A fermentation process control method of ecological compound microbial bacterial fertilizer

Abstract

The application discloses a fermentation process control method of ecological compound microbial bacterial fertilizer, and relates to the technical field of biological agricultural organic fertilizer processing.The method is characterized in that a distributed electronic nose and a multi-dimensional sensor matrix arranged in multiple control partitions of a fermentation pile are used to collect temperature, moisture content, oxygen content and characteristic metabolic gas concentration signals, the signals are input into a digital twin mapping model based on a long short-term memory network, and estimated values of the number of viable bacteria, antagonistic strength and sporulation rate of core functional bacteria are obtained, the estimated values and physical sensing data are combined and input into a reinforcement learning strategy network, control actions such as variable frequency oxygen supplement, pile turning and heat dissipation and pulse type atomization liquid supplement are output, and the control actions are converted by a programmable logic controller to drive field execution mechanisms, so that a sensing, modeling, decision-making and execution closed loop is formed.

Description

Technical Field

[0001] This invention relates to the field of bio-agricultural organic fertilizer processing technology, specifically to a method for controlling the fermentation process of an ecological compound microbial fertilizer. Background Technology

[0002] Ecological compound microbial fertilizer is a new type of fertilizer product that combines multiple functional microbial strains with an organic substrate and achieves organic matter decomposition and functional microbial propagation through aerobic solid-state fermentation. It has functions such as improving soil structure, enhancing crop disease resistance, and promoting nutrient conversion. It belongs to the field of bio-agricultural organic fertilizer processing. Aerobic solid-state fermentation is usually carried out in windrow or trough fermentation systems, providing oxygen to the microorganisms and dissipating fermentation heat through methods such as turning and forced ventilation. It is a typical multi-strain compound fermentation system, involving the intersection of multiple technical fields such as bio-agricultural organic fertilizer processing, agricultural microbiology technology, and agricultural environmental protection industrial control systems.

[0003] In industrial production, to achieve automated management of the fermentation process, the industry has developed online monitoring and control systems based on physical parameters such as temperature, humidity, and oxygen concentration. These systems are typically equipped with distributed temperature and humidity probes, dissolved oxygen or oxygen content sensors, and are linked with actuators such as variable frequency Roots blowers, automatic turning machines, and spray systems. Through the industrial control system, operations such as ventilation, turning, and water replenishment are triggered at timed intervals or based on temperature thresholds, constituting a typical Internet of Things (IoT) industrial control system.

[0004] The closest existing technical solution to this invention in the industry is an automatic control method for aerobic solid-state fermentation based on a temperature-time static curve or temperature threshold triggering. The implementation steps of this method are usually as follows: several temperature probes and humidity probes are arranged in the fermentation pile to collect the temperature and moisture content of the core area of ​​the pile in real time.

[0005] The collected temperature signals are compared with the preset fermentation stage temperature curve;

[0006] When the detected temperature exceeds the preset upper limit threshold, the control system triggers the turning machine to turn the pile for heat dissipation, or starts the variable frequency fan to increase the ventilation volume for forced cooling.

[0007] When the moisture content of the pile is detected to be lower than the preset lower limit, the spray system is triggered to replenish water quantitatively.

[0008] The process is repeated cyclically according to a pre-set fixed fermentation cycle until the preset total number of fermentation days is reached, at which point fermentation is considered complete.

[0009] The control logic of the existing solution is essentially an open-loop or simple feedback control that drives the actuator action based on a single physical quantity threshold judgment. It does not involve online sensing and modeling of the microbial metabolic state and community succession. Its control objective is only to maintain physical parameters such as temperature and humidity within a safe range, without distinguishing or taking into account the survival status of different functional microbial species. Since the control system can only sense macroscopic physical parameters, it cannot obtain online information on the synergistic and antagonistic metabolic state of multiple microbial species within the fermentation system and the sporulation status of core functional bacteria. This leads to the long-term reliance on static empirical timetables for fermentation control, making it difficult to ensure the maturity of the material while simultaneously considering the survival rate and sporulation rate of the core functional bacteria. Consequently, problems such as inflated viable bacteria counts in microbial fertilizer products, poor storage stability, and unstable field growth-promoting effects are urgently needed. There is an urgent need for a fermentation process control method that can sense the metabolic succession state of the microbial community online and achieve synergistic regulation of both maturity and viable bacteria. Summary of the Invention

