Livestock and poultry manure fermentation monitoring method based on digital twinning

Abstract

The invention discloses a livestock and poultry manure fermentation monitoring method based on digital twinning, and relates to the technical field of livestock and poultry manure treatment. By constructing a high-precision digital twinborn model, fermentation field sensing data are synchronized in real time, the fermentation state is simulated, and the development trend is predicted; setting an intelligent control strategy fusing data relevance and control priority, and dynamically identifying and avoiding rule conflicts; on this basis, control actions such as pile turning, spraying and ventilation are automatically executed, feedback data are recorded, and twin model parameters are continuously optimized based on feedback results; according to the method, intelligent regulation and control and model self-adaptive updating of the fermentation process are realized, and the stability, the energy utilization efficiency and the prediction accuracy of the system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of livestock and poultry manure treatment, and in particular to a livestock and poultry manure fermentation monitoring method based on digital twins. Background Art

[0002] Livestock and poultry manure fermentation monitoring refers to the process of real-time or periodic monitoring and regulation of various parameters of livestock and poultry manure during the fermentation process, aiming to ensure that the fermentation process is efficient, stable and meets environmental protection requirements. The system is equipped with intelligent control based on the rule engine, such as "if the temperature is >65℃ and the humidity is <30%, turn on the composting + spraying", and there is another rule "if the temperature is <55℃ after turning the compost, stop spraying".

[0003] The existing technology has the following shortcomings: Under extreme conditions such as continuous high-temperature drying, the two rules of the system may conflict in a loop, constantly triggering mutually exclusive conditions, thus entering a logical dead loop, causing the execution layer to frequently start and shut down the compost turning / spraying equipment. In addition, if a logical conflict occurs in the rule engine, it will not only cause frequent start and stop of the equipment and a sharp increase in energy consumption, but may also seriously interfere with the digital twin model's accurate mapping and prediction of the fermentation status. Summary of the invention

[0004] The purpose of the present invention is to provide a livestock and poultry manure fermentation monitoring method based on digital twins to address the shortcomings of the background technology.

[0005] In order to achieve the above-mentioned purpose, the present invention provides the following technical solution: a method for monitoring livestock and poultry manure fermentation based on digital twin, comprising: Build a digital twin model of the livestock and poultry manure fermentation process, collect real-time data of the fermentation pile through sensors deployed at the fermentation site, and synchronize the real-time data to the digital twin model; Simulate the actual fermentation status in the digital twin model to predict fermentation development trends and potential anomalies; Setting an intelligent control strategy, wherein the control strategy combines the mutual influence degree between real-time data and the execution conditions of the control action priority to determine whether there is a logical conflict in the rule engine, and if so, optimizes the control strategy; The control strategy constructs a comprehensive feature vector based on the coupling relationship between the predicted value of oxygen utilization efficiency and the estimated value of surface drying rate, and combines the preset control action priority to predict the risk value of logical conflict in the rule engine through a machine learning model, and compares it with a predetermined threshold to determine whether there is a logical conflict; According to the optimized control strategy, the pile turning, spraying and ventilation operations are automatically performed, and the control behavior and feedback data are recorded in real time. The digital twin model is continuously adjusted based on the recorded feedback data.

[0006] Preferably, the real-time data collected by the sensor include temperature, humidity, oxygen concentration, pH value and gas release amount at different layers of the pile, which are used to construct a multi-dimensional, hierarchical twin model input data set.

[0007] Preferably, a graph neural network is used to predict fermentation development trends and potential anomalies, specifically including: The entire fermentation pile is divided into multiple monitoring sub-areas, each of which corresponds to a node in the graph; the connection relationship between nodes is established based on the spatial adjacency relationship of the pile and the heat or moisture migration path; each node is associated with multiple sensor data to form a feature vector of the node; the initial graph structure G=(V,E,X) of the pile is constructed, where: V is the node set; E is the edge set; X is the node feature matrix; The state of the heap at each moment is represented as a graph Gt, and the continuous time forms a graph sequence; combining the node feature changes in each graph with the graph structure invariance, the time series graph data input is constructed; the label data is the target variable or abnormal state mark of each node in a certain time window in the future; Use graph convolutional networks or graph attention networks to extract the spatiotemporal features of nodes; the GNN model performs convolution operations on the graph Gt at each time step to obtain the embedded representation Ht of the node; The node representation Ht is input into the subsequent time series prediction module; the temperature or humidity variables of each node in the future time step are predicted; the prediction results are scored for anomalies based on the actual historical distribution; if the predicted value of a node deviates from the normal change trend, it is marked as a potential anomaly.

[0008] Preferably, the method for obtaining the predicted value of oxygen utilization efficiency is as follows: constructing a causal relationship diagram of oxygen utilization, determining variable nodes related to oxygen utilization efficiency, including fermentation pile temperature, humidity or moisture content, oxygen concentration, and ventilation volume, establishing a causal dependency relationship between variables, forming a directed acyclic graph, and setting the predicted value of oxygen utilization efficiency as a target variable node therein; Collect real-time monitoring data from historical fermentation processes for training Bayesian networks; During the fermentation process, the observed values ​​of temperature, humidity, oxygen concentration and ventilation volume are collected in real time; the observed values ​​are input into the Bayesian network as known nodes; the conditional propagation algorithm in the Bayesian network is triggered using the input of the observed nodes; the network automatically updates the posterior probabilities of all relevant nodes; the probability distribution result of the oxygen utilization efficiency prediction value node is obtained as the current oxygen utilization efficiency prediction value.

[0009] Preferably, the method for obtaining the surface drying rate estimate is: High-sensitivity humidity sensors are deployed on the surface of the fermentation pile to collect data at fixed time intervals; each data sequence represents the change trajectory of surface humidity within a fixed time window, forming a humidity time series sample; an autoencoder is used to perform unsupervised feature learning on the humidity time series: the encoder part automatically extracts low-dimensional feature vectors representing the drying change trend in each time series; the decoder part attempts to restore the feature vector to the original humidity curve; after training, the encoder compresses any newly collected humidity change sequence into a fixed-dimensional feature representation; A clustering algorithm is used to cluster the feature vectors extracted by the autoencoder; each clustering result represents a typical drying mode; after clustering, a corresponding drying rate label is assigned to each type of drying mode; during the fermentation process, the system continuously collects the current surface humidity time series; the trained autoencoder is used to perform feature compression on the current sequence; the feature vector is matched with the cluster center to determine which drying mode the current sequence belongs to; based on the category, the surface drying rate estimate for the current time period is output accordingly.

[0010] Preferably, the predicted value of oxygen utilization efficiency and the estimated surface drying rate are converted into a comprehensive feature vector, and the comprehensive feature vector is used as the input of the machine learning model. Combined with the currently set control action priority, the machine learning model uses each group of comprehensive feature vectors to predict the risk value label of logical conflict in the rule engine as the prediction target, and takes minimizing the sum of prediction errors of risk value labels of logical conflict in all rule engines as the training target. The machine learning model is trained until the sum of prediction errors converges and the model training is stopped. The risk value of logical conflict in the rule engine is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0011] Preferably, the acquired risk value of the logical conflict in the rule engine is compared with a predetermined threshold value. If the risk value of the logical conflict in the rule engine is greater than or equal to the predetermined threshold value, it means that the logical conflict exists in the rule engine. At this time, a warning signal is generated and the control strategy is optimized; if the risk value of the logical conflict in the rule engine is less than the predetermined threshold value, it means that the logical conflict does not exist in the rule engine. At this time, no warning signal is generated and no additional adjustment is required.

[0012] Preferably, the digital twin model is continuously adjusted based on the recorded feedback data, specifically including: in the initial stage, a digital twin model is constructed to simulate the fermentation process of livestock and poultry manure, and a set of data structures for recording are established to associate the correspondence between each control behavior, environmental response and model output; multiple small sample task sets are constructed using historical feedback data, each task represents a specific fermentation sub-scenario, and each task contains a training set and a verification set; the digital twin model is gradually fine-tuned in each task; the initial parameters of the model are optimized through repeated iterations of multiple tasks. In actual operation, after each control behavior is executed, the collected feedback data is immediately constructed as a new task; the new task is input into the trained meta-learning framework for iteration, the current twin model is locally fine-tuned, and the fine-tuned model is retained as the current scene sub-model; the system regularly verifies the prediction performance of the current twin sub-model; if the error decreases or meets the self-set threshold, the system will extract the fine-tuned parameter changes, feed them back to the main model parameter set, update the main model or add it to the experience model pool.

[0013] In the above technical solution, the technical effects and advantages provided by the present invention are: 1. The present invention constructs a digital twin model to perform high-fidelity and dynamic simulation of the entire process of livestock and poultry manure fermentation, and combines advanced algorithms such as graph neural networks, Bayesian networks, and autoencoders to achieve intelligent acquisition of key parameters such as fermentation trend prediction, potential anomaly identification, oxygen utilization efficiency evaluation, and surface drying rate judgment. By integrating the above key features into a comprehensive vector and introducing a polynomial regression model to predict the risk of rule engine conflict, the logical mutual exclusion problems existing in traditional rule engines under extreme working conditions can be effectively identified, and early warnings can be generated and control strategies optimized in advance to avoid the risk of the system entering an infinite loop or frequent start and stop of equipment.

[0014] 2. The present invention introduces a meta-learning algorithm to continuously and adaptively optimize the digital twin model, so that the model can be quickly adjusted under variable fermentation materials, climate environments and operating conditions, and improve the adaptability and prediction accuracy of the model in new scenarios. This method realizes a closed-loop intelligent control process, which not only enhances the stability, intelligence and energy efficiency of the fermentation process, but also provides a scalable and highly reliable digital solution for the resource utilization of livestock and poultry breeding waste. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0016] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] For examples, see Figure 1 As shown, the livestock and poultry manure fermentation monitoring method based on digital twins described in this embodiment includes: Build a digital twin model of the livestock and poultry manure fermentation process, collect real-time data of the fermentation pile through sensors deployed at the fermentation site, and synchronize the real-time data to the digital twin model; Simulate the actual fermentation status in the digital twin model to predict fermentation development trends and potential anomalies; Setting an intelligent control strategy, wherein the control strategy combines the mutual influence degree between real-time data and the execution conditions of the control action priority to determine whether there is a logical conflict in the rule engine, and if so, optimizes the control strategy; According to the optimized control strategy, the pile turning, spraying and ventilation operations are automatically performed, and the control behavior and feedback data are recorded in real time. The digital twin model is continuously adjusted based on the recorded feedback data.

[0019] Building a digital twin model of the livestock and poultry manure fermentation process refers to high-fidelity modeling and simulation of the actual fermentation system in a virtual environment, including: A geometric model of the fermentation pile was established, including parameters such as the length, width, height, hierarchical structure, and material stacking density of the pile; the heat conduction path, moisture migration channel, and ventilation hole arrangement inside the pile were simulated; and the effects of different material types (such as pig manure, chicken manure, and cow manure) on the structural stability and heat flow characteristics of the pile were considered.

[0020] Based on the principle of aerobic fermentation, a microbial metabolic kinetic model is introduced to simulate processes such as organic matter decomposition, carbon-nitrogen cycle, and heat release; the reaction rate in different fermentation stages (warming period, high temperature period, cooling period) is simulated as the temperature, pH, and moisture change; an oxygen consumption model and by-product release models such as carbon dioxide and ammonia are introduced to enhance the environmental response capability of the twin model.

[0021] Simulate the impact of ambient temperature, humidity, wind speed, etc. on the fermentation process of the pile; introduce weather data interface or edge meteorological station information to enable the model to have dynamic perception of the external environment; consider the differences in ventilation efficiency and heat exchange efficiency caused by site layout.

[0022] In order to achieve dynamic updates of the digital twin model, it is necessary to deploy various types of sensors at the fermentation site to continuously monitor the internal and environmental parameters of the fermentation pile, mainly including: Temperature sensor: used to monitor the temperature distribution at different depths of the pile (surface, middle and bottom layers); humidity sensor: monitors the moisture content of the material or the relative humidity of the pile to determine the drying trend; oxygen sensor: monitors the ventilation condition of the pile and whether it is in an aerobic state; gas sensor (NH3, CH4, etc.): used to detect odorous substances or anaerobic reaction by-products.

[0023] Multi-point arrangements are carried out according to the depth and horizontal distribution of the pile structure to obtain more comprehensive pile parameters; sensors are deployed in different fermentation areas (starting section, middle section, and final section) to achieve regional comparison and abnormality identification; all sensors are connected to the edge acquisition gateway or data acquisition controller (DAQ) via wireless (such as LoRa, NB-IoT) or wired methods.

[0024] The collected multi-source heterogeneous data needs to be standardized and synchronized to the digital twin platform in real time to drive model updates and predictive analysis: the sensor data must be formatted, denoised, time-aligned, and anomalies eliminated; filters, sliding average algorithms, and data interpolation techniques can be used to enhance data stability; a real-time data caching mechanism must be established to ensure data integrity and traceability.

[0025] Edge computing and cloud platform collaboration are used to achieve two-way synchronization between data reporting and twin models; support updating model input variables at a time granularity of seconds or minutes; and high-frequency synchronization strategies can be triggered for key indicators (such as sudden temperature changes) to achieve dynamic response.

[0026] Every time the data is updated, the model will automatically run simulation calculations to deduce the current internal state of the pile and predict future fermentation trends. If the predicted value deviates from the set target range, it will be used as a decision support input for the control layer. The model is also used to identify abnormal patterns (such as fermentation hypothermia, local anaerobic conditions, etc.) to achieve early warning.

[0027] Graph neural networks can be used to simulate actual fermentation status in digital twin models and predict fermentation development trends and potential anomalies. The advantages of their application are: the pile is regarded as a "spatial graph" composed of multiple points, with nodes representing sensor points and edges representing heat conduction or moisture migration paths; the spatial correlation and mutual influence between different areas of the pile can be captured; and the ability to model spatial heterogeneous features in multi-sensor data is strong.

[0028] The entire fermentation pile is divided into multiple monitoring sub-areas, each of which corresponds to a node in the graph; the connection relationship (edge) between nodes is established based on the spatial adjacency relationship of the pile and the heat or moisture migration path; each node is associated with multiple sensor data (such as temperature, humidity, oxygen concentration) to form a feature vector of the node; the initial graph structure G=(V,E,X) of the pile is constructed, where: V is the node set (fermentation sub-area); E is the edge set (spatial coupling relationship); X is the node feature matrix (multidimensional environmental variable).

[0029] The state of the pile at each moment is represented as a graph Gt, and continuous time forms a graph sequence; combining the node feature changes in each graph with the graph structure invariance, a time series graph data input is constructed; the label data is the target variable (such as temperature, humidity, etc.) or abnormal state mark of each node in a certain time window in the future.

[0030] A graph convolutional network (GCN) or a graph attention network (GAT) is used to extract the spatiotemporal features of nodes. The GNN model performs a convolution operation on the graph Gt at each time step to obtain the embedded representation Ht of the node. The representation of each node is not only based on local sensor data, but also integrates the features of adjacent nodes to reflect the spatial coupling relationship inside the stack.

[0031] The node representation Ht is input into the subsequent time series prediction module (such as LSTM, GRU or Transformer module); the temperature / humidity and other variables are predicted for each node's future time step (such as t+1 to t+6 hours); the prediction results are scored for anomalies (such as residual analysis and probability distribution shift judgment) based on the actual historical distribution; if the predicted value of a node deviates from the normal change trend, it is marked as a potential anomaly.

[0032] The prediction results and abnormal identification results are fed back to the digital twin platform to update the thermal map, trend line, and status label in the twin pile body; trigger control strategies in advance, such as early warning of pile turning and spraying operations; and compare the identification results with subsequent actual monitoring results to dynamically adjust the GNN model parameters (online learning or retraining) to improve prediction accuracy.

[0033] There are multiple control actions in the system, such as compost turning, spraying, ventilation, etc. The priority of each control action is preset (such as spraying takes precedence over compost turning), which is used as the judgment weight in conflict detection.

[0034] Extract real-time sensor data closely related to control actions, including the predicted value of oxygen utilization efficiency inside the pile, which is used to reflect the real-time metabolic ability of microorganisms to oxygen during the fermentation process, which is highly related to whether the pile needs to be turned, and reflects whether the pile ventilation is effective, as well as the estimated drying rate of the pile surface, which indicates the rate of surface water loss per unit time, is closely related to whether spraying is needed, and can reflect the drying rate rather than just the absolute humidity value.

[0035] Traditional systems mostly use oxygen concentration as the basis for judgment, but this value is greatly affected by ventilation and location and is inaccurate. The predicted value of oxygen utilization efficiency can more directly reflect "whether the pile needs to be turned over" and is a more dynamic indicator that reflects the status of microorganisms.

[0036] The method for obtaining the predicted value of oxygen utilization efficiency is as follows: construct a causal relationship diagram of oxygen utilization, determine the variable nodes related to oxygen utilization efficiency, including fermentation pile temperature, humidity or moisture content, oxygen concentration (intake, inside the pile), and ventilation volume; establish the causal dependency relationship between variables based on fermentation principles and field experience to form a directed acyclic graph (DAG); set the predicted value of oxygen utilization efficiency as the target variable node, which is jointly affected by other variables; Collect a large amount of real-time monitoring data from historical fermentation processes for training Bayesian networks; For each pair of dependent variables, a conditional probability distribution is established through statistical methods; if some data are missing or sparse, prior knowledge can be introduced to set the initial probability, and then corrected through Bayesian updating.

[0037] During the fermentation process, the observed values ​​of temperature, humidity, oxygen concentration and ventilation volume are collected in real time; the observed values ​​are input into the Bayesian network as known nodes; the conditional propagation algorithm in the Bayesian network is triggered using the input of the observed nodes; the network automatically updates the posterior probabilities of all relevant nodes; the probability distribution result of the oxygen utilization efficiency prediction value node is obtained as the current oxygen utilization efficiency prediction value.

[0038] Unlike absolute humidity, the estimated surface drying rate reflects the dynamic trend of water loss and can determine whether it is "continuously dry". It is the core variable for determining whether spraying is needed to avoid the misjudgment of "low humidity but not actually dry".

[0039] An estimate of the surface drying rate is obtained as follows: High-sensitivity humidity sensors are deployed on the surface of the fermentation pile to collect data at fixed time intervals (such as every 5 minutes or 10 minutes); each data sequence represents the change trajectory of surface humidity within a certain time window, forming a large number of humidity time series samples; The autoencoder is used to perform unsupervised feature learning on the above humidity time series: the encoder part automatically extracts the low-dimensional feature vector that best represents the dryness change trend of each time series; the decoder part attempts to restore the feature vector to the original humidity curve for training stability; after training, the encoder can compress any newly collected humidity change sequence into a feature representation of fixed dimensionality.

[0040] The feature vectors extracted by the autoencoder are clustered using a clustering algorithm (such as K-Means, DBSCAN, or DTW-based hierarchical clustering); each clustering result represents a typical drying mode, such as: fast drying, slow drying, drying stagnation, and rewetting after drying (such as after rain); after clustering, a corresponding drying rate label (such as "high", "medium", "low" or numerical level) is assigned to each type of drying mode.

[0041] During the fermentation process, the system continuously collects the current surface humidity time series; uses the trained autoencoder to compress the features of the current sequence; matches the feature vector with the cluster center to determine which drying mode the current sequence belongs to; and outputs the estimated surface drying rate for the current time period based on the category.

[0042] The predicted value of oxygen utilization efficiency and the estimated surface drying rate are converted into a comprehensive feature vector, which is used as the input of the machine learning model. The feature vector reflects the linkage trend between the two types of control behaviors (such as high OUE + high SDR indicates that compost turning and spraying may be triggered at the same time). OUE is the predicted value of oxygen utilization efficiency, and SDR is the estimated surface drying rate. Combined with the currently set control action priority (such as spraying takes precedence over compost turning), the machine learning model uses each group of comprehensive feature vectors to predict the risk value label of logical conflict in the rule engine as the prediction target, and minimizes the sum of prediction errors of risk value labels with logical conflicts in all rule engines as the training target. The machine learning model is trained until the sum of prediction errors reaches convergence, and the model training is stopped. The risk value of logical conflict in the rule engine is determined based on the model output results. Among them, the machine learning model is a polynomial regression model.

[0043] The obtained risk value of logical conflict in the rule engine is compared with the preset threshold. If the risk value of logical conflict in the rule engine is greater than or equal to the preset threshold, it means that there is a logical conflict in the rule engine. At this time, a warning signal is generated and the control strategy is optimized; if the risk value of logical conflict in the rule engine is less than the preset threshold, it means that there is no logical conflict in the rule engine. At this time, no warning signal is generated and no additional adjustment is required.

[0044] Dynamically adjust the execution order of control actions (priority adjustment): reallocate the priority of control actions according to the actual working conditions reflected by the current OUE and SDR; for example, if the current OUE is small → increase the priority of compost turning; if the current SDR is large → increase the priority of spraying; avoid executing two conflicting actions (such as compost turning and spraying) when both meet the trigger conditions at the same time; if the weights of two actions are similar, the system can insert a "delay window" to execute the second action (refer to the "shortest cooling time" model).

[0045] A "control mutual exclusion time lock mechanism" is introduced to set the minimum interval time threshold between actions. For example, spraying is not allowed within 20 minutes after turning the pile; turning the pile again is not allowed within 15 minutes after spraying; avoid rapid switching that causes frequent start and stop of equipment or abnormal pile disturbance; the time lock threshold can be dynamically adjusted based on historical feedback.

[0046] After the system detects the risk of conflict, it automatically optimizes the following thresholds: increase the temperature threshold for triggering pile turning (for example, from 65°C to 68°C); lower the lower limit of humidity for triggering spraying (for example, from 30% to 28%); stagger the trigger conditions to reduce the probability of meeting all conditions at the same time; the range of optimization value changes is controlled by the intensity of the "risk value". The higher the risk, the greater the adjustment.

[0047] Rather than triggering two conflicting actions in parallel, a progressive control strategy of “light first, heavy later” is adopted: first, only execute an action with a higher priority or a more significant risk; decide whether to execute the second one by observing the changes in feedback parameters; for example: observe whether the SDR decreases 5 minutes after the initial spraying; if it is still dry, trigger the next stage of turning the pile; in essence, a dynamic feedback closed-loop control logic is introduced.

[0048] The optimized control strategy is generated by the central control module, including: control action type (turning, spraying, ventilation); execution time (immediate execution or delayed execution); action parameters (such as spraying duration, turning depth, air volume intensity); the control strategy is encoded as a standard control instruction and sent to the execution device through the PLC or remote IO module, such as: compost turning robot arm or compost turner; spray solenoid valve control system; ventilation fan unit.

[0049] The control system monitors the action status of the execution layer in real time to ensure that the instructions are responded to accurately; for continuous control actions (such as sprinkler and ventilation), the system starts duration control or closed-loop feedback adjustment according to the set parameters; if the equipment executes abnormally (such as startup failure, action interruption), the system will trigger the abnormal termination mechanism and record the fault log.

[0050] The system records the information of each control action execution in a structured manner, including but not limited to: action type (turning / spraying / ventilation); execution start and end time; action intensity parameters (such as wind speed, water spraying volume, turning frequency); control reasons (such as low OUE triggering turning, and high SDR triggering spraying); decision sources (such as model output / manual intervention); data is stored in the "control behavior log module" for historical backtracking and model learning.

[0051] After the action is executed, the system collects the corresponding environmental feedback data in real time, including: whether the temperature distribution is evenly improved after turning the pile; the trend of surface humidity changes after spraying; whether the oxygen concentration and OUE are improved after ventilation; compare the parameter changes before and after control, and quantify the control effect; write the feedback data and execution data into the "Feedback Data Archiving Module" as the basis for subsequent control strategy optimization and model calibration.

[0052] The system performs closed-loop control analysis based on feedback data: if the feedback result reaches the expected goal, the current strategy is solidified; if the feedback deviation is large, it triggers model update, parameter fine-tuning or recommends manual intervention; at the same time, feedback data is used to train or fine-tune prediction models (such as OUE prediction, drying rate identification, and conflict risk assessment model) to improve the accuracy of the next round of strategies.

[0053] This method builds a rapid adaptation mechanism between control behavior and feedback data, so that the digital twin model has the ability to self-adjust and adapt to different fermentation scenarios, material differences and environmental disturbances, including: In the initial stage, a differentiable and transferable digital twin model is constructed to simulate the fermentation process of livestock and poultry manure (temperature and humidity dynamics, microbial metabolism, etc.); at the same time, a set of data structures for recording is established to associate the correspondence between each control behavior, environmental response and model output; each record contains: control input (such as spraying duration, turning frequency); model prediction value (such as temperature rise rate); actual feedback value (sensor measurement); environmental context variables (such as outside temperature, ventilation mode, etc.).

[0054] Historical feedback data is used to construct multiple small sample task sets, each of which represents a specific fermentation sub-scenario, such as turning over wet pig manure; rapid drying in high temperatures in summer; rapid ventilation in local hypoxia, etc.; each task contains a training set (for model updating) and a validation set (for effect evaluation) to evaluate the generalization and adaptability of the twin model; and simulates the goal of "whether a model can quickly adapt to different fermentation environments".

[0055] The digital twin model is gradually fine-tuned in each task; through repeated iterations of multiple tasks, the initial parameters of the model are optimized so that it can quickly adapt to new tasks; in essence, it is not about training "an optimal model" but training "a model that is easy to fine-tune."

[0056] In actual operation, after each control action is executed, the collected feedback data is immediately constructed as a new task; the task is input into the trained meta-learning framework, and the current twin model can be locally fine-tuned with only a few iterations; the goal of fine-tuning is to make the model's mapping relationship between the current stack and the control response more accurate; retain the fine-tuned model as the current scene sub-model for short-term prediction and control.

[0057] The system verifies the predictive performance of the current twin model periodically or by triggering mechanism (for example, through subsequent real sensor data); if the error drops significantly or meets the self-set threshold, the system will: extract the fine-tuned parameter changes; feedback to the main model parameter set, update the main model or add it to the empirical model pool; realize the gradual evolution of the main model and enhance its generalization capabilities.

[0058] Through continuous feedback, fine-tuning and induction processes, the system gradually builds a digital twin sub-model library suitable for: different fermentation materials; different climatic environments; different pile structures; when formulating future control strategies, the system can quickly call the most matching sub-model from the model pool for high-precision prediction and control strategy deduction.

[0059] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0060] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

Claims

1. A method for monitoring livestock and poultry manure fermentation based on digital twins, characterized in that: include: Build a digital twin model of the livestock and poultry manure fermentation process, collect real-time data of the fermentation pile through sensors deployed at the fermentation site, and synchronize the real-time data to the digital twin model; Simulate the actual fermentation status in the digital twin model to predict fermentation development trends and potential anomalies; Set up an intelligent control strategy, which combines the mutual influence between real-time data and the execution conditions of the control action priority to determine whether there is a logical conflict in the rule engine. If so, optimize the control strategy, specifically: The control strategy constructs a comprehensive feature vector based on the coupling relationship between the predicted value of oxygen utilization efficiency and the estimated value of surface drying rate, and combines the preset control action priority to predict the risk value of logical conflict in the rule engine through a machine learning model, and compares it with a predetermined threshold to determine whether there is a logical conflict; According to the optimized control strategy, the pile turning, spraying and ventilation operations are automatically performed, and the control behavior and feedback data are recorded in real time. The digital twin model is continuously adjusted based on the recorded feedback data.

2. The method for monitoring livestock and poultry manure fermentation based on digital twin according to claim 1, characterized in that: The real-time data collected by the sensor include temperature, humidity, oxygen concentration, pH value and gas release at different layers of the pile, which are used to construct a multi-dimensional, hierarchical twin model input data set.

3. The method for monitoring livestock and poultry manure fermentation based on digital twin according to claim 1, characterized in that: Graph neural networks are used to predict fermentation trends and potential anomalies, including: The entire fermentation pile is divided into multiple monitoring sub-areas, each of which corresponds to a node in the graph; the connection relationship between nodes is established based on the spatial adjacency relationship of the pile and the heat or moisture migration path; each node is associated with multiple sensor data to form a feature vector of the node; the initial graph structure G=(V,E,X) of the pile is constructed, where: V is the node set; E is the edge set; X is the node feature matrix; The state of the heap at each moment is represented as a graph Gt, and the continuous time forms a graph sequence; combining the node feature changes in each graph with the graph structure invariance, the time series graph data input is constructed; the label data is the target variable or abnormal state mark of each node in a certain time window in the future; Use graph convolutional networks or graph attention networks to extract the spatiotemporal features of nodes; the GNN model performs convolution operations on the graph Gt at each time step to obtain the embedded representation Ht of the node; The node representation Ht is input into the subsequent time series prediction module; the temperature or humidity variables of each node in the future time step are predicted; the prediction results are scored for anomalies based on the actual historical distribution; if the predicted value of a node deviates from the normal change trend, it is marked as a potential anomaly.

4. The method for monitoring livestock and poultry manure fermentation based on digital twin according to claim 3 is characterized in that: The method for obtaining the predicted value of oxygen utilization efficiency is as follows: constructing a causal relationship diagram of oxygen utilization, determining variable nodes related to oxygen utilization efficiency, including fermentation pile temperature, humidity or moisture content, oxygen concentration, and ventilation volume, establishing a causal dependency relationship between variables, forming a directed acyclic graph, and setting the predicted value of oxygen utilization efficiency as the target variable node therein; Collect real-time monitoring data from historical fermentation processes for training Bayesian networks; During the fermentation process, the observed values ​​of temperature, humidity, oxygen concentration and ventilation volume are collected in real time; the observed values ​​are input into the Bayesian network as known nodes; The input of the observation node is used to trigger the conditional propagation algorithm in the Bayesian network; the network automatically updates the posterior probabilities of all relevant nodes; and the probability distribution result of the oxygen utilization efficiency prediction value node is obtained as the current oxygen utilization efficiency prediction value.

5. The method for monitoring livestock and poultry manure fermentation based on digital twin according to claim 4, characterized in that: An estimate of the surface drying rate is obtained as follows: High-sensitivity humidity sensors are deployed on the surface of the fermentation pile to collect data at fixed time intervals; each data sequence represents the change trajectory of surface humidity within a fixed time window, forming a humidity time series sample; an autoencoder is used to perform unsupervised feature learning on the humidity time series: the encoder part automatically extracts low-dimensional feature vectors representing the drying change trend in each time series; the decoder part attempts to restore the feature vector to the original humidity curve; after training, the encoder compresses any newly collected humidity change sequence into a fixed-dimensional feature representation; A clustering algorithm is used to cluster the feature vectors extracted by the autoencoder; each clustering result represents a typical drying mode; after clustering, a corresponding drying rate label is assigned to each type of drying mode; During the fermentation process, the system continuously collects the current surface humidity time series and uses the trained autoencoder to perform feature compression on the current series; Match the feature vector with the cluster center to determine which drying mode the current sequence belongs to; According to the category, the estimated surface drying rate for the current time period is output accordingly.

6. The method for monitoring livestock and poultry manure fermentation based on digital twin according to claim 5, characterized in that: The predicted value of oxygen utilization efficiency and the estimated surface drying rate are converted into a comprehensive feature vector, which is used as the input of the machine learning model. Combined with the currently set control action priority, the machine learning model uses each group of comprehensive feature vectors to predict the risk value label of logical conflict in the rule engine as the prediction target, and takes minimizing the sum of prediction errors of risk value labels of logical conflict in all rule engines as the training target. The machine learning model is trained until the sum of prediction errors converges, and the model training is stopped. The risk value of logical conflict in the rule engine is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

7. The method for monitoring livestock and poultry manure fermentation based on digital twin according to claim 6, characterized in that: The obtained risk value of logical conflict in the rule engine is compared with the preset threshold. If the risk value of logical conflict in the rule engine is greater than or equal to the preset threshold, it means that there is a logical conflict in the rule engine. At this time, a warning signal is generated and the control strategy is optimized; if the risk value of logical conflict in the rule engine is less than the preset threshold, it means that there is no logical conflict in the rule engine. At this time, no warning signal is generated and no additional adjustment is required.

8. The method for monitoring livestock and poultry manure fermentation based on digital twin according to claim 1, characterized in that: The digital twin model is continuously adjusted based on the recorded feedback data, specifically including: in the initial stage, a digital twin model is constructed to simulate the fermentation process of livestock and poultry manure, and a set of data structures for recording is established to associate the correspondence between each control behavior, environmental response and model output; multiple small sample task sets are constructed using historical feedback data, each task represents a specific fermentation sub-scenario, and each task contains a training set and a verification set; the digital twin model is gradually fine-tuned in each task; the initial parameters of the model are optimized through repeated iterations of multiple tasks. In actual operation, after each control behavior is executed, the collected feedback data is immediately constructed as a new task; the new task is input into the trained meta-learning framework for iteration, the current twin model is locally fine-tuned, and the fine-tuned model is retained as the current scene sub-model; the system regularly verifies the prediction performance of the current twin sub-model; if the error decreases or meets the self-set threshold, the system will extract the fine-tuned parameter changes, feed them back to the main model parameter set, and update the main model or add it to the experience model pool.

Images