New energy hybrid energy storage system dynamic adaptive control method based on machine learning

Abstract

The invention provides a new energy hybrid energy storage system dynamic adaptive control method based on machine learning. The method comprises the following steps: acquiring sensor data in an organic solid waste anaerobic fermentation treatment process, wherein the sensor data comprises temperature data, pH value data and gas component data; preprocessing the sensor data by adopting wavelet transform to obtain preprocessed sensor data; inputting the preprocessed sensor data into a pre-trained integrated classifier, wherein the integrated classifier comprises an XGBoost classifier, a K neighbor classifier and a BP neural network classifier; performing classification processing on the preprocessed sensor data through the integrated classifier to obtain a fermentation state classification result; and determining control parameters of the anaerobic fermentation process according to the fermentation state classification result, wherein the fermentation state classification result comprises a normal fermentation state, an acid accumulation state, a low methane yield state, a temperature abnormal state and a pH value abnormal state.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a dynamic adaptive control method for a new energy hybrid energy storage system based on machine learning. Background Technology

[0002] A core technical challenge in the anaerobic fermentation of organic solid waste is how to achieve accurate identification and real-time intelligent control of the fermentation state. The complexity of this problem lies in several aspects: First, the fermentation system is a complex biochemical process characterized by multiple variables, nonlinearity, and strong coupling. The various parameters influence each other, making it difficult to establish an accurate mathematical model. Second, the raw data collected by sensors often contains noise and interference, affecting the accuracy of state identification. Third, a single identification method cannot comprehensively capture the dynamic characteristics of the fermentation process, easily leading to misjudgments. Finally, changes in the fermentation state have a lag effect, and traditional fixed-parameter control strategies are ill-suited to fluctuations in operating conditions. These factors make it difficult to accurately determine whether fermentation is in a normal state or experiencing abnormalities such as acid accumulation or low gas production during actual operation, thus hindering timely and effective control measures. Failure to solve this problem will directly affect the efficiency and stability of organic solid waste resource utilization, restricting the widespread application of the entire treatment process. Therefore, overcoming the limitations of traditional methods and achieving intelligent monitoring and precise control of the anaerobic fermentation process is a critical technical challenge that urgently needs to be addressed. Summary of the Invention

[0003] This invention provides a dynamic adaptive control method for a new energy hybrid energy storage system based on machine learning.

[0004] The dynamic adaptive control method for new energy hybrid energy storage systems based on machine learning, as described in this invention, includes the following steps:

[0005] S101. Acquire sensor data during the anaerobic fermentation treatment of organic solid waste, wherein the sensor data includes temperature data, pH value data, and gas composition data;

[0006] S102. The sensor data is preprocessed using wavelet transform to obtain preprocessed sensor data;

[0007] S103. Input the preprocessed sensor data into a pre-trained ensemble classifier, which includes an XGBoost classifier, a K-nearest neighbor classifier, and a BP neural network classifier;

[0008] S104. The preprocessed sensor data is classified using the integrated classifier to obtain the fermentation state classification result;

[0009] S105. Determine the control parameters of the anaerobic fermentation process based on the fermentation state classification results, wherein the fermentation state classification results include normal fermentation state, acid accumulation state, low methane production state, abnormal temperature state, and abnormal pH state.

[0010] Furthermore, in step S102, the preprocessing of the sensor data using wavelet transform specifically involves:

[0011] S1021. The sensor data is decomposed into multiple levels using the Daubechies wavelet basis function.

[0012] S1022. Remove noise interference from the sensor data through multi-level decomposition;

[0013] S1023. Extract useful features from the sensor data based on the results of the multi-level decomposition;

[0014] S1024. Generate preprocessed sensor data based on useful features. The preprocessed sensor data has a higher signal-to-noise ratio and clearer feature representation.

[0015] Furthermore, in step S103, the specific steps for training the sensor data using the XGBoost classifier are as follows:

[0016] S1031. Construct multiple regression trees as weak learners;

[0017] S1032. Use the gradient boosting method to integrate multiple regression trees;

[0018] S1033. Set the target loss function and add a regularization penalty term;

[0019] S1034. Train the XGBoost classifier by iteratively optimizing the target loss function;

[0020] S1035. Determine the fermentation state category corresponding to the preprocessed sensor data based on the output of the XGBoost classifier. The fermentation state category is used for subsequent control parameter adjustment.

