the system, encompasses all physical entities within the anaerobic fermentation plant. In this embodiment, the physical layer includes N (N≥2) anaerobic fermentation reactors, along with supporting monitoring and diagnostic systems and actuators. The anaerobic fermentation reactors are the core sites for biological reactions. Specifically, they include:

[0049] The monitoring and diagnostic system is a comprehensive, three-dimensional sensing system equipped for each anaerobic fermentation reactor. It includes multi-type sensor arrays, non-destructive tomography units, and data acquisition and fault diagnosis modules.

[0050] In this embodiment, the multi-type sensor array includes basic parameter sensors and online analyzers for key biochemical parameters.

[0051] Basic parameter sensors: Standard industrial sensors deployed on each anaerobic fermentation reactor are used to measure in real time temperature (thermocouples or PT100), pH value (industrial online pH meter), tank pressure, liquid level (ultrasonic or radar level gauge), inlet and outlet flow rates (electromagnetic flow meter), and agitator motor current (to reflect stirring torque and viscosity changes), etc. These data form the basis of process control.

[0052] Key Biochemical Parameter Online Analyzer: To gain deeper insights into the state of biological reactions, the system integrates a dedicated online analyzer. For example, a near-infrared spectroscopy (NIR) probe or automatic titrator is used to monitor the concentrations of total volatile fatty acids (VFA) and alkalinity online, which are key indicators for early warning of system acidification. Simultaneously, an online biogas analyzer (such as an infrared or laser gas analyzer) is equipped to detect the concentrations of components such as CH4, CO2, and H2S in biogas in real time, directly reflecting methanogenesis efficiency and system health.

[0053] In this embodiment, a non-destructive tomographic imaging unit is innovatively deployed on the exterior or interior of each anaerobic fermentation reactor tank wall. The preferred option is resistive tomography (ERT) technology. ERT injects a safe, weak current into the medium through multiple rings of electrodes arranged circumferentially around the tank, measures the boundary potential distribution, and then uses an image reconstruction algorithm to reconstruct the conductivity distribution of the tank's internal cross-section. Due to the significant differences in conductivity between the biogas bubbles, solid particles, and liquid phases, ERT images can clearly reveal spatial structural information such as the gas holdup distribution, solid-liquid mixing uniformity, and the presence of dead zones or material agglomeration within the tank. An alternative option is ultrasonic tomography (UT) technology.

[0054] The data acquisition and fault diagnosis module is used to collect and fuse data from the multi-type sensor array and non-destructive tomography unit in real time, performing signal filtering and normalization. Then, based on a data-driven approach, the monitoring data is analyzed in real time to achieve early warning and diagnosis of abnormal operating conditions such as acidification, blockage, and foaming. Specifically, data from all sensors and analyzers is collected to the edge computing gateway via fieldbus (e.g., Profibus-DP) or wireless network (e.g., LoRa). The gateway has built-in data preprocessing algorithms (filtering, denoising, normalization) and a preliminary fault diagnosis model. This model can monitor key parameters based on statistical process control (SPC) methods (e.g., Shewhart control charts), triggering an alarm if a parameter exceeds a preset control limit. Meanwhile, more advanced diagnostics, such as image-based diagnostics, will be performed in the digital twin layer.

[0055] Independently adjustable actuators, devices associated with each pressurized oxygen fermentation reactor that can receive control signals and change the reactor's operating state, such as variable frequency feed pumps, adjustable speed stirrers, solenoid valves for the heating circulation system connected to a PLC, and electric regulating valves on the biogas pipeline.

[0056] 2. Digital Twin Layer: As a digital mirror and simulation platform of the physical world, it is responsible for high-fidelity modeling, simulation, and extrapolation of the physical factory (it constructs a complete, dynamic, and high-fidelity digital copy of the physical factory in virtual space). This layer achieves bidirectional real-time data connection with the physical layer through Industrial Internet of Things (IIoT) technology. It mainly consists of two types of models:

[0057] High-fidelity virtual tank model: A "sensory-diagnostic" enhanced digital twin is created for each pressurized oxygen fermentation reactor. This twin is the product of deep fusion of multiple models, including:

[0058] Mechanistic Model: The core framework is based on the Anaerobic Digestion Model 1 (ADM1) published by the International Water Association (IWA). This model, grounded in the laws of mass and energy conservation, uses a series of complex differential-algebraic equations (DAEs) to describe the intricate biochemical reaction network (such as hydrolysis, acid production, acetic acid production, and methanogenesis) and physicochemical processes (such as gas-liquid equilibrium and acid-base equilibrium) during anaerobic fermentation. It can simulate the core dynamics of feeding, gas production, and microbial growth at a fundamental mechanistic level.

[0059] ADM1 is a complex model containing 26 dynamic state variables and 19 biochemical kinetic processes. Its general mass balance equation can be expressed as:

[0060]

[0061] in:

[0062] It is the concentration of component i in the reactor (a state variable).

[0063] It's time.

[0064] It is the feed flow rate. It is the effective volume of the reactor.

[0065] It is the concentration of component i in the feed.

[0066] This represents the sum of the contributions of all 19 biochemical reactions to the change in the concentration of component i.

[0067] This represents the reaction rate of the j-th biochemical process (such as hydrolysis rate or methanogenesis rate). It is typically a Monod-type kinetic equation and incorporates various influencing factors such as pH, temperature, and inhibitors. For example, the rate of a typical substrate degradation process can be expressed as:

[0068]

[0069] in It is the maximum reaction rate constant. It is the biomass of functional microorganisms. It is the substrate concentration. It is the half-saturation constant. These are various suppression functions.

[0070] This mechanistic model is driven by actual feed composition, flow rate, and temperature data transmitted from the physical layer, and uses historical data and genetic algorithms for parameter calibration, enabling it to predict key outputs such as gas production and VFA concentration with high accuracy.

[0071] A three-dimensional state visualization model is constructed based on a three-dimensional geometric model (such as a CAD model) of the pressurized oxygen fermentation reactor, driven in real time by the mechanistic model and physical layer sensor data. This model can intuitively and dynamically present the distribution of physical fields within the reactor that are difficult to measure directly, such as temperature field, velocity field (caused by agitation), substrate concentration gradient, and bubble distribution, in the form of three-dimensional cloud maps and streamline diagrams. This provides a powerful visualization tool for understanding and diagnosing spatial inhomogeneities. Specifically, the model first establishes a precise geometric shell by importing the reactor's design drawings (such as STEP or IGES format). Then, it uses computational fluid dynamics (CFD) software (such as ANSYS Fluent or OpenFOAM) to perform offline simulation of the flow field within the reactor, obtaining typical velocity field distributions under different agitation modes. During online operation, the model maps scalar field data such as component concentration distribution and temperature distribution calculated by the mechanistic model, as well as point data measured by the physical layer sensors, onto the three-dimensional geometric model through interpolation or data fusion algorithms. Simultaneously, combined with CFD simulation results, it dynamically displays the streamlines during agitator rotation, the suspension trajectory of solid particles, and the rising path of biogas bubbles. Users can engage in immersive interaction through VR / AR devices, freely cutting and scaling the virtual tank to intuitively observe its internal state, greatly enhancing their understanding of complex processes.

[0072] Radiomics Diagnostic Model: This is an innovative diagnostic model that draws inspiration from medical image analysis. It utilizes non-destructive tomographic imaging units (such as ERT / UT) to acquire cross-sectional images of the interior of a pressurized oxygen fermentation reactor. Image processing algorithms then extract a series of quantified radiomics features (such as texture features describing mixing uniformity, shape features describing foam or crust regions, and grayscale statistical features describing the degree of solid-liquid stratification). These features are then input into a pre-trained machine learning model (such as Support Vector Machine (SVM) or Convolutional Neural Network (CNN)) to achieve accurate diagnosis and quantitative assessment of complex spatial structural faults that traditional point sensors cannot detect, such as crust formation, short-circuiting, dead zones, and severe foaming.

[0073] Radiomics diagnostic models are key to achieving in-depth diagnostics. Their workflow is as follows:

[0074] Data acquisition: Real-time acquisition of 2D or 3D conductivity distribution image data from the ERT unit of the physical layer.

[0075] Feature extraction: Advanced image analysis algorithms are applied to extract high-throughput, quantified radiomics features from raw tomographic images. These features can be categorized into several types:

[0076] First-order statistical features: describe the distribution of pixel intensity, such as mean, variance, kurtosis, and entropy, and can reflect the overall uniformity of mixing.

[0077] Shape and size characteristics: The area, perimeter, roundness, elongation, etc. of specific regions identified by image segmentation (such as low-conductivity gas masses or high-conductivity solid-phase deposition areas) are used to quantify the size and morphology of crusts and foam layers.

[0078] Texture features: Using algorithms such as Gray-Level Co-occurrence Matrix (GLCM) or Local Binary Pattern (LBP), features such as contrast, correlation, energy, and homogeneity are extracted to describe the complexity and spatial correlation of the medium distribution, and can effectively identify stirring dead zones or material short-circuiting.

[0079] Model diagnostics: The extracted multidimensional feature vectors are input into a pre-trained machine learning classifier (such as a random forest or gradient boosting decision tree). By learning from a large amount of image omics feature data under normal and fault conditions, this classifier can automatically identify and diagnose whether the current reactor has spatial structural faults such as crusting, severe foaming, stirring failure, and solid-liquid stratification, and output the probability and severity level of the fault.

[0080] Virtual Factory Scheduling Model: A higher-level plant-level scheduling model is built upon individual virtual tank models. This model connects all virtual tank models as sub-modules and integrates a material flow network (simulating the distribution paths of waste with different characteristics between tanks), an energy flow network (simulating the generation, consumption, storage, and conversion of biogas, hot water, and electricity), and interfaces for interaction with external markets (such as time-of-use electricity pricing from the power grid and heat prices from heating companies). This model provides a "digital sandbox" for higher-level intelligent decision-making, enabling global, multi-factor collaborative simulation.

[0081] This model uses multiple high-fidelity virtual tank models as nodes to construct a plant-wide dynamic simulation network. It defines:

[0082] Material flow: Waste from different sources (such as canteen A and slaughterhouse B) enters the pretreatment pool of the plant. The model simulates the path and flow of waste to be allocated to different virtual tanks based on its characteristics (including solids content, C / N ratio, etc.) and inventory.

[0083] Energy Flow: The model simulates the biogas produced by each virtual tank flowing into the main pipe. Part of it is used for boiler heating (hot water is redistributed to each tank for insulation), and part enters the combined heat and power (CHP) unit for power generation. The model simulates the optimal biogas distribution strategy based on real-time electricity prices and heat load.

[0084] Information flow: It integrates external information inputs, such as hourly temperature forecasts for the next 72 hours provided by the meteorological service API, and time-of-use electricity price curves for the next 24 hours published by the power trading center.

[0085] This model constitutes a dynamic, high-fidelity virtual environment for the hybrid intelligent decision-making layer to perform global optimization simulations.

[0086] 3. Hybrid Intelligence Decision Layer: As the system's "intelligent brain," this layer is responsible for global and forward-looking decision optimization within the virtual environment provided by the digital twin layer. This layer is the core of the system's decision-making, using intelligent algorithms to find the optimal control strategy within the digital twin environment. Internally, it includes a real-time adaptive optimizer and a swarm knowledge transfer module. The real-time adaptive optimizer employs a multi-objective reinforcement learning algorithm to simulate and extrapolate within the virtual factory scheduling model, generating a collaborative control strategy that optimizes global benefits. The swarm knowledge transfer module is used to safely transfer optimization strategies learned from a single high-fidelity virtual tank model to other high-fidelity virtual tank models after verification.

[0087] Real-time Adaptive Optimizer: Using the virtual factory scheduling model of the digital twin layer as the simulation environment, it employs advanced intelligent optimization algorithms, such as Multi-Objective Reinforcement Learning (MORL) or evolutionary algorithms. The optimizer acts as an agent, learning through high-speed trial and error in the virtual environment. Based on the current global state obtained from the digital twin layer (including the health status of each tank, material inventory, energy demand, and future weather and electricity price information), and around a comprehensive multi-objective optimization function (e.g., maximizing the total gas production value in the next 24 hours, or achieving the highest overall plant operational stability), it automatically explores and evaluates thousands of possible combinations of control actions (e.g., prioritizing the pumping of material A into tank 1, adjusting the stirring mode of tank 2 to intermittent high-intensity, and starting the insulation of tank 3 ahead of schedule). Ultimately, it outputs the currently optimal collaborative control strategy.

[0088] In this embodiment, the optimizer employs an advanced multi-objective deep reinforcement learning algorithm, such as a multi-objective version of Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO), because they are adept at handling high-dimensional complex control problems with continuous states and action spaces, such as anaerobic fermentation.

[0089] Its implementation process is modeled as a plant-wide Markov decision process (MDP):

[0090] State space (S): Defines the state of an agent at any given moment. All observable information, including current operating parameters, health status indices output by the radiomics diagnostic model, material inventory information, and a joint vector of weather forecasts and time-of-use electricity prices for the near future. It is a high-dimensional vector. Its composition is as follows:

[0091]

[0092] in: It is the state vector of the i-th reactor, including: [temperature, pH, VFA, alkalinity, gas production rate, CH4 concentration, liquid level, stirring power, health score output by the radiomics diagnostic model].

[0093] It is a vector of the plant's inventory and resource status, including: [material inventory in each pretreatment tank, biogas tank pressure, and hot water tank temperature].

[0094] It is a vector of future forecast information, including: [the ambient temperature sequence for the next 24 hours, and the electricity price sequence for the next 24 hours].

[0095] Action Space (A): Defines the space in which the agent operates within a state. This refers to all control actions that can be taken. The control includes the feed rate of each reactor, feed type and ratio, stirring power and start / stop strategy, and heating power; it is a continuous, multi-dimensional action vector. :

[0096]

[0097] in:

[0098] It is the control action vector for the i-th reactor, including: [target feed rate, proportion of each material, target stirring power, target heating power].

[0099] It refers to the control actions of the overall energy system, including: [the proportion of biogas allocated to CHP, CHP power generation load].

[0100] Multi-objective reward function (R): This is the "command stick" of reinforcement learning, guiding the agent to learn the desired behavior. This invention designs a composite reward function that integrates economic benefits, operational stability, and energy consumption costs.

[0101]

[0102] Economic benefits:

[0103]

[0104] This calculation covers the instantaneous revenue from biogas power generation, of which... It is the total gas production. It represents the current electricity price, reflecting the incentive to generate more electricity during periods of high electricity prices.

[0105] Stability penalty term Penalty:

[0106]

[0107] This measure imposes an exponential penalty on key stability metrics (such as pH and VFA / Alkalinity ratio) that deviate from their optimal range (opt), ensuring safe system operation.

[0108] Energy cost item:

[0109]

[0110] This calculation covers the total energy cost of stirring and heating all tanks.

[0111] These are weighting coefficients, which managers can adjust according to production goals at different times (e.g., pursuing maximum profit or pursuing the most stable operation), reflecting the flexibility of multi-objective optimization.

[0112] The optimizer's Actor-Critic network uses Input, output action This action is first executed in virtual factory model 202, where the model deduces the state at the next moment based on its internal mechanistic equations. and rewards These transferred samples They are stored in the experience replay pool for continuous training of the Actor and Critic networks, allowing their strategies to be iteratively optimized.

[0113] Group Knowledge Transfer Module: To address the pain point of traditional systems where "knowledge cannot be transferred across tanks," this module utilizes the concept of transfer learning, particularly policy distillation. After the real-time adaptive optimizer has learned an efficient operating strategy for a specific type of material on a source tank through long-term learning, this module is responsible for extracting and generalizing the core parameters and decision logic from this well-trained policy network ("teacher network"). Then, in a virtual model of a target tank handling similar materials, the "teacher network" guides a new, not yet fully trained policy network ("student network") to learn. In the simulation environment of the target virtual tank model, the output action distribution of the "teacher network" guides the training of the "student network." Its loss function comprises two parts: one is the reinforcement learning loss generated by the student network's interaction with the environment, and the other is the KL divergence loss, which measures the difference between the output action distributions of the student network and the teacher network. By minimizing this composite loss function, the student network can quickly learn and inherit the excellent strategies of the teacher network, thereby accelerating the learning process on the target tank. This "knowledge transfer" is not a simple data copying, but rather, under the premise of ensuring safety boundaries (all migration strategies must first undergo sufficient verification simulation in the virtual twin of the target tank), it efficiently and safely transfers and adapts successful "experiences," thereby greatly shortening the learning cycle under new tanks or new operating conditions and achieving a synchronous and rapid improvement in the overall plant's intelligence level.

[0114] In this embodiment: When the optimizer (agent) of tank 1 has been trained for a long time and its policy network (Actor network) is already very efficient in processing Category A food waste, if it is planned to start tank 3 to process Category A waste as well, this module will be activated:

[0115] Policy extraction: The Actor network with the best performance on the virtual model of tank 1 is identified as... .

[0116] Security verification: The simulation was directly applied to a virtual model of tank No. 3 for comprehensive testing. Its performance on the tank No. 3 model was evaluated, and it was checked whether any safety boundaries (such as drastic pH fluctuations) were triggered. This step is crucial for ensuring the safety of knowledge transfer.

[0117] Strategy distillation: Initialize a new Actor network for jar #3. During training When, its loss function Designed as:

[0118]

[0119] in:

[0120] yes The standard reinforcement learning loss generated by the interaction with the virtual environment of tank 3 (such as the policy gradient loss of DDPG).

[0121] This is the strategy distillation loss, used to measure the similarity between student and teacher strategies, and is usually calculated using KL divergence.

[0122]

[0123] here This represents the action (or action distribution) output by the policy network in state s.

[0124] It is a hyperparameter that balances the two types of losses.

[0125] By minimizing this composite loss function, the student network not only learns from the environment but also directly "imitates" the decision-making behavior of the teacher network, thereby enabling it to quickly master efficient control strategies at a speed far exceeding that of learning from scratch.

[0126] 4. Coordinated Execution Layer: Serving as a bridge between decision-making and the physical world, this layer is responsible for translating macroscopic strategies generated by the upper layer into specific operations at the lower level. This layer receives the optimal coordinated control strategy sequence generated by the hybrid intelligent decision-making layer and then uses model predictive control to decompose these long-term, high-level strategies (e.g., "increase the gas production load of tank 1 by 20% within the next 8 hours") into specific, time-sequential, and executable control command sequences (e.g., the feed pump speed and heating valve opening per minute within the next 30 minutes). These refined commands are securely sent to the corresponding actuators in the physical layer, and the execution effect is monitored in real time. The execution results (i.e., the actual response in the physical world) are fed back to the digital twin layer in real time via sensors for continuous calibration and updating of the twin model, thus forming a complete adaptive optimization closed loop of "perception-modeling-decision-execution-feedback".

[0127] The collaborative execution layer adopts a model predictive control (MPC) architecture. The long-cycle, high-level collaborative control strategy generated by the real-time adaptive optimizer is used as a dynamic setpoint or optimization target. Based on this dynamic setpoint and combined with the short-term predictive capability of the high-fidelity virtual tank model, the MPC controller performs rolling optimization within a relatively short prediction time domain to calculate a series of refined, time-sequential control command sequences for the future. The first control command in the sequence is then sent to the physical layer's execution mechanism, and this process is repeated in the next control cycle.

[0128] In this embodiment: when the optimizer provides a coordinated control strategy, such as "increase the gas production rate of tank 1 by 15% over the next 4 hours, while stabilizing the temperature of tank 3 at 38±0.2℃", the coordinated execution layer works as follows:

[0129] Target decomposition: This macroscopic, long-term target is decomposed into the dynamic setpoints of the MPC controller. That is, the setpoint curve of the gas production rate of tank 1 is set to increase linearly by 15% over the next 4 hours, and the setpoint of the temperature of tank 3 is set to 38℃.

[0130] Rolling optimization: Taking the feed control of tank No. 1 as an example, its MPC controller will perform an optimization calculation once in each control cycle (e.g., 5 minutes).

[0131] Prediction: Using a high-fidelity virtual tank model (mechanism model), starting from the current state, predict the impact of different feed rate sequences on the gas production rate within a relatively short time domain (e.g., 30 minutes).

[0132] Optimization: Solve an optimization problem to find the optimal feed rate control sequence so that the predicted gas production rate can track the upper setpoint curve as quickly and smoothly as possible, while satisfying the constraints of the actuator (such as the maximum flow rate of the feed pump).

[0133] Execution: Only the first action in the calculated optimal control sequence (i.e. the average feed rate for the next 5 minutes) is converted into a frequency signal of the variable frequency pump and sent to the execution mechanism of the physical layer.

[0134] Feedback Correction: After 5 minutes, the system enters the next control cycle. The MPC controller first obtains the latest actual gas production rate from the physical layer, compares it with the predicted value, and calculates the model error. Then, it uses this error to correct the next prediction and performs rolling optimization again based on the new state.

[0135] Through this rolling cycle of "prediction-optimization-execution-feedback", the collaborative execution layer can execute upper-level decisions accurately and robustly, while effectively responding to various disturbances in the real world.

[0136] Example 2, as Figure 2 As shown, a collaborative control method for anaerobic fermentation based on digital twins and hybrid intelligence comprises the following steps:

[0137] Step 1: Digital Twin Construction and State Mapping: Through the physical layer's monitoring and diagnostic system, real-time operational data and internal state images of multiple anaerobic fermentation reactors are collected. This data is then injected into the digital twin layer, driving the high-fidelity virtual tank model and the virtual factory scheduling model to be updated synchronously, thus completing the real-time state mapping from the physical factory to the virtual factory.

[0138] Specifically: First, the geometric foundation of the digital twin layer is constructed by performing 3D scanning and CAD modeling of the equipment in the physical plant. Then, a stable, low-latency data link is established between the advanced monitoring and diagnostic systems deployed in the physical layer and the digital twin layer. Real-time operational data (temperature, pH, VFA, gas production components, ERT images, etc.) is continuously injected into the digital twin layer. This data is used, on the one hand, to drive the 3D state visualization model and the radiomics diagnostic model, achieving a real-time mirroring of the physical plant's state; on the other hand, it is used for continuous data assimilation and parameter calibration of the mechanistic model, ensuring the predictive accuracy of the virtual tank model. All virtual tank models, together with the virtual plant scheduling model, constitute a virtual plant that evolves synchronously with the physical world.

[0139] Step 2: Future Scenario and Strategy Generation: External environment prediction information is input into the hybrid intelligent decision layer. The real-time adaptive optimizer of the hybrid intelligent decision layer uses the virtual factory as a simulation sandbox. Based on a preset multi-objective reward function, it uses reinforcement learning algorithm to perform tens of thousands of rapid simulations, automatically explores and generates a set of collaborative control strategy sequences that can achieve global optimality.

[0140] Based on future information and a multi-objective function, a collaborative control strategy is automatically generated using an intelligent optimization algorithm. Specifically, the system triggers a decision optimization process periodically (e.g., once per hour) or when a major event occurs (e.g., upon receiving a new material supply plan). The hybrid intelligent decision layer obtains environmental forecast information (weather, electricity price) for a future period (e.g., 24 hours) from external sources. The real-time adaptive optimizer takes the current state of the virtual factory and this future information as input, using a preset multi-objective reward function (as described in Example 1) as the optimization objective. Within its internal reinforcement learning framework, the optimizer drives the virtual factory model through tens of thousands or even millions of accelerated simulations ("rollouts"). In these rollouts, the agent continuously tries different combinations of collaborative control actions and updates its policy network based on the accumulated rewards. This process is fully automated and ultimately converges to a collaborative control strategy that maximizes the overall reward under the current state and future expectations.

[0141] Step 3: Virtual Verification and Security Assessment: The coordinated control strategy sequence generated in Step 2 is simulated and executed in a complete cycle in the virtual factory of the digital twin layer to assess its expected benefits, impact on system stability, and whether there is a risk of violating security constraints. If the verification fails, the optimizer parameters are adjusted and the process returns to Step 2 to regenerate the strategy.

[0142] The "optimal" strategy generated in step two must undergo a rigorous "virtual review" before being deployed to the physical factory. The system will execute the strategy throughout a complete simulation cycle (e.g., 24 hours) in the digital twin virtual factory. During this process, the system will closely monitor all key safety indicators (KPIs), such as:

[0143] Process safety: Is the pH value of each virtual tank always within the safe range (e.g., 6.8-7.8)? Is the VFA / Alkalinity ratio below the warning threshold (e.g., 0.4)? Does the tank pressure exceed the design limit?

[0144] Equipment safety: Do the control commands require the actuators (such as agitators) to operate under overload for extended periods?

[0145] Benefit assessment: Does the total economic benefit, total energy consumption, and total material handling volume of this strategy meet the expected targets?

[0146] A policy is considered "validated" only if it passes all security checks and meets the expected benefits. Any policy that does not meet the conditions will be rejected, triggering the optimizer to adjust its parameters and generate a new policy.

[0147] Step 4: Command Decomposition and Closed-Loop Execution: The validated collaborative control strategy is distributed to the collaborative execution layer. As described in Example 1, the collaborative execution layer uses methods such as MPC to decompose the strategy into time-sequential control commands for each anaerobic fermentation reactor actuator and issues them to each actuator in the physical layer for execution. After the physical layer executes the commands, its state changes are captured in real time by the monitoring system and fed back to the digital twin layer. These real feedback data are compared with the predicted data of the twin model, and the resulting deviations are used to continuously calibrate and correct the twin model. For example, if the actual gas production is consistently lower than the predicted value of the mechanistic model, the system may automatically fine-tune a certain kinetic parameter in the model. This continuous "physical-virtual-feedback-calibration" closed loop ensures that the digital twin model can always approximate physical reality to the greatest extent possible, thereby guaranteeing the accuracy and effectiveness of subsequent decisions.

[0148] Specific application example: Suppose a food waste treatment plant with four 3000m³ anaerobic digesters uses the system of this invention.

[0149] Scenario: Current time is 4 PM. The system retrieved a forecast from the meteorological service API: Starting at 8 PM tomorrow night, the local temperature will plummet by 10°C, lasting for 6 hours. Simultaneously, information from the power grid company API indicates that tomorrow afternoon from 2 PM to 5 PM will be peak electricity consumption, and industrial electricity prices will increase by 50%.

[0150] System operation:

[0151] Information Input and Status Assessment: Future weather and electricity price information is input into the virtual plant scheduling model. Meanwhile, the digital twin layer shows that tanks 1 and 2 are currently at their peak gas production period with high microbial activity; tank 3, due to the low calorific value of the feed material in the past few days, is currently at the lower limit of the normal range (36.5℃), with moderate gas production activity; tank 4 is in the later stages of its acidification recovery period, with a low load, requiring stable operation.

[0152] Initiating Optimization Decision: The real-time adaptive optimizer of the hybrid intelligent decision layer is triggered. The weights in its multi-objective reward function are set to prioritize maximizing "tomorrow's total revenue (biogas power generation revenue - energy consumption cost)".

[0153] Virtual simulation: In the digital twin "virtual factory", the optimizer, guided by this goal, begins a large-scale, rapid simulation.

[0154] Simulation Scenario A (Routine Operation): Keep all parameters unchanged. The simulation results show that the temperature drop tomorrow night will cause a sharp increase in heating energy consumption in all tanks, especially tank No. 3, thereby eroding profits; at the same time, the gas production during the afternoon peak electricity price period will not increase significantly, resulting in limited growth in power generation revenue.

[0155] Simulation Scenario B (Exploratory Strategy 1): To cope with the cooling, the insulation power of all tanks will be increased comprehensively starting now. The simulation results show that although this can withstand the cooling, the excessive heating several hours in advance leads to a significant increase in total energy costs, making it counterproductive.

[0156] Simulation Scenario C (Exploratory Strategy 2): To capitalize on peak electricity prices, the feed load of all tanks will be increased comprehensively starting now. The simulation results show that gas production in tanks 1 and 2 does increase, but tank 4, still recovering, may experience acidification recurrence due to the high load impact, leading to a surge in stability penalties and a decrease in overall reward.

[0157] ...After tens of thousands of similar but more refined simulations...

[0158] Generate the optimal cooperative strategy: The system ultimately arrives at a globally optimal cooperative strategy.

[0159] Strategy a) For tanks No. 1 and No. 2 (high activity): Starting at 10 PM tonight (off-peak electricity price period), moderately increase their stirring power and insulation setting temperature (from 37℃ to 38℃). The aim is to utilize their high metabolic activity to "store" more bioheat and metabolic potential before cooling down, so as to smoothly pass through the cooling period and ensure that they can maintain maximum gas production capacity during tomorrow afternoon's peak electricity price period.

[0160] Strategy b) For Tank No. 3 (low temperature): A batch of materials with high solids content and known high heat generation efficiency (such as oil-water separator waste with high oil content) scheduled to enter the plant tomorrow will be prioritized for allocation to Tank No. 3 through path planning in the virtual factory scheduling model. The biothermal effect of these materials will be used to assist in raising the temperature, rather than relying solely on electric heating, thus achieving "increasing revenue and reducing expenditure".

[0161] Strategy c) For tank #4 (which needs to be stabilized): Maintain its current low feed load and slightly reduce the stirring intensity over the next 24 hours to ensure a smooth recovery and avoid any aggressive operations.

[0162] Verification and Execution: The strategy passed security verification in the virtual environment. Subsequently, the collaborative execution layer decomposed it into specific timing instructions (such as increasing the opening of the heating valve of tank 1 by 15% at 22:00, and starting the A material feed pump of tank 3 at 10:00 tomorrow morning) and automatically issued them for execution.

[0163] Execution and Results:

[0164] The following day, despite a sharp drop in nighttime temperature, due to pre-emptive measures, the temperature fluctuations in tanks 1 and 2 were far less than usual, maintaining stable gas production. During the afternoon peak electricity price period, they operated at full capacity, generating significantly higher electricity sales revenue than usual. After receiving high-calorific-value materials, tank 3's temperature gradually returned to normal without consuming excessive additional energy. Tank 4 recovered smoothly. According to system backend statistics, compared to simulation predictions using traditional constant-parameter operation, implementing this collaborative strategy increased the plant's overall net profit by approximately 12% over 24 hours.

[0165] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An anaerobic fermentation collaborative management and control system based on digital twinning and hybrid intelligence, characterized in that, include: Physical layer, digital twin layer, hybrid intelligent decision-making layer, and collaborative execution layer; The physical layer includes multiple anaerobic fermentation reactors, each equipped with a monitoring and diagnostic system for monitoring internal state parameters and an independently adjustable actuator for executing control commands. The digital twin layer interacts with the physical layer in real time. It contains multiple high-fidelity virtual tank models that correspond one-to-one with multiple anaerobic fermentation reactors, as well as a virtual factory scheduling model that characterizes the material and energy flow between multiple anaerobic fermentation reactors. Each high-fidelity virtual tank model deeply integrates a mechanism model based on the anaerobic digestion model ADM1, a three-dimensional state visualization model for presenting the multidimensional physical field distribution inside the tank, and a radiomics diagnostic model based on non-destructive tomography technology. The hybrid intelligent decision-making layer optimizes decisions based on the virtual environment provided by the digital twin layer. It includes a real-time adaptive optimizer and a group knowledge transfer module. The real-time adaptive optimizer employs a multi-objective reinforcement learning algorithm to simulate and extrapolate within the virtual factory scheduling model, generating a collaborative control strategy that optimizes global benefits. The group knowledge transfer module is used to safely transfer optimization strategies learned from a single high-fidelity virtual tank model to other high-fidelity virtual tank models after verification. The collaborative execution layer is used to receive the collaborative control strategy generated by the hybrid intelligent decision-making layer, decompose it into specific and time-sequential control instructions for each execution mechanism, and send them to the execution mechanism corresponding to the physical layer for execution. At the same time, the execution results are fed back to the digital twin layer to form a closed-loop control.

2. The anaerobic fermentation collaborative management system based on digital twin and hybrid intelligence according to claim 1, characterized in that, The monitoring and diagnostic system includes a multi-type sensor array, a non-destructive tomography unit, and a data acquisition and fault diagnosis module. The multi-type sensor array includes general-purpose sensors for measuring temperature, pH value, pressure, liquid level, and flow rate, as well as specialized chemical sensors for online detection of volatile fatty acids, biogas components, and redox potential. The non-destructive tomography unit uses electrical resistive tomography (ERT) or ultrasonic tomography (UT) technology to acquire media distribution image data of the internal cross section of the anaerobic fermentation reactor. The data acquisition and fault diagnosis module is used to acquire and fuse data from the multi-type sensor array and non-destructive tomography unit in real time, and perform signal filtering and normalization processing; then, based on the data-driven method, the monitoring data is analyzed to achieve early warning and diagnosis of abnormal operating conditions; among which, abnormal operating conditions include acidification, blockage, and foaming.

3. The anaerobic fermentation collaborative control system based on digital twin and hybrid intelligence according to claim 2, characterized in that, The implementation method of the radiomics diagnostic model in the high-fidelity virtual tank model is as follows: extracting radiomics features that characterize the mixing uniformity of materials in the tank, the distribution state of bubbles, and the solid-liquid stratification from the cross-sectional image obtained by the non-destructive tomography unit. The radiomics features include texture features, shape features, and grayscale distribution statistical features. Then, the image omics features are input into a pre-trained machine learning classification or regression model to achieve quantitative diagnosis and evaluation of complex spatial anomalies such as crusting, short-flow, and stirring dead zones.

4. The anaerobic fermentation collaborative control system based on digital twin and hybrid intelligence according to claim 1, characterized in that, The real-time adaptive optimizer employs a reinforcement learning algorithm that defines the entire anaerobic fermentation plant as a Markov decision process (MDP), where: The state space is defined as a joint vector containing all current operating parameters of the pressurized oxygen fermentation reactor, health status index output by the radiomics diagnostic model, material inventory information, and weather forecast and time-of-use electricity price information for a future period. The action space is defined as a joint action vector that includes the regulation of all pressurized oxygen fermentation reactors, including the feed rate, feed type ratio, stirring power and start-stop strategy, and heating power of each reactor. The reward function is defined as a multi-objective composite function, which is a weighted sum of the factory's total economic benefits, operational stability, and total energy consumption costs over a preset time period. The specific mathematical expression is as follows: ； in, State at time t Next action Instant rewards received The item is based on biogas production and real-time electricity prices, calculating the revenue from electricity generation. This is a penalty term that is generated when a critical operating parameter deviates from the optimal range. The item is the energy cost of stirring and heating operations. These are adjustable weighting coefficients used to balance different optimization objectives.

5. The anaerobic fermentation collaborative control system based on digital twin and hybrid intelligence according to claim 1, characterized in that, The group knowledge transfer module uses a policy distillation algorithm to achieve knowledge transfer, and its implementation method includes: The reinforcement learning policy network that has been trained on the source virtual tank model is used as the "teacher network"; Initialize a new policy network as the "student network" for the target virtual tank model; In the simulation environment of the target virtual tank model, the output action distribution of the "teacher network" is used to guide the training of the "student network". Its loss function consists of two parts: one part is the reinforcement learning loss generated by the interaction between the student network and the environment, and the other part is the KL divergence loss that measures the difference between the output action distribution of the student network and the teacher network. By minimizing this composite loss function, the student network can quickly learn and inherit the excellent strategies of the teacher network, thereby accelerating the learning process on the target tank.

6. The anaerobic fermentation collaborative control system based on digital twin and hybrid intelligence according to claim 1, characterized in that, The collaborative execution layer adopts a model predictive control (MPC) architecture. The long-cycle, high-level collaborative control strategy generated by the real-time adaptive optimizer is used as a dynamic setpoint or optimization target. Based on this dynamic setpoint and combined with the short-term predictive capability of the high-fidelity virtual tank model, the MPC controller performs rolling optimization within a relatively short prediction time domain to calculate a series of refined, time-sequential control command sequences for the future. The first control command in the sequence is then sent to the physical layer's execution mechanism, and this process is repeated in the next control cycle.

7. A method for synergistic control of anaerobic fermentation based on digital twins and hybrid intelligence as described in any one of claims 1-6, characterized in that, The steps are as follows: Step 1: Digital Twin Construction and State Mapping: Through the physical layer monitoring and diagnostic system, real-time operation data and internal state images of multiple anaerobic fermentation reactors are collected and injected into the digital twin layer to drive the high-fidelity virtual tank model and the virtual factory scheduling model to be updated synchronously, thus completing the real-time state mapping from the physical factory to the virtual factory. Step 2: Future Scenario and Strategy Generation: External environment prediction information is input into the hybrid intelligent decision layer. The real-time adaptive optimizer of the hybrid intelligent decision layer uses the virtual factory as a simulation sandbox. Based on the preset multi-objective reward function, it uses reinforcement learning algorithm to perform tens of thousands of rapid simulations, automatically explores and generates a set of collaborative control strategy sequences that can achieve global optimality. Step 3: Virtual Verification and Security Assessment: The collaborative control strategy sequence generated in Step 2 is simulated and executed in a complete cycle in the virtual factory of the digital twin layer to assess its expected benefits, impact on system stability, and whether there is a risk of violating security constraints; if the verification fails, the optimizer parameters are adjusted and the process returns to Step 2 to regenerate the strategy. Step 4: Command Decomposition and Closed-Loop Execution: The verified collaborative control strategy is sent to the collaborative execution layer, which decomposes it into time-sequential control commands for the actuators of each anaerobic fermentation reactor and sends them to the physical layer. After the physical layer executes the commands, its state changes are captured by the monitoring system and fed back to the digital twin layer for continuous calibration and correction of the twin model, thus forming an adaptive and self-optimizing closed-loop control process.

Images