[0010] The existing aerobic solid-state fermentation control methods have the following technical problems: First, because the existing technology only relies on macroscopic physical parameters such as temperature and humidity for control, it fails to detect the characteristic gas signals produced by the metabolism of microorganisms inside the pile. Therefore, it cannot identify the antagonistic and competitive imbalance of multiple microorganisms caused by changes in nutrient substrate and local pH fluctuations in the middle and late stages of fermentation. This leads to the death of a large number of core functional bacteria when the total number of viable bacteria appears to be high, resulting in the rapid decline of viable bacteria and unstable field growth promotion effect of microbial fertilizer products during the later storage process.

[0011] Secondly, due to the lack of mathematical models that link gas metabolism signals with the state of microbial communities in existing technologies, even if gas sensing data is acquired, it cannot be converted into biological quantitative indicators such as the number of functional bacteria and the strength of antagonism that can be used for control decisions.

[0012] Third, because existing technologies use fixed timetables or single threshold triggering control, their control objectives are singular. They do not simultaneously optimize the degree of decomposition, the sporulation rate of functional bacteria, and the survival rate. This leads to an excessive pursuit of high decomposition, which can easily cause nutrient depletion and premature death of functional bacteria in the later stages. On the other hand, ending fermentation prematurely to preserve the number of surviving bacteria can easily lead to secondary fermentation and seedling burn after fertilizer application. The two are contradictory and existing technologies cannot achieve a dynamic balance online.

[0013] The technical problem to be solved by this invention is to provide an ecological compound microbial fertilizer fermentation process control method that can sense the metabolic succession state of the microbial community online and achieve dual-objective closed-loop adaptive regulation of material maturity and core functional bacteria spore formation rate and survival rate.

[0014] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a fermentation process control method for ecological compound microbial fertilizer, which uses a distributed electronic nose and a multi-dimensional sensor matrix deployed in multiple control zones of the fermentation pile to collect temperature, moisture content, oxygen content and characteristic metabolic gas concentration signals of each zone. The characteristic metabolic gas includes at least two of short-chain fatty acids, nitrogen-containing compounds and characteristic alcohols.

[0015] The collected signals are input into a pre-trained metabolite-microbial community state digital twin mapping model based on a long short-term memory network to obtain the estimated number of viable core functional bacteria, the estimated antagonistic strength, and the estimated sporulation rate for each partition.

[0016] The above biological state estimates are combined with physical sensor data to form the current system state, which is then input into a pre-trained reinforcement learning strategy network to obtain the control actions corresponding to each partition. The control actions include the partition variable frequency oxygenation target air volume, the partition turning and heat dissipation trigger signal, and the dosage and triggering timing of the partition pulse atomization liquid replenishment. The reinforcement learning strategy network is trained with a reward function consisting of a maturity progression score, sporulation rate evaluation, antagonistic strength evaluation, and the energy consumption cost of the corresponding control action.

[0017] The above control actions are converted into control instructions that can be executed by the programmable logic controller and sent to the variable frequency Roots blower, automatic turning and turning machine and pulse atomizing spray system for execution.

[0018] Repeat the above process until the maturity and sporulation rate of each zone reach the preset endpoint criteria, at which point fermentation is considered complete, thus forming a complete closed-loop control loop consisting of four layers: perception, modeling, decision-making, and execution.

[0019] Based on the above, the present invention can also divide the fermentation pile into multiple independent control zones according to the spatial grid, and independently deploy sensing nodes in each control zone and configure execution actions respectively to achieve zoned precise variable frequency oxygenation and pile turning heat dissipation.

[0020] It is also possible to set criteria for increasing antagonistic intensity and excessive progress of decomposition. When the corresponding criteria are met, reverse regulation is triggered, and the strategy network outputs targeted instructions for oxygen supplementation, turning over, or liquid supplementation.

[0021] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0022] 1. This invention, by embedding a distributed electronic nose and a multi-dimensional sensor matrix into the fermentation pile space, captures characteristic gas signals that can indicate the synergistic and antagonistic metabolic states of the microbial community in real time and uses them as online biomarkers. This breaks through the limitations of traditional solid-state fermentation, which only monitors physical parameters such as temperature, humidity, and oxygen, and achieves the technical effect of being able to sense the dynamics of microbial community metabolic succession online.

[0023] 2. This invention constructs a metabolite-microbial community state digital twin mapping model based on a long short-term memory network, realizing a dynamic mapping from environmental physical parameters and characteristic metabolic gases to functional bacterial viability, antagonistic strength, and sporulation rate, thereby achieving the technical effect of real-time quantitative inference of the biological state inside the pile.

[0024] 3. This invention constructs a reinforcement learning closed-loop adaptive control algorithm by taking the material maturity, sporulation rate, and survival rate of core functional bacteria as optimization targets. When the antagonistic trend is predicted to intensify or become over-matured, the algorithm automatically outputs control commands such as zoned variable frequency oxygenation, turning and heat dissipation, or pulsed atomization liquid replenishment. This achieves end-to-end coordinated control of the fermentation endpoint maturity and the storage stability of functional bacteria. Compared with existing technologies that rely on fixed time schedules or single physical thresholds, this invention can achieve coordinated control of the fermentation endpoint maturity and the sporulation rate and survival rate of functional bacteria, while reducing ineffective ventilation and liquid replenishment energy consumption. It also has the added benefits of reducing ineffective ventilation and liquid replenishment energy consumption and improving the quality of soil improvement-type green agricultural inputs. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall system architecture in an embodiment of the present invention;

[0026] Figure 2 This is a flowchart illustrating the training and inference process of the digital twin model in an embodiment of the present invention.

[0027] Figure 3 This is a flowchart of the reinforcement learning closed-loop fermentation control in an embodiment of the present invention. Detailed Implementation

[0028] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0029] The technical solution of the present invention will be further described below with reference to specific embodiments. Those skilled in the art should understand that the following embodiments are merely illustrative and are not intended to limit the scope of protection of the present invention.

[0030] I. System Composition

[0031] The fermentation process control method described in this embodiment is based on a closed-loop system consisting of a distributed sensing layer, a digital twin modeling layer, a dual-objective reinforcement learning decision-making layer, and an execution control layer. The windrow fermentation pile is divided into multiple longitudinal partitions along its length and multiple levels along its depth, forming multiple control partitions. Each control partition deploys a set of sensing nodes, which include temperature probes, humidity probes, oxygen content probes, and an electronic nose gas sensor array composed of gas sensor units with various sensitive membrane materials. The gas sensor units are sensitive to short-chain fatty acids, nitrogen-containing compounds, and characteristic alcohols, respectively. The types of gas sensor units include metal oxide semiconductor gas sensors and electrochemical gas sensors. The data collected by the sensing nodes in each partition are aggregated to the edge data acquisition gateway via fieldbus or wireless means. The edge data acquisition gateway uploads the data to the computing unit running the digital twin model and reinforcement learning decision-making module, and receives the control commands issued by the computing unit and converts them into specific action commands that can be executed by the programmable logic controller (PLC). The PLC is connected to the variable frequency Roots blower, automatic turning machine, and pulse atomizing spray system through control lines, forming a complete closed-loop loop of sensing, modeling, decision-making, and execution.

[0032] II. Construction and Training of Metabolite-Microbial Community State Digital Twin Mapping Model

[0033] The input to the digital twin mapping model described in this embodiment is the first... Each control partition in The multidimensional feature vector composed of environmental physical parameters and characteristic metabolic gas concentrations at any given time is represented as:

[0034] ;

[0035] In the formula, Indicates the first Each partition Temperature at any moment;

[0036] Indicates the first Each partition Moisture content at any given time;

[0037] Indicates the first Each partition Oxygen content at any given time;

[0038] to They represent the first Each partition Detected by the electronic nose array at all times The concentration signal of a characteristic metabolic gas, the The characteristic metabolic gases include short-chain fatty acids, nitrogen-containing compounds, and characteristic alcohols.

[0039] The model's output is this partition. The biological state estimation vector corresponding to time step 1 is represented as:

[0040] ;

[0041] In the formula, This represents the number of viable core functional bacteria estimated by the model;

[0042] This represents the antagonistic strength estimated by the model;

[0043] This represents the core functional sporulation rate estimated by the model.

[0044] The digital twin mapping model described in this embodiment is constructed using a Long Short-Term Memory (LSTM) network. LSTM is a recurrent neural network with memory units. Each memory unit contains a forget gate, an input gate, and an output gate. By performing a gated weighted operation on the hidden state of the previous time step and the input feature vector of the current time step, the updated cell state and hidden state of the current time step are obtained. The hidden states of each time step are passed sequentially along the time dimension. The hidden state of the last time step is mapped through a fully connected layer and outputs the three predicted quantities of the aforementioned core functional bacteria: viable bacterial count, antagonistic strength, and sporulation rate. The specific gating operation method of the LSTM network is a well-known technique familiar to those skilled in the art and will not be elaborated here.

[0045] The model training uses the mean squared error loss function for supervised learning, expressed as:

[0046] ;

[0047] In the formula, This is the complete set of trainable parameters for the Long Short-Term Memory network.

[0048] The model predicts the output;

[0049] The offline detection true values ​​at the corresponding time point are composed of functional bacterial viability, antagonistic strength, and sporulation rate tags obtained by laboratory methods such as metagenomic sequencing and 16S rRNA absolute quantification.

[0050] This represents the total number of time steps for the training samples.

[0051] This is the regularization coefficient, used to suppress model overfitting.

[0052] The offline training of the model was carried out in the following manner: During the pilot and intermediate fermentation stages, environmental physical parameters and electronic nose characteristic gas concentration data of each zone were collected synchronously at fixed time intervals, and metagenomic sequencing and 16S rRNA absolute quantification were performed at the corresponding time points to obtain offline true value labels of functional bacterial viability, antagonistic strength and sporulation rate, and to construct a labeled time-series training sample set.

[0053] The collected raw data is normalized, the time series data is sliced ​​using a sliding time window, and missing data points are interpolated to complete the data, forming standardized model input samples.

[0054] Construct a long short-term memory network structure, set the number of input layer nodes to be consistent with the feature dimension, set the number of hidden layer units and layers, and set the number of output layer nodes to be consistent with the number of biological indicators to be predicted;

[0055] The model parameters are iteratively trained by using a backpropagation algorithm in time combined with an adaptive moment estimation optimizer to narrow down the value of the aforementioned loss function.

[0056] Cross-validation is used to optimize hyperparameters such as the number of hidden layer units, learning rate, and time window length. The model with smaller validation set error is selected as the deployment model, and the model parameters are fixed and deployed to the computing unit for online inference.

[0057] III. Dual-objective closed-loop adaptive regulation of decomposition degree and sporulation rate

[0058] Based on the real-time output of biological state estimates for each zone by the digital twin model, this implementation further constructs a dual-objective closed-loop control module based on reinforcement learning, which takes the material maturity and the core functional bacterial spore formation rate and survival rate as optimization objectives to dynamically generate zone control commands.

[0059] The core functional values ​​output by the digital twin model—estimated viable bacterial count, estimated antagonistic strength, and estimated sporulation rate—together with physical sensor data constitute the current system state, which is then input into the reinforcement learning policy network.

[0060] The strategy network outputs control actions, including the target air volume for variable frequency oxygenation in each zone, the trigger signal for heat dissipation during turning, and the dosage and triggering timing of pulsed atomization liquid replenishment;

[0061] The control actions are converted and executed by the programmable logic controller and applied to the fermentation pile. The state of the pile changes and is fed back to the sensor data acquisition at the next moment. At the same time, the reward is calculated according to the reward function and fed back to the policy network for parameter update, forming a complete closed loop.

[0062] The reward function is designed as a comprehensive evaluation function that takes into account the progress of composting, the survival of the inoculum, and the control of energy consumption, and is expressed as:

[0063] ;

[0064] In the formula, express A quantitative score for the degree of decay progress at any given moment;

[0065] This represents the overall evaluation of the sporulation rate estimated by the digital twin model across the entire field.

[0066] This is the corresponding antagonistic intensity evaluation; the higher the value, the more severe the antagonism, which serves as a penalty.

[0067] This indicates the energy consumption cost corresponding to the current control action, including the energy consumption corresponding to the fan power and the amount of spray.

[0068] , , , These are the preset positive weighting coefficients for the corresponding items, which are determined by process experts based on experience and pilot test data. After calibration, these values ​​are kept constant during the deployment phase.

[0069] The maturity advancement score is calculated based on the degree of deviation between the predicted maturity and the target maturity, and is expressed as follows:

[0070] ;

[0071] In the formula, for The current decomposition maturity is estimated by comprehensively considering physicochemical indicators, including the rate of temperature change, moisture content, and oxygen consumption rate.

[0072] This is the preset target final value for the degree of decay.

[0073] The parameter iteration update of the policy network adopts a value iteration method based on temporal difference learning, which updates the value function estimate corresponding to the current state and the current action by approximating the sum of the immediate reward and the maximum value estimate of the next state.

[0074] For continuous action variables such as variable frequency oxygen supplementation and nebulized fluid supplementation in this invention, in specific implementation, a policy gradient algorithm with continuous action space can be used to replace the aforementioned discretized value iteration method. Its training objective is also to increase the expected cumulative reward defined by the aforementioned reward function.

[0075] IV. Specific Implementation Process of Control Methods

[0076] The fermentation process control method described in this embodiment is implemented cyclically in the following manner: the edge computing unit obtains the latest temperature, moisture content, oxygen content and characteristic metabolic gas concentration data from the sensing nodes of each partition according to the preset sampling period, inputs them into the deployed digital twin mapping model, and obtains real-time estimates of the number of functional bacteria, antagonistic strength and sporulation rate of each partition.

[0077] The above biological state estimates are combined with physical sensor data to form the current system state, which is then input into the reinforcement learning policy network.

[0078] The strategy network outputs control actions for each zone based on the current state, including the target air volume for zone variable frequency oxygenation, the zone turning and heat dissipation trigger signal, and the dosage and triggering timing of pulse atomization liquid replenishment.

[0079] Calculate the immediate reward corresponding to the current decision based on the aforementioned reward function, and evaluate the comprehensive impact of the current control action on the progress of composting, the survival of the microbial strain, and energy consumption costs.

[0080] Determine whether the current state has reached the preset antagonism aggravation threshold or the over-decomposition trend criterion. When it is determined that the estimated value of antagonism intensity continues to rise and exceeds the preset antagonism intensity threshold, reverse regulation is triggered, and the strategy network outputs at least one of the following commands for the corresponding partition: oxygen supplementation, turning over, or liquid supplementation.

[0081] When the rate of progress of decomposition exceeds the preset rate of progress and the estimated value of sporulation rate is lower than the preset target value of sporulation rate, the reverse regulation is also triggered.

[0082] The control actions output by the strategy network are converted into specific control instructions that can be executed by the programmable logic controller, including the frequency value of the fan inverter, the start and stop signal of the turner and the opening duration of the spray valve, and sent to the field actuators for execution.

[0083] After the actuator completes its actions, the system enters the next sampling cycle and re-executes the above data collection, model reasoning, reward evaluation, and instruction issuance process, forming a continuous closed loop of perception, modeling, decision-making, execution, and feedback, until the maturity and sporulation rate of each zone simultaneously reach the preset endpoint criteria. At this point, the system determines that fermentation has ended and outputs a termination signal.

[0084] As a specific embodiment, in a certain bar stack ecological compound microbial fertilizer fermentation workshop, the length direction is divided into multiple longitudinal partitions, and a set of sensing nodes is deployed in each partition. In the early stage of fermentation, the temperature of each partition gradually rises, the concentration signal of short-chain fatty acids detected by the electronic nose array is at a low level, the antagonistic strength estimate output by the digital twin model is in a low range, and the strategy network outputs a conventional frequency conversion oxygen supplementation command to maintain the aerobic state of the stack.

[0085] As fermentation progresses into the middle and late stages, some zones experience uneven consumption of local nutrient substrates. The electronic nose array detects abnormal fluctuations in the concentration signal of nitrogen-containing compounds. Based on this, the digital twin model outputs an estimated value of antagonistic intensity that continues to rise and exceeds a preset threshold, triggering reverse regulation. The strategy network outputs an increased variable frequency oxygenation airflow for this zone and simultaneously triggers a pulsed atomization liquid replenishment command. After conversion by the programmable logic controller, the command drives the corresponding actuator to suppress further aggravation of antagonism while replenishing nutrients and maintaining the normal progress of the core function of bacterial spore formation.

[0086] When the estimated maturity and sporulation rate of each zone simultaneously reach the preset endpoint criterion, the system determines that fermentation has ended.

[0087] V. Alternative Implementation Methods

[0088] In the construction of the digital twin mapping model, in addition to using the long short-term memory network structure, gated recurrent unit network can also be used, or a time-series prediction network based on self-attention mechanism can be used to achieve the same time-series mapping relationship modeling from environmental physical parameters and characteristic metabolic gases to functional bacterial viability, antagonistic strength, and sporulation rate. Although the internal structure of the network is different, the input and output definitions and training methods are consistent with the scheme described in this embodiment, which are equivalent alternative technical means.

[0089] In the dual-objective closed-loop control decision module, in addition to using reinforcement learning algorithms, a multi-objective decision method based on fuzzy logic rule base can also be used in application scenarios with limited computing resources or labeled data. That is, corresponding partitioned oxygenation, turning, and liquid replenishment control rule tables are set in advance for different antagonistic intensity intervals and decomposition progress intervals. By integrating multiple biological state estimates output by the digital twin model through fuzzy inference, equivalent partitioned control actions are output to replace the reinforcement learning policy network to realize the dual-objective closed-loop control function. This can also achieve the technical effect of the present invention that takes into account both decomposition degree and functional fungal sporulation rate.

[0090] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for controlling the fermentation process of an ecological compound microbial fertilizer, characterized in that, include: By deploying a distributed electronic nose and a multi-dimensional sensor matrix in multiple control zones of the fermentation pile, the temperature, moisture content, oxygen content, and characteristic metabolic gas concentration signals of each control zone are collected. The characteristic metabolic gases include at least two of the following: short-chain fatty acids, nitrogen-containing compounds, and characteristic alcohols. The collected temperature, moisture content, oxygen content, and characteristic metabolic gas concentration signals are input into a pre-trained digital twin mapping model based on a long short-term memory network to obtain the estimated number of viable core functional bacteria, the estimated antagonistic strength, and the estimated sporulation rate for each control zone. The estimated values ​​of the viable bacterial count, antagonistic strength, and sporulation rate of the core functional bacteria in each control zone are combined with the temperature, moisture content, and oxygen content signals to form the current system state. This state is then input into a pre-trained reinforcement learning strategy network to obtain the control actions corresponding to each control zone. The control actions include the target air volume for variable frequency oxygenation in the zone, the trigger signal for heat dissipation by turning the pile in the zone, and the dosage and triggering timing of pulsed atomized liquid replenishment in the zone. The reinforcement learning strategy network is trained using a reward function consisting of a maturity progression score, a sporulation rate evaluation, an antagonistic strength evaluation, and the energy cost of executing the corresponding control action. The control actions are converted into control instructions that can be executed by the programmable logic controller (PLC) and sent to the variable frequency Roots blower, automatic turning machine and pulse atomizing spray system connected to the PLC for execution, so that the corresponding control zone completes the corresponding actions in variable frequency oxygenation, turning heat dissipation and pulse atomizing liquid replenishment. Repeat the above steps until the maturity and sporulation rate of each control zone reach the preset endpoint criteria, at which point fermentation is considered complete.

2. The method according to claim 1, characterized in that, The input to the digital twin mapping model is the first... Each control partition in The feature vector at time step, the feature vector including the control partition Temperature, moisture content, oxygen content, and other parameters at any given time Concentration signals of characteristic metabolic gases; The output of the digital twin mapping model is the control partition. The biological state estimation vector is composed of the estimated number of viable core functional bacteria, the estimated antagonistic strength, and the estimated sporulation rate at the corresponding time.

3. The method according to claim 1, characterized in that, The digital twin mapping model was trained in the following way: during the pilot and intermediate fermentation stages, the temperature, moisture content, oxygen content and characteristic metabolic gas concentration signals of each control zone were collected simultaneously, and at the corresponding time points, the label values ​​of the viable number of core functional bacteria, antagonistic strength and sporulation rate were obtained by metagenomic sequencing and 16S rRNA absolute quantitative detection, and a labeled time-series training sample set was constructed. The time-series training sample set is subjected to normalization, time window slicing, and missing value interpolation. With the goal of reducing the error between the predicted and labeled values, the parameters of the digital twin mapping model are iteratively trained using a backpropagation algorithm in time combined with an adaptive moment estimation optimizer.

4. The method according to claim 1, characterized in that, The degree of decomposition progression score is determined based on the deviation between the estimated decomposition value calculated from the comprehensive physicochemical indicators at time t and the preset target final decomposition value. The physicochemical indicators include the rate of temperature change, moisture content, and oxygen consumption rate.

5. The method according to claim 1, characterized in that, Also includes: When the estimated antagonistic intensity of any control zone shows an upward trend within a consecutive preset number of sampling periods and the value after the rise exceeds the preset antagonistic intensity threshold, reverse regulation is triggered, and the reinforcement learning strategy network outputs at least one of the oxygen supplementation control command, the turning control command, and the liquid supplementation control command for the corresponding control zone. When the rate of decay progression of any control zone exceeds the preset rate of progression threshold and the estimated sporulation rate of that control zone is lower than the preset sporulation rate target value, the reverse regulation is also triggered.

6. The method according to claim 1, characterized in that, The fermentation pile is divided into multiple control zones along its length and depth. Each control zone independently deploys a sensing node consisting of a temperature probe, a humidity probe, an oxygen content probe, and an electronic nose gas sensor array. The signals collected by each sensing node are aggregated through an edge data acquisition gateway and then uploaded to a computing unit running the digital twin mapping model and the reinforcement learning policy network.

7. The method according to claim 6, characterized in that, The electronic nose gas sensor array includes a gas sensor unit sensitive to short-chain fatty acids, a gas sensor unit sensitive to nitrogen-containing compounds, and a gas sensor unit sensitive to characteristic alcohols. The gas sensor unit includes a metal oxide semiconductor gas sensor and an electrochemical gas sensor.

8. The method according to claim 1, characterized in that, The digital twin mapping model based on long short-term memory network includes an input layer, a hidden layer composed of several layers of long short-term memory network units, and an output layer connected in sequence. The hidden state of the last time step of the hidden layer is mapped by the fully connected layer and outputs the estimated values ​​of the viable number of core functional bacteria, the estimated value of the antagonistic strength, and the estimated value of the sporulation rate.

9. A fermentation process control system for an ecological compound microbial fertilizer, characterized in that, include: A distributed electronic nose and a multi-dimensional sensor matrix are deployed in multiple control zones of the fermentation pile to collect temperature, moisture content, oxygen content, and characteristic metabolic gas concentration signals of each control zone. An edge data acquisition gateway, connected to the distributed electronic nose and multi-dimensional sensor matrix, is used to aggregate the acquisition signals from each control zone; The computing unit, connected to the edge data acquisition gateway, runs a digital twin mapping model and a reinforcement learning policy network built on a long short-term memory network. It is used to obtain the estimated values ​​of the viable number of core functional bacteria, the estimated value of antagonism intensity, and the estimated value of sporulation rate of each control zone based on the converged acquisition signals, and to generate control actions for each control zone based on the estimated values ​​of the viable number of core functional bacteria, the estimated value of antagonism intensity, the estimated value of sporulation rate, and the temperature, moisture content, and oxygen content signals. A programmable logic controller, connected to the computing unit, is used to convert the control actions into control instructions; The variable frequency Roots blower, automatic turning machine, and pulse atomizing spray system are respectively connected to the programmable logic controller and are used to perform variable frequency oxygenation, turning and heat dissipation, and pulse atomizing liquid replenishment according to the control instructions.

10. The system according to claim 9, characterized in that, The sensing nodes corresponding to each control zone in the distributed electronic nose and multi-dimensional sensor matrix include a temperature probe, a humidity probe, an oxygen content probe, and an electronic nose gas sensor array. The electronic nose gas sensor array includes a metal oxide semiconductor gas sensor and an electrochemical gas sensor.

Images