[0021] Furthermore, in step S103, the specific steps for training the sensor data using the K-nearest neighbor classifier are as follows:

[0022] S1036. Calculate the distance between the sample to be classified and the training samples;

[0023] S1037. Determine the K nearest neighbor samples based on distance;

[0024] S1038. Calculate the category distribution of K neighbor samples;

[0025] S1039. Use a voting mechanism to determine the category of the sample to be classified;

[0026] S10310 Output the classification results of the sample to be classified. The classification results reflect the fermentation process status corresponding to the preprocessed sensor data.

[0027] Furthermore, in step S103, the specific steps for training the sensor data using the BP neural network classifier are as follows:

[0028] S10311. Construct a multi-layer feedforward neural network structure, which includes an input layer, hidden layers, and an output layer.

[0029] S10312. Set Dropout layers between hidden layers to prevent overfitting;

[0030] S10313. Introduce nonlinear transformation using the ReLU activation function;

[0031] S10314. Use the softmax activation function in the output layer for multi-class classification;

[0032] S10315. The BP neural network classifier is trained by the backpropagation algorithm. The BP neural network classifier is used to identify the fermentation state pattern of the preprocessed sensor data.

[0033] Furthermore, in step S105, the control parameters for the anaerobic fermentation process are determined based on the fermentation state classification results as follows:

[0034] S1051. If the integrated classifier outputs a normal fermentation state, then the current control parameters remain unchanged.

[0035] S1052. If the integrated classifier outputs an acid accumulation status, then adjust the pH control parameter.

[0036] S1053. If the integrated classifier outputs a low methane production status, adjust the temperature and stirring control parameters.

[0037] S1054. If the output temperature of the integrated classifier is abnormal, adjust the heating or cooling control parameters.

[0038] S1055. If the integrated classifier outputs an abnormal pH value, adjust the buffer dosing control parameters.

[0039] Furthermore, in step S101, acquiring sensor data during the anaerobic fermentation treatment of organic solid waste specifically involves:

[0040] S1011. Real-time monitoring of temperature changes inside the anaerobic fermentation reactor using a temperature sensor;

[0041] S1012. Real-time monitoring of pH changes within the anaerobic fermentation reactor using a pH sensor;

[0042] S1013. Real-time monitoring of the gas composition generated during the anaerobic fermentation process using a gas composition sensor;

[0043] S1014. Collect data on temperature changes, pH changes, and gas composition according to a preset sampling frequency;

[0044] S1015. The collected data is time-series labeled to form sensor data for subsequent machine learning analysis and processing.

[0045] Furthermore, the adjustment and determination of the control parameters specifically involves:

[0046] S1056. Query the preset control strategy database based on the fermentation state classification results;

[0047] S1057. Obtain the corresponding parameter adjustment scheme from the control strategy database;

[0048] S1058, Calculate the execution priority of the parameter adjustment scheme;

[0049] S1059. Adjust control parameters according to execution priority order;

[0050] S10510, Monitor the system response after the control parameters are adjusted. The system response is used to verify the control effect and provide feedback information for the next round of control decisions.

[0051] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0052] This invention discloses a dynamic adaptive control method for a new energy hybrid energy storage system based on machine learning. Addressing the core problem of accurate identification and real-time control of fermentation states during anaerobic fermentation, this invention acquires multi-dimensional sensor data such as temperature, pH, and gas composition, and uses Daubechies wavelet transform to preprocess the raw data for noise reduction, effectively removing noise interference and baseline drift. Based on this, three classifiers—XGBoost, BP neural network, and K-nearest neighbors—are constructed, utilizing gradient boosting, backpropagation, and instance similarity principles respectively to achieve multi-angle identification of fermentation states. The weight coefficients of each model are determined through performance evaluation indicators, and an ensemble classifier is constructed to accurately classify five states: normal fermentation, acid accumulation, low methane production, abnormal temperature, and abnormal pH. Corresponding control commands are automatically generated based on the classification results, achieving adaptive adjustment of pH, temperature, and stirring speed. This invention significantly improves the accuracy of anaerobic fermentation state identification and the intelligence level of the control system, providing an efficient technical solution for the resource utilization of organic solid waste. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 The effect of ECG data preprocessing;

[0055] Figure 3 This is a diagram of the BP neural network structure.

[0056] Figure 4 This is a schematic diagram of the confusion matrix results for the XGBoost model;

[0057] Figure 5 This is a schematic diagram of the XGBoost_ROC curve results;

[0058] Figure 6 This is a schematic diagram of the confusion matrix results of the BP neural network model;

[0059] Figure 7 This is a schematic diagram of the ROC curve results for a BP neural network.

[0060] Figure 8 This is a schematic diagