A real-time collaborative system for dynamic optimization of transport pathways and fermentation control

Abstract

The application discloses a kind of transport path dynamic optimization and fermentation control real-time collaborative system.The application first collects fermentation system and transport network data, builds micro-macro-macro multi-scale digital twin base, completes the real-time fusion and stability processing of multi-source heterogeneous data;Based on the total data after fusion, the transport-fermentation coupling state perception and multi-level hierarchical risk warning are completed.Construct a multi-objective optimization model with real-time fermentation state as constraint, and obtain the optimal transport execution scheme, and support real-time rescheduling under abnormal scene.Based on the transport prediction data, a feedforward-feedback composite control architecture is built to realize adaptive control, multi-reactor collaborative load distribution and robust control in the fermentation process.Finally, with the help of global collaborative decision engine, the transport optimization and fermentation control two types of decision instructions are coordinated, the decision execution effect is quantitatively evaluated, and the closed-loop self-optimization of model and strategy is realized.

Description

Technical Field

[0001] This invention relates to the field of dynamic optimization of transport paths and real-time coordination of fermentation regulation, and particularly to a system for dynamic optimization of transport paths and real-time coordination of fermentation regulation. Background Technology

[0002] Bio-fermentation, as a core production link in pillar industries of the national economy such as pharmaceuticals, food additives, and biochemicals, directly depends on the precise control of the fermentation process and the efficient coordination of the material supply chain for its production efficiency and product quality. Currently, the automation level of the fermentation industry has significantly improved. Most enterprises use distributed control systems (DCS) to achieve online monitoring and feedback control of core process parameters such as fermenter temperature, pH value, and dissolved oxygen concentration. Some enterprises have introduced digital twin technology to construct fermentation process mechanism models to improve parameter prediction accuracy. In the field of material handling within the plant, warehouse management systems (WMS) and transportation management systems (TMS) have been widely applied. These systems can complete basic logistics tasks such as raw material warehousing, pipeline transportation, and intermediate tank buffering based on static network topology and fixed scheduling cycles. Intelligent optimization algorithms such as genetic algorithms and ant colony algorithms are also used to plan and solve single logistics paths.

[0003] However, in the existing technological system, fermentation process control and material transfer scheduling have long operated independently, lacking a deep real-time collaboration mechanism, which has become a key bottleneck restricting the overall efficiency improvement of fermentation production. At the data fusion level, there is a serious information silo problem with multi-source heterogeneous data from fermentation DCS systems, logistics WMS / TMS systems, and equipment operation and maintenance systems. Data sovereignty is isolated among systems, and interface standards are inconsistent, making real-time interaction and fusion of cross-system data impossible. Furthermore, real-time fusion suffers from stability issues such as high data noise, high missing rates, and spatiotemporal asynchrony, making it difficult to support high-precision global state perception and real-time decision-making. It is also difficult to construct a multi-scale digital twin foundation covering microscopic biochemical reactions, mesoscopic equipment operation, and macroscopic transfer networks, resulting in the inability to accurately perceive the overall production status.

[0004] At the collaborative decision-making and control level, existing transport route optimization only aims at minimizing transport time and cost, without considering bioprocess parameters such as fermentation cell growth stage, substrate residue, and product synthesis rate as constraints. The low match between material arrival time and fermentation feed requirements easily leads to substrate starvation or inhibition, resulting in decreased product yield. Furthermore, fermentation regulation is still primarily based on feedback control, unable to anticipate upstream disturbances such as transport delays and equipment failures. Adjustments are only made passively after process parameters deviate from standard ranges, resulting in significant control lag. Simultaneously, the lack of a dynamic load allocation mechanism among multiple parallel fermentation reactors leads to uneven allocation of production resources and limited overall capacity utilization. Moreover, existing systems lack sufficient responsiveness in abnormal scenarios. When transport failures occur, they cannot achieve coordinated execution of route rescheduling and emergency fermentation control. They also lack quantitative evaluation of decision-making effects and closed-loop self-optimization mechanisms, and model accuracy gradually decreases with changing operating conditions, making them unsuitable for complex and variable industrial production environments. Summary of the Invention

[0005] One objective of this invention is to propose a real-time collaborative system for dynamic optimization of transport paths and fermentation regulation, in order to solve problems such as data silos in various systems, the disconnect between fermentation regulation and material transport, insufficient coordination and anti-disturbance capabilities, and the lack of a closed-loop self-optimization mechanism.

[0006] A real-time coordinated system for dynamic optimization of transport paths and fermentation regulation according to an embodiment of the present invention includes the following steps: S1. Collect fermentation system operation data and transport network data, construct a multi-scale digital twin base of micro-meso-macro, and complete the real-time fusion and stability processing of multi-source heterogeneous data; S2. Based on the fused full data, complete the perception of the transport-fermentation coupled state, construct a coupled evaluation index system, and realize multi-level risk warning of the system; S3. Construct a multi-objective optimization model with real-time fermentation status as a constraint, solve for the optimal transfer execution scheme, and support real-time rescheduling in abnormal scenarios; S4. Receive the optimization results of the transport path as a feedforward input to realize adaptive regulation of the fermentation process and complete the collaborative load distribution and robust control of multiple reactors. S5. Build a global collaborative decision-making engine to coordinate two types of decision-making instructions: transfer optimization and fermentation regulation. Complete the quantitative evaluation of decision execution effect and realize the closed-loop self-optimization of model and strategy. In this process, S3 and S4 are executed synchronously and iteratively in each scheduling cycle. The transfer path optimization result output by S3 is transmitted to S4 in real time as a feedforward input, and the fermentation status feedback information output by S4 is transmitted back to S3 in real time as a path optimization constraint, thus constructing a two-way closed-loop collaborative control logic for transfer-fermentation.

[0007] Optionally, S1 specifically includes: S11. Collect online sensor data, offline test data, and equipment operating parameters in the fermentation reactor, combine them with the fermentation kinetic mechanism equation to construct a basic mechanism model, introduce a neural network to correct the deviation of the mechanism model, and build a fermentation process mixing mechanism-data fusion model. S12, extract the network topology of raw material warehouse, transfer equipment, conveying pipeline, intermediate storage tank and fermentation workshop, and build a digital model of the transfer system using discrete event simulation and timed Petri net, and characterize the random disturbance factors of the transfer process based on Markov chain; S13. Access data from multi-source heterogeneous industrial systems, use extended Kalman filtering to reduce noise in sensor data online, and use a sliding window algorithm combined with interpolation and neural network prediction to complete missing data; S14. Adopting an edge-cloud collaborative architecture, cross-system data fusion is completed through federated learning, a unified data interface and parameter mapping relationship are established, and a multi-scale digital twin foundation is formed. The multi-scale digital twin base is divided into three layers: the first layer is the fermentation micro-biochemical reaction layer, the second layer is the meso-scale reactor equipment status layer, and the third layer is the macro-scale transport network layer. The spatiotemporal synchronization and linkage of the three-layer model are achieved through timestamp alignment and spatial interpolation. Through the above noise reduction, completion, and timestamp alignment steps, the stability and consistency of real-time fused data are ensured, providing reliable input for subsequent state perception and decision-making.

[0008] Optionally, S2 specifically includes: S21. Extract the core state indicators of the fermentation system and the transport system respectively, and define three types of cross-system coupled correlation indicators: substrate starvation risk index, product degradation risk index and batch synchronization rate, to form a complete coupled evaluation index system. S22. Extract the time series features of the fermentation process and the spatial topology features of the transport network, input them into the spatiotemporal prediction model, and simultaneously output the trend of fermentation parameter changes and the prediction results of transport path delays within a preset time period. S23. The objective weights of each indicator are determined by the entropy weight method, and the system coupling health degree is calculated by combining the fuzzy comprehensive evaluation method. Based on the numerical range of the coupling health degree, the system operation status is divided into four risk levels: normal, attention, warning and emergency. S24. Build a hierarchical decision tree and risk transmission matrix, automatically trigger corresponding level warnings based on the real-time risk level of the system, and push matching preliminary emergency response plans. S2, through the above-mentioned process of indicator extraction, coupling quantification, health calculation and risk classification, completes the comprehensive perception of the transport-fermentation coupling state. The output real-time fermentation state parameters include cell growth rate, substrate concentration, product synthesis rate, fermentation operation stage, cell activity and remaining substrate amount, which are directly used as constraints for S3 path optimization.

[0009] Optionally, S3 specifically includes: S31. Real-time reception of fermentation operation stage, cell activity, remaining substrate amount output by S2 and control deviation information fed back by S4 as path planning constraints; with the shortest transfer time, lowest transfer cost and highest resource utilization rate as logistics optimization objectives, and the minimum substrate concentration fluctuation, shortest fermentation interruption time and maximum product yield as biological optimization objectives, a multi-objective optimization function is constructed, and the decision preference weight of each optimization objective is determined by the analytic hierarchy process. S32. The Pareto optimal solution set for multi-objective optimization is obtained by using the non-dominated sorting genetic algorithm II, and the global optimal transfer execution scheme is selected from the optimal solution set by combining the TOPSIS superior-inferior solution distance method. S33. Build a multi-agent deep reinforcement learning model, construct a joint state space and action space, combine the proximal policy optimization algorithm to complete the model training, and set the material transfer priority according to the different growth stages of fermentation. S34. Set up an event-driven rescheduling trigger mechanism. When the triggering condition is met, start the rolling window optimization algorithm to resolve the transfer scheme within the preset optimization time domain. Use the Hungarian algorithm to complete the secondary matching of transfer units and delivery tasks to achieve real-time rescheduling of transfer paths.

[0010] Optionally, S4 specifically includes: S41. Build a feedforward-feedback composite control architecture, adopt a multi-level cascade feedback control loop, receive predicted data such as the expected arrival time of materials and the total amount of materials transported from the results of the transfer path optimization, and design an adaptive feedforward compensator to realize the advance adjustment of the core process parameters of fermentation feeding, temperature and dissolved oxygen. S42. For multiple fermentation reactors arranged in parallel, game theory and dynamic programming algorithms are introduced to complete dynamic material allocation by combining the real-time status of each reactor, so as to realize load transfer and balanced allocation of production resources among multiple reactors. S43. Based on the multi-level cascade feedback control loop, H∞ robust control and sliding mode control strategies are integrated to form an interference suppression layer, thereby improving the anti-disturbance capability and operational stability of the fermentation control system. S44. When the transfer system experiences a continuous interruption fault exceeding the preset duration, the fermentation emergency operation mode will be automatically switched to ensure continuous operation of fermentation production.

[0011] Optionally, S5 specifically includes: S51. Build a global collaborative decision engine to receive the transfer path optimization plan and fermentation control instructions, complete the conflict verification and coordination of the two types of instructions, output unified and executable collaborative control instructions, and retain the manual intervention interface. S52. Collect real-time running data after command execution, construct a multi-attribute decision evaluation system, and quantify the execution accuracy and application effect of single collaborative decision-making; S53. Collect all historical operation data and use incremental learning algorithms to perform online parameter iterative updates on the multi-scale digital twin model, state prediction model, reinforcement learning strategy model, and feedforward compensator parameter model; S54. Construct an industrial production decision knowledge graph, and use the knowledge graph reasoning algorithm to retrieve similar historical cases as incremental learning prior samples to improve the online convergence efficiency of the policy model and realize the closed-loop self-optimization of the model and policy.

[0012] Optionally, in S23, the system coupling health is calculated by weighted fusion of three coupling-related indicators: substrate starvation risk index, product degradation risk index, and batch synchronization rate. The weights of each indicator are objectively assigned using the entropy weight method. A quantitative mapping relationship is set between the coupling health value range and the four risk levels. The boundary thresholds of each range are determined based on the statistical characteristics of historical operating data and the process safety margin. The coupling health value is positively correlated with the synergistic stability of the transfer-fermentation system. The weights of each indicator determined by the entropy weight method are updated on a rolling basis according to the latest operating data within a preset time window.

[0013] Optionally, the multi-agent deep reinforcement learning model in S33 is configured with a multi-dimensional composite reward function. The reward function includes a transportation energy consumption penalty, a path congestion penalty, a fermentation substrate fluctuation penalty, a fermentation product yield increase reward, an emergency scheduling compliance reward, and a feedforward compensation trigger frequency penalty. The feedforward compensation trigger frequency penalty is calculated based on the activation count of the S4 adaptive feedforward compensator, thereby guiding the transportation path optimization scheme to reduce the disturbance to the fermentation control end. A policy gradient pruning mechanism is introduced during the model training phase to constrain the policy parameter update amplitude and ensure the overall scheduling operation is stable.

[0014] Optionally, the control logic of the adaptive feedforward compensator in S41 is as follows: read the expected arrival time of materials and the total amount of materials transported in the optimization results of the transfer path in real time, and calculate the process compensation amount in combination with the current concentration of the bottom material of the fermenter and the specific growth rate of the cells; when it is predicted that the materials will arrive early, the feeding rate will be gradually reduced; when it is predicted that the materials will arrive late, the dissolved oxygen level will be increased and the fermentation temperature will be finely adjusted simultaneously; the compensation amount can be dynamically and adaptively corrected according to the transfer prediction error.

[0015] The beneficial effects of this invention are: By constructing a multi-scale digital twin foundation and an edge-cloud collaborative data fusion architecture, this project addresses the issues of data silos, poor information exchange, and disconnect between the two business modules in the fermentation production and material transfer systems. It combines noise reduction, completion, and spatiotemporal synchronization technologies to achieve cross-system heterogeneous data governance, enabling spatiotemporal synchronization and linkage of three-layer models: microscopic biochemical reactions, mesoscopic equipment status, and macroscopic transfer network. This achieves a digital mapping of the entire production status, effectively solving the stability problem of real-time multi-source heterogeneous data fusion and providing a reliable data foundation for high-precision status perception and decision-making. A complete transfer-fermentation coupled status perception system has been established. Through quantification of coupled evaluation indicators and a four-level risk classification and early warning system, comprehensive, real-time, and accurate perception of the global production status is achieved, allowing for early identification and avoidance of production risks. A combination of multi-objective optimization and differentiated scheduling maximizes fermentation production benefits while ensuring transfer efficiency. Optional deep reinforcement learning can be used to adapt scheduling strategies to the fermentation growth stage, further improving collaborative accuracy. An event-driven mechanism enables real-time rescheduling for abnormal scenarios such as faults and congestion. A feedforward-feedback composite control architecture was constructed, combining robust control and sliding mode control techniques to suppress different types of disturbances. Game theory and dynamic programming were used to achieve balanced load distribution among multiple reactors. An adaptive feedforward compensator was used to proactively adjust process parameters based on transport prediction data, and an emergency operation mode was set up to ensure continuous production during transport interruptions, significantly improving the fermentation system's resistance to disturbances. Finally, a global collaborative decision engine was built to complete instruction conflict verification and overall management. A multi-attribute evaluation system was used to quantify the decision execution effect, and incremental learning and knowledge graph reasoning were combined to complete online iterative optimization of various models. A two-way closed-loop collaborative control logic for transport and fermentation was constructed, ultimately achieving deep real-time collaboration between dynamic optimization of material transport paths and adaptive regulation of the fermentation process. This effectively reduced process parameter fluctuations, avoided production risks, and improved product yield and resource utilization, while also giving the system strong anti-interference capabilities and closed-loop self-optimization capabilities for models and strategies. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a real-time collaborative system for dynamic optimization of transport paths and fermentation regulation proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1A real-time collaborative system for dynamic optimization of transport pathways and fermentation regulation includes the following steps: S1. Collect fermentation system operation data and transport network data, construct a multi-scale digital twin base of micro-, meso-, and macro-scale, and complete the real-time fusion of multi-source heterogeneous data.

[0019] S11. Collect online sensor data, offline laboratory data, and equipment operating parameters within the fermentation reactor. Combine this with the fermentation kinetic mechanism equation to construct a basic mechanistic model of cell growth, substrate consumption, and product formation. Introduce an attention mechanism LSTM network to correct operational deviations in the mechanistic model, and build a fermentation process mixing mechanism-data fusion model. The online sensor data includes fermentation temperature, pH, dissolved oxygen concentration, stirring speed, tank pressure, and exhaust CO2 concentration, with a sampling frequency of 1 time / second. The offline laboratory data includes cell concentration, substrate concentration, and product concentration, with a sampling frequency of 1 time / 2 hours. The fermentation kinetic mechanism equation uses the Monod equation to describe the specific cell growth rate and the Luedeking-Piret equation to describe the product formation rate. The 12 input features include fermentation temperature, pH, dissolved oxygen concentration, stirring speed, tank pressure, exhaust CO2 concentration, cell concentration, substrate concentration, product concentration, feed rate, ventilation volume, and cooling water flow rate. The Monod equation has the following specific form: ,in The specific growth rate of the cells. Maximum specific growth rate (range of values) In this embodiment, ), The value is the substrate half-saturation constant (ranging from 5 to 20 g / L, and 10 g / L in this example). The real-time concentration of fermentable substrate within the fermentation system (in this example, the glucose concentration in corn starch hydrolysate) is the core variable determining the specific growth rate of the cells; corresponding to this... The substrate half-saturation constant refers to the substrate concentration at which the cell growth rate reaches half of its maximum value. The specific form of the Luedeking-Piret equation is: ; in The concentration of the product. This refers to the bacterial cell concentration. The synthesis coefficient of growth-related products is 0.8 in this embodiment. The synthesis coefficient of non-growth-related products (in this embodiment, it is...) The attention mechanism LSTM network has an input feature dimension of 12, a hidden layer dimension of 64, a time step of 20, and uses the Adam optimizer with a learning rate of 0.001 for model training. This optimizer is used to correct the prediction bias caused by parameter drift and environmental disturbances in the mechanism model. The proportion of the mechanism model's predicted value with a relative error of less than 5% compared to the actual value reaches more than 95%.

[0020] S12 extracts the network topology relationships of raw material warehouses, transfer equipment, pipelines, intermediate storage tanks, and fermentation workshops. A digital model of the transfer system is built using discrete event simulation and timed Petri nets. Markov chains are used to characterize random disturbances such as equipment failures and path congestion during the transfer process. The transfer equipment includes AGVs, belt conveyors, and pipeline pumps; the number of devices and operating parameters are configured according to the actual production scale. The timed Petri nets have 52 storage nodes across 5 categories: raw material inventory status, AGV idle status, pipeline occupancy status, intermediate storage tank level status, and fermentation tank replenishment status. Transition nodes have 38 nodes across 6 categories: raw material outbound, AGV start-up, pipeline switching, material conveying, and replenishment initiation, used to accurately describe the temporal logic and resource constraints of material transfer. The Markov chain state transition matrix has a dimension of 12×12 and includes four core states: normal equipment, faulty equipment, idle path, and congested path. The matrix elements are calculated by statistically analyzing the duration and frequency of each state over the past 18 months, and are used to predict the probability and duration of random disturbances.

[0021] S13. Integrate multi-source heterogeneous data from fermentation DCS system, logistics WMS / TMS system, equipment operation and maintenance system, and edge acquisition terminal. Use extended Kalman filter to reduce noise online for sensor data of fermentation temperature, pH value, dissolved oxygen concentration, and stirring speed. Use sliding window algorithm to complete missing data. When the proportion of missing data in the window is lower than the first preset proportion threshold, use linear interpolation to complete it. When it is higher than the second preset proportion threshold, use LSTM network based on independent training attention mechanism to predict and complete it. When the missing proportion falls between two thresholds, a combination of linear interpolation and an LSTM network is used for imputation. The first preset proportion threshold is 10%, and the second preset proportion threshold is 30%. The sliding window duration is 10 minutes, and the sliding step size is 1 minute. The state variables of the extended Kalman filter are... , representing fermentation temperature, pH value, dissolved oxygen concentration, and stirring speed, respectively; the observation equation is: ,in It is the identity matrix. Indicates the first The system state vector at the sampling time. To mitigate noise, the process noise covariance is set to 0.01, and the measurement noise covariance is set to 0.1, effectively suppressing high-frequency noise and random interference from the sensor. The sliding window data completion logic is as follows: when the proportion of missing data within the window is less than 10%, linear interpolation between two adjacent valid data points is used for completion. When the missing proportion is greater than 30%, the historical data of the first 3 windows are input into the independently trained attention mechanism LSTM network for prediction and completion; when the missing proportion is between 10% and 30%, the linear interpolation result and the LSTM prediction result are weighted and averaged at a ratio of 7:3.

[0022] S14. Employing an edge-cloud collaborative architecture, under the constraint of multi-entity data sovereignty isolation, a federated averaging algorithm is used to achieve cross-system multi-source heterogeneous data fusion. Each system completes model training locally, transmitting only model parameter updates to the cloud. A model structure adaptation layer eliminates model differences between heterogeneous systems, establishing a unified data interface and parameter mapping relationship, forming a multi-scale digital twin foundation encompassing fermentation, transport, and environmental states. The local model of the federated averaging algorithm uses a 3-layer fully connected neural network with an input dimension of 12, a hidden layer dimension of 32, and an output dimension of 6. The parameter aggregation method is a weighted average based on the proportion of local data volume in each system to the total data volume. The model structure adaptation layer includes a dimension transformation layer and a feature alignment layer. The dimension transformation layer maps different dimensional features output by different systems into a unified 24-dimensional vector. The feature alignment layer achieves heterogeneous feature alignment by minimizing the KL divergence of feature distributions in different systems. The multi-scale digital twin base is divided into three layers: the first layer is a fermentation micro-biochemical reaction layer with a time granularity of seconds to minutes, used to simulate the dynamic processes of cell growth and substrate metabolism; the second layer is a meso-scale reactor equipment status layer with a time granularity of minutes, used to monitor the operating status of reactors, agitators, and heat exchangers; the third layer is a macro-transfer network layer with a time granularity of minutes to hours, used to simulate the material flow and resource allocation of the entire plant area. Through a cross-scale parameter transfer algorithm of timestamp alignment and spatial interpolation, the three-layer model is synchronized and linked in time and space, with a parameter transfer delay of less than 100 milliseconds; the global model update cycle of the federated averaging algorithm is 15 minutes, and the training batch size of each local model is 32.

[0023] S2. Based on the fused full data, the transport-fermentation coupled state perception is completed, a coupled evaluation index system is constructed, and a spatiotemporal prediction model is used to perform joint state prediction to achieve multi-level risk warning of the system.

[0024] S21. Extract fermentation system status indicators including cell growth rate deviation, substrate concentration fluctuation range, product synthesis rate, and transport system status indicators including transport time delay deviation, equipment availability, and path saturation. Define cross-system coupling correlation indicators, including substrate starvation risk index, product degradation risk index, and batch synchronization rate, to form a complete cross-system coupling evaluation index system.

[0025] The formula for calculating the substrate starvation risk index is as follows: ; in This represents the current substrate concentration. This is the minimum substrate concentration required for bacterial growth (15 g / L, determined by experiments on the growth characteristics of the bacterial strain). The formula for calculating the degradation risk index of the product is as follows: ; in Due to delays in material transfer, The critical delay time at which the product begins to degrade significantly (2 hours; when the material delay exceeds 2 hours, the product degradation rate will increase by more than 3 times); the batch synchronization rate is the percentage of reactors in the same production batch whose fermentation progress deviation is less than ±5%.

[0026] S22. Extract the time-series features of the fermentation process and the spatial topology features of the transport network, and input them into the spatiotemporal prediction model. Integrate a graph convolutional network and a gated recurrent unit (GRU) to complete multi-dimensional feature extraction. Employ a rolling time-domain prediction method to simultaneously output the trends of fermentation parameter changes and the predicted results of transport path delays within a preset time period. The spatiotemporal prediction model adopts a GCN-GRU structure, containing a 2-layer graph convolutional network (GCN) and a 2-layer gated recurrent unit (GRU). The GCN is used to extract the spatial topology features of the transport network, and the GRU is used to extract the time-series features of the fermentation process. The loss function is mean squared error (MSE), the optimizer is Adam, the learning rate is 0.001, the batch size is 32, and the training set to test set ratio is 8:2. The preset time period is 24 hours, and the prediction step size is 15 minutes. The window update cycle of the rolling time-domain prediction is 15 minutes, and each update retains the historical data from the previous 12 hours for model calibration.

[0027] S23. The objective weights of each evaluation index are determined using the entropy weight method. The system coupling health degree is calculated using the fuzzy comprehensive evaluation method. Based on the quantified numerical range of the coupling health degree, the system operating status is divided into four risk levels: normal, attention, warning, and emergency. The fuzzy comprehensive evaluation method uses a triangular membership function, with an evaluation set of {excellent, good, moderate, poor, very poor}, corresponding to coupling health degree ranges of [0.8, 1.0], [0.6, 0.8), [0.4, 0.6), [0.2, 0.4), and [0, 0.2], respectively.

[0028] The fuzzy relation matrix is ​​constructed by statistically analyzing the correspondence between the values ​​of various indicators and the actual operating status of the system over the past 24 months. Matrix elements represent the membership degree of a certain indicator value to a certain evaluation level. The quantitative mapping relationship between the coupled health value range and the risk level is as follows: a health value of [0.8, 1.0] corresponds to a normal level, [0.6, 0.8) corresponds to a concern level, [0.4, 0.6) corresponds to a warning level, and [0, 0.4) corresponds to an emergency level. The boundary thresholds for each range are determined based on the statistical characteristics of historical operating data over the past 24 months and the process safety margin. The weights of each indicator determined by the entropy weight method are updated on a rolling basis according to the latest operating data within a 7-day time window, allowing the weight allocation to adaptively adjust with changes in system operating conditions. The mapping relationship between the fuzzy comprehensive evaluation result and the risk level is as follows: an evaluation result of "excellent" or "good" corresponds to a normal level, "medium" corresponds to a concern level, "poor" corresponds to a warning level, and "very poor" corresponds to an emergency level.

[0029] S24. Construct a hierarchical decision tree and risk transmission matrix to quantify the impact weight and transmission degree of abnormal transport-fermentation conditions on the fermentation production process. Automatically trigger corresponding level warnings based on the system's real-time risk level and push matching preliminary emergency response plans. The hierarchical decision tree contains three layers of decision nodes and a total of 16 decision paths. The first layer of decision nodes represents the coupling health degree, the second layer represents the risk type, and the third layer represents the degree of impact.

[0030] For example, when the coupling health score is <0.4, the risk type is substrate starvation, and the impact level is severe, an emergency warning is triggered. The risk transmission matrix has a dimension of 12×8, which can quantify the transmission relationship between 12 typical transport anomalies and 8 typical fermentation anomalies. Among them, the transmission weight of transport equipment failure on fermentation interruption is 0.9, and the transmission weight of path congestion on substrate fluctuation is 0.7. Warning methods include system pop-ups, audible and visual alarms, and SMS push notifications. There is no warning for the normal level, a system pop-up is pushed for the attention level, an audible and visual alarm is triggered for the warning level, and an emergency level triggers both an audible and visual alarm and an SMS push notification to the responsible person.

[0031] S3. Construct a multi-objective optimization function with fermentation state parameters as constraints, use an evolutionary algorithm to solve the Pareto solution set for multi-objective optimization of transport paths, and combine deep reinforcement learning to design a differentiated scheduling strategy adapted to the fermentation growth stage; support real-time rescheduling in abnormal scenarios based on event-driven mechanism and rolling optimization algorithm.

[0032] S31. Real-time reception of fermentation process parameters, control deviation information, and fermentation status assessment results from S4. The fermentation operation stage, cell activity, remaining substrate amount, and S4 feedback parameters are used as path planning constraints. The logistics optimization objectives are the shortest transfer time, the lowest transfer cost, and the highest resource utilization rate. The biological optimization objectives are the minimum substrate concentration fluctuation, the shortest fermentation interruption time, and the maximum product yield. A multi-objective optimization function is constructed, and the decision preference weights of each optimization objective are determined through the analytic hierarchy process (AHP) for subsequent screening of the global optimal solution in the Pareto solution set.

[0033] The expression for the multi-objective optimization function is: ; in Transit time (unit: minutes). For transportation costs (unit: yuan). For resource utilization, The standard deviation of substrate concentration. The duration of fermentation interruption (in minutes). The yield of the product; to The decision preference weights for each objective are determined using the analytic hierarchy process (AHP). The default weights are 0.15, 0.1, 0.1, 0.25, 0.2, and 0.2, respectively, and can be manually adjusted according to production needs.

[0034] The resource utilization rate The calculation formula is: ; in For the first The actual usage time of the transfer equipment. Total available time, This refers to the total number of all transfer equipment, including AGVs, transfer pumps, pipeline valves, etc.

[0035] S32. The Pareto optimal solution set is obtained using the non-dominated sorting genetic algorithm II. Combined with the decision preference weights determined by the analytic hierarchy process in S31, the globally optimal transfer execution scheme is selected from the optimal solution set using the TOPSIS distance method. The non-dominated sorting genetic algorithm II employs a two-point crossover operator with a crossover probability of 0.8; a multinomial mutation operator with a mutation probability of 0.1 and a mutation distribution exponent of 20; a tournament selection operator with a tournament size of 2; a population size of 100; and 50 iterations. The TOPSIS method uses Euclidean distance to calculate the closeness between each Pareto solution and the ideal solution; the solution with the highest closeness is the globally optimal scheme.

[0036] S33. Construct a multi-agent deep reinforcement learning model with the transfer execution unit as the intelligent agent. This model includes a joint state space and a path selection action space encompassing fermentation operation status, transfer resource status, and road network operation status. The model is trained using a proximal strategy optimization algorithm, and material transfer priorities are set according to the different growth stages of fermentation. The 24-dimensional joint state space includes features such as cell growth rate, remaining substrate quantity, AGV position, path congestion level, and intermediate storage tank level. The action space includes 8 preset transfer paths and 3 transfer speed levels (low speed 0.5 m / s, medium speed 1.0 m / s, high speed 1.5 m / s).

[0037] The expression for the multi-dimensional composite reward function is as follows: ; in This is an incentive for increased production of fermentation products. This is a compliance reward item for emergency dispatch. As a penalty item for energy consumption during transportation, For route congestion penalties, This is a penalty term for fluctuations in the fermentation substrate. This is a penalty term for the frequency of feedforward compensation triggering.

[0038] The calculation methods for each term in the reward function are as follows: ; in This represents the actual product yield. Based on the benchmark yield; ,in Energy consumption for transport (kWh); ,in The number of times feedforward compensation is triggered. The pruning threshold of the near-end strategy optimization algorithm is 0.2, the learning rate is 0.0003, and the batch size is 64; the material transfer priority is set as follows: fermentation logarithmic growth phase > fermentation stationary phase > fermentation lag phase > fermentation decay phase.

[0039] S34. An event-driven rescheduling trigger mechanism is set up. Triggering conditions include fermentation state parameters deviating from the process standard range by more than ±10%, transfer equipment malfunction and shutdown, and transfer path congestion delay exceeding a preset time threshold. After the triggering conditions are met, a rolling window optimization algorithm is started to resolve the transfer scheme within a preset optimization time domain. The Hungarian algorithm is used to complete the secondary matching of transfer units and transport tasks, achieving real-time rescheduling of the transfer path within a preset scheduling delay range. The preset time threshold is 30 minutes; the preset optimization time domain is 2 hours; the sliding step size of the rolling window optimization algorithm is 15 minutes; and the preset scheduling delay is 5 minutes, meaning the time from rescheduling trigger to the issuance of the new scheme does not exceed 5 minutes. The input matrix of the Hungarian algorithm is the cost matrix, and the matrix elements... Representing the Taiwan AGV execution The total cost of the task is calculated by weighting the transit time, transit energy consumption, and route congestion penalties.

[0040] S4. Receive the optimized transfer path results, extract predicted data such as the expected arrival time of materials and the total amount of materials transported as feedforward inputs, build a feedforward-feedback composite control architecture, realize adaptive regulation of the fermentation process, and complete the collaborative load distribution and robust control of multiple reactors.

[0041] S41. Construct a feedforward-feedback composite control architecture, using an outermost model predictive control, an intermediate fuzzy PID control, and an inner loop actuator loop to form a multi-level cascade feedback control loop. Receive the optimized transfer path results, extract predicted data such as the expected arrival time of materials and the total amount of materials transported, and calculate the process compensation amount by combining the current fermenter substrate concentration and cell growth rate. Design an adaptive feedforward compensator to achieve advanced adjustment of core process parameters such as fermentation feeding, temperature, and dissolved oxygen.

[0042] The state variables for the model predictive control are substrate concentration, cell concentration, and product concentration; the control variables are feed rate, ventilation rate, and stirring speed. The constraints are feed rate ∈ [0, 100] L / h, ventilation rate ∈ [0, 200] m³ / h, and stirring speed ∈ [0, 300] rpm. The objective function is to minimize the deviation between the substrate concentration and the setpoint. The prediction time domain is 10, the control time domain is 3, and the sampling period is 1 minute. The fuzzy PID control uses a triangular membership function containing 7 fuzzy subsets, and can adaptively adjust the PID parameters according to the nonlinear characteristics of the fermentation process. Example of fuzzy rule representation: When the error is large and the error change rate is large, the proportional coefficient increases, the integral coefficient decreases, and the derivative coefficient increases; when the error is large and the error change rate is large, the proportional coefficient decreases, the integral coefficient increases, and the derivative coefficient decreases. The control logic of the adaptive feedforward compensator is as follows: real-time reading of the expected arrival time of materials and the total amount of materials transported from the optimized transfer path results, combined with the current concentration of the fermenter bottom material and the specific growth rate of the cells, to calculate the process compensation amount.

[0043] The formula for calculating the adjustment amount of the feeding rate is: ; in This is the difference (in hours) between the actual arrival time and the planned arrival time of the material; delays are positive, and early arrivals are negative. The compensation coefficient is 0.05 in this embodiment. This refers to the volume of the fermentation broth; The formula for calculating dissolved oxygen adjustment is: ; in The compensation coefficient is 2% / hour in this embodiment. When the predicted material is delivered more than 15 minutes ahead of schedule ( The feeding rate is gradually reduced by 12%-22% to avoid substrate inhibition and cell overload; when a material delivery delay of more than 30 minutes is predicted ( Simultaneously, it increases dissolved oxygen levels by 6%-12% and fine-tunes fermentation temperature by 1-2℃ to slow down cell metabolism; the compensation amount can be dynamically and adaptively corrected according to transport prediction errors, with a correction cycle of 5 minutes.

[0044] S42. For multiple parallel fermentation reactors, game theory and dynamic programming algorithms are introduced. Combining the real-time fermentation stage, product value weights, and substrate shortage of each reactor, dynamic material allocation is achieved to realize load transfer and balanced allocation of production resources among the reactors. The participants in the non-cooperative game model are 12 parallel fermentation reactors, and the strategy space represents the material allocation ratio of each reactor. The payoff function is ; in For the first Product yield of the reactor. This represents the fluctuation range of substrate concentration. and The weighting coefficients are used. The optimal material allocation scheme is obtained by iteratively solving the Nash equilibrium. The state dimension of the dynamic programming algorithm is the number of reactors × 2, and the decision dimension is the material allocation ratio, which is used to optimize the global material allocation strategy for multi-batch continuous production.

[0045] S43. Based on the multi-level cascade feedback control loop described in S41, the middle layer fuzzy PID control adapts to the nonlinear fluctuation characteristics of the fermentation process, and the outermost layer integrates H∞ robust control and sliding mode control strategies to form an interference suppression layer. Robust control suppresses low-frequency disturbances caused by uncertain transfer time, and sliding mode control addresses high-frequency disturbances caused by fluctuations in material supply, thereby improving the anti-disturbance capability and operational stability of the fermentation control system.

[0046] The sensitivity function weighting function of the H∞ robust control is: ; The weighting function of the supplementary sensitivity function is With an interference suppression index γ≤0.5, it can effectively suppress low-frequency disturbances with amplitude less than ±20%.

[0047] The sliding surface is designed as follows: ; in For substrate concentration bias, The rate of change of deviation , The exponential reaching law is adopted, with the reaching law parameter k=10. =0.5, which can suppress high-frequency disturbances with amplitude less than ±30% within 0.5 minutes; where This is a Laplace complex frequency domain operator used for frequency domain modeling of transfer functions.

[0048] S44. When the transfer system experiences a continuous interruption exceeding the preset duration, the system automatically switches to the fermentation emergency operation mode, adjusts process parameters to extend the cell metabolism cycle, and simultaneously calls upon the reserve materials in the backup storage tank to complete emergency feeding, ensuring continuous fermentation production. The preset duration is 2 hours; the process parameters of the emergency operation mode are adjusted as follows: fermentation temperature reduced by 2-3℃, stirring speed reduced by 10%-15%, and dissolved oxygen concentration maintained at 30%-40%; the reserve material in the backup storage tank can meet the normal feeding needs of all fermentation reactors for 4 hours.

[0049] S5. Build a global collaborative decision-making engine to coordinate two types of decision-making instructions: transfer optimization and fermentation regulation. Complete the quantitative evaluation of decision execution effect and achieve closed-loop self-optimization of model and strategy based on incremental learning.

[0050] S51. Build a global collaborative decision engine to receive the transfer path optimization plan and fermentation control instructions. Based on time window overlap detection and resource occupation mutual exclusion analysis, complete the conflict verification of the two types of instructions. When a conflict occurs, the principle of fermentation production continuity priority is higher than transfer efficiency to coordinate instructions. Combined with the global production benefit assessment, output unified and executable collaborative control instructions, while retaining the manual intervention interface.

[0051] The time window overlap detection algorithm compares the execution time intervals of two instructions. and ,like and If multiple instructions occupy the same resource at the same time, it is considered an overlap. The resource occupancy mutual exclusion analysis is achieved by maintaining a resource occupancy schedule. When multiple instructions occupy the same resource at the same time, it is considered mutually exclusive. The conflict coordination rules are: fermentation feeding instruction priority > material transfer instruction priority > equipment maintenance instruction priority; instructions with the same priority are sorted from high to low production efficiency. The manual intervention interface supports instruction modification, priority adjustment, and emergency stop operations. All manual operations are recorded and used for subsequent model training.

[0052] S52. Real-time collection of fermentation process parameters, logistics operation status, production energy consumption, and product yield data after command execution; construction of a multi-attribute decision evaluation system; comparison of parameter prediction values ​​with actual operating values; quantification of the execution accuracy and application effect of a single collaborative decision. The multi-attribute decision evaluation system includes four primary indicators and twelve secondary indicators: decision accuracy rate, execution delay rate, parameter deviation rate, and production efficiency improvement rate.

[0053] The formula for calculating the decision accuracy is as follows: ; in This represents the total number of collaborative decision-making predictions made by the system within the statistical period. The number of times the predicted value deviates from the actual value by less than 5%; The formula for calculating the execution latency rate is: ; in: Represents the execution latency rate; This represents the total number of collaborative decision-making instructions within the statistical period. This refers to the sequence number of a single decision instruction, with a value range of... ; For the first The actual execution time of each instruction; For the first The standard execution time for each instruction is pre-planned; The numerator is the total delay time when the actual execution time of all instructions exceeds the planned execution time, and the denominator is the sum of the planned execution times of all instructions.

[0054] The formula for calculating the productivity improvement rate is: ; in: Represents the rate of increase in production efficiency; This represents the actual production revenue per cycle under the collaborative scheduling mode of this invention. This represents the benchmark production revenue for the same period under the traditional manual scheduling model.

[0055] The accuracy of decision execution is expressed as the relative error between the predicted value and the actual value, and the relative error of the core parameters is required to be less than 5%.

[0056] S53. Collect all historical operational data, decision-making data, and anomaly handling data, and use an incremental learning algorithm to perform online parameter iterative updates on the multi-scale digital twin model, state prediction model, reinforcement learning strategy model, and feedforward compensator parameter model. The incremental learning algorithm employs an elastic weight consolidation algorithm to prevent the model from forgetting historical knowledge when learning new data.

[0057] Importance weights are calculated using the Fisher information matrix, and the loss function is... ; in The loss function for the new data, This is the regularization coefficient (1000 in this example). For the first Fisher information values ​​for each parameter These are the parameter values ​​for the old model. The model is updated online every 24 hours, with each update using incremental data from the past 7 days. The update process does not affect the normal operation of the system.

[0058] S54. Construct an industrial production decision-making knowledge graph. Entity nodes in the knowledge graph include operating condition types, process parameter combinations, decision schemes, and disposal results. Relationship edges include causal relationships, temporal relationships, and similarity relationships. A knowledge graph reasoning algorithm integrating graph embedding and attention mechanisms is employed to calculate the similarity between the decision-making scenario and historical cases. Typical historical cases with similarity exceeding a preset threshold are retrieved and used as incremental learning prior samples to improve the online convergence efficiency of the strategy model and achieve closed-loop self-optimization of the model and strategy. Entities and relationships in the knowledge graph are automatically extracted from historical production data using a BERT-based named entity recognition and relationship extraction model. The knowledge graph contains over 1200 entity nodes and over 3500 relationship edges, and is automatically updated monthly. The preset similarity threshold is 0.85.

[0059] The similarity calculation uses cosine similarity, and the formula is as follows: ; in and These are the graph embedding vectors for two scenes; the graph embedding algorithm uses the TransE model with an embedding dimension of 128.

[0060] In this invention, S3 and S4 are executed synchronously and iteratively in each scheduling cycle, which is 15 minutes. The transfer path optimization result output by S3 is transmitted to S4 in real time as a feedforward input, and the fermentation status feedback information output by S4 is transmitted back to S3 in real time as a path optimization constraint, thus constructing a two-way closed-loop collaborative control logic for transfer-fermentation.

[0061] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of the real-time coordinated method for dynamic optimization of transport paths and real-time control of fermentation as described above. The computer-readable storage medium can be a non-volatile storage medium such as a hard disk, USB flash drive, optical disk, or flash memory, or a network storage device such as a server or cloud storage platform. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0062] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A real-time collaborative system for dynamic optimization of transport paths and fermentation regulation, characterized in that, Includes the following steps: S1. Collect fermentation system operation data and transport network data, construct a multi-scale digital twin base of micro-meso-macro, and complete the real-time fusion and stability processing of multi-source heterogeneous data; S2. Based on the fused full data, complete the perception of the transport-fermentation coupled state, construct a coupled evaluation index system, and realize multi-level risk warning of the system; S3. Construct a multi-objective optimization model with real-time fermentation status as a constraint, solve for the optimal transfer execution scheme, and support real-time rescheduling in abnormal scenarios; S4. Receive the optimization results of the transport path as a feedforward input to realize adaptive regulation of the fermentation process and complete the collaborative load distribution and robust control of multiple reactors. S5. Build a global collaborative decision-making engine to coordinate two types of decision-making instructions: transfer optimization and fermentation regulation. Complete the quantitative evaluation of decision execution effect and realize the closed-loop self-optimization of model and strategy. In this process, S3 and S4 are executed synchronously and iteratively in each scheduling cycle. The transfer path optimization result output by S3 is transmitted to S4 in real time as a feedforward input, and the fermentation status feedback information output by S4 is transmitted back to S3 in real time as a path optimization constraint, thus constructing a two-way closed-loop collaborative control logic for transfer-fermentation.

2. The real-time collaborative system for dynamic optimization of transport paths and fermentation regulation according to claim 1, characterized in that, S1 specifically includes: S11. Collect online sensor data, offline test data, and equipment operating parameters in the fermentation reactor, combine them with the fermentation kinetic mechanism equation to construct a basic mechanism model, introduce a neural network to correct the deviation of the mechanism model, and build a fermentation process mixing mechanism-data fusion model. S12, extract the network topology of raw material warehouse, transfer equipment, conveying pipeline, intermediate storage tank and fermentation workshop, and build a digital model of the transfer system using discrete event simulation and timed Petri net, and characterize the random disturbance factors of the transfer process based on Markov chain; S13. Access data from multi-source heterogeneous industrial systems, use extended Kalman filtering to reduce noise in sensor data online, and use a sliding window algorithm combined with interpolation and neural network prediction to complete missing data; S14. Adopting an edge-cloud collaborative architecture, cross-system data fusion is completed through federated learning, a unified data interface and parameter mapping relationship are established, and a multi-scale digital twin foundation is formed. The multi-scale digital twin base is divided into three layers: the first layer is the fermentation micro-biochemical reaction layer, the second layer is the meso-scale reactor equipment status layer, and the third layer is the macro-scale transport network layer. The spatiotemporal synchronization and linkage of the three-layer model are achieved through timestamp alignment and spatial interpolation. Through the above noise reduction, completion, and timestamp alignment steps, the stability and consistency of real-time fused data are ensured, providing reliable input for subsequent state perception and decision-making.

3. The real-time collaborative system for dynamic optimization of transport paths and fermentation regulation according to claim 1, characterized in that, S2 specifically includes: S21. Extract the core state indicators of the fermentation system and the transport system respectively, and define three types of cross-system coupled correlation indicators: substrate starvation risk index, product degradation risk index and batch synchronization rate, to form a complete coupled evaluation index system. S22. Extract the time series features of the fermentation process and the spatial topology features of the transport network, input them into the spatiotemporal prediction model, and simultaneously output the trend of fermentation parameter changes and the prediction results of transport path delays within a preset time period. S23. The objective weights of each indicator are determined by the entropy weight method, and the system coupling health degree is calculated by combining the fuzzy comprehensive evaluation method. Based on the numerical range of the coupling health degree, the system operation status is divided into four risk levels: normal, attention, warning and emergency. S24. Build a hierarchical decision tree and risk transmission matrix, automatically trigger corresponding level warnings based on the real-time risk level of the system, and push matching preliminary emergency response plans. S2, through the above-mentioned process of indicator extraction, coupling quantification, health calculation and risk classification, completes the comprehensive perception of the transport-fermentation coupling state. The output real-time fermentation state parameters include cell growth rate, substrate concentration, product synthesis rate, fermentation operation stage, cell activity and remaining substrate amount, which are directly used as constraints for S3 path optimization.

4. The real-time collaborative system for dynamic optimization of transport paths and fermentation regulation according to claim 1, characterized in that, S3 specifically includes: S31. Real-time reception of fermentation operation stage, cell activity, remaining substrate amount output by S2 and control deviation information fed back by S4 as path planning constraints; with the shortest transfer time, lowest transfer cost and highest resource utilization rate as logistics optimization objectives, and the minimum substrate concentration fluctuation, shortest fermentation interruption time and maximum product yield as biological optimization objectives, a multi-objective optimization function is constructed, and the decision preference weight of each optimization objective is determined by the analytic hierarchy process. S32. The Pareto optimal solution set for multi-objective optimization is obtained by using the non-dominated sorting genetic algorithm II, and the global optimal transfer execution scheme is selected from the optimal solution set by combining the TOPSIS superior-inferior solution distance method. S33. Build a multi-agent deep reinforcement learning model, construct a joint state space and action space, combine the proximal policy optimization algorithm to complete the model training, and set the material transfer priority according to the different growth stages of fermentation. S34. Set up an event-driven rescheduling trigger mechanism. When the triggering condition is met, start the rolling window optimization algorithm to resolve the transfer scheme within the preset optimization time domain. Use the Hungarian algorithm to complete the secondary matching of transfer units and delivery tasks to achieve real-time rescheduling of transfer paths.

5. The real-time collaborative system for dynamic optimization of transport paths and fermentation regulation according to claim 1, characterized in that, S4 specifically includes: S41. Build a feedforward-feedback composite control architecture, adopt a multi-level cascade feedback control loop, receive predicted data such as the expected arrival time of materials and the total amount of materials transported from the results of the transfer path optimization, and design an adaptive feedforward compensator to realize the advance adjustment of the core process parameters of fermentation feeding, temperature and dissolved oxygen. S42. For multiple fermentation reactors arranged in parallel, game theory and dynamic programming algorithms are introduced to complete dynamic material allocation by combining the real-time status of each reactor, so as to realize load transfer and balanced allocation of production resources among multiple reactors. S43. Based on the multi-level cascade feedback control loop, H∞ robust control and sliding mode control strategies are integrated to form an interference suppression layer, thereby improving the anti-disturbance capability and operational stability of the fermentation control system. S44. When the transfer system experiences a continuous interruption fault exceeding the preset duration, the fermentation emergency operation mode will be automatically switched to ensure continuous operation of fermentation production.

6. The real-time collaborative system for dynamic optimization of transport paths and fermentation regulation according to claim 1, characterized in that, S5 specifically includes: S51. Build a global collaborative decision engine to receive the transfer path optimization plan and fermentation control instructions, complete the conflict verification and coordination of the two types of instructions, output unified and executable collaborative control instructions, and retain the manual intervention interface. S52. Collect real-time running data after command execution, construct a multi-attribute decision evaluation system, and quantify the execution accuracy and application effect of single collaborative decision-making; S53. Collect all historical operation data and use incremental learning algorithms to perform online parameter iterative updates on the multi-scale digital twin model, state prediction model, reinforcement learning strategy model, and feedforward compensator parameter model; S54. Construct an industrial production decision knowledge graph, and use the knowledge graph reasoning algorithm to retrieve similar historical cases as incremental learning prior samples to improve the online convergence efficiency of the policy model and realize the closed-loop self-optimization of the model and policy.

7. The real-time collaborative system for dynamic optimization of transport paths and fermentation regulation according to claim 3, characterized in that, In S23, the system coupling health is calculated by weighted fusion of three coupling correlation indicators: substrate starvation risk index, product degradation risk index and batch synchronization rate. The weight of each indicator is objectively assigned by the entropy weight method. A quantitative mapping relationship is established between the coupled health value range and the four risk levels. The boundary thresholds of each range are determined based on the statistical characteristics of historical operating data and the process safety margin. The coupled health value is positively correlated with the synergistic stability of the transfer-fermentation system. The weights of each indicator determined by the entropy weight method are updated on a rolling basis based on the latest operating data within a preset time window.

8. The real-time collaborative system for dynamic optimization of transport paths and fermentation regulation according to claim 4, characterized in that, The multi-agent deep reinforcement learning model in S33 is configured with a multi-dimensional composite reward function, which includes a transportation energy consumption penalty, a path congestion penalty, a fermentation substrate fluctuation penalty, a fermentation product yield increase reward, an emergency scheduling compliance reward, and a feedforward compensation trigger frequency penalty. The feedforward compensation trigger frequency penalty is calculated based on the activation count of the S4 adaptive feedforward compensator, thereby guiding the transportation path optimization scheme to reduce the disturbance to the fermentation control end. A policy gradient pruning mechanism is introduced during the model training phase to constrain the policy parameter update magnitude and ensure the overall scheduling and operation stability.

9. The real-time collaborative system for dynamic optimization of transport paths and fermentation regulation according to claim 5, characterized in that, The control logic of the adaptive feedforward compensator in S41 is as follows: real-time reading of the expected arrival time of materials and the total amount of materials transported in the optimization results of the transfer path, combined with the current concentration of the bottom material in the fermenter and the specific growth rate of the cells, to calculate the process compensation amount; when it is predicted that the materials will arrive early, the feeding rate will be gradually reduced; when it is predicted that the materials will arrive late, the dissolved oxygen level will be increased and the fermentation temperature will be finely adjusted; the compensation amount can be dynamically and adaptively corrected according to the transfer prediction error.

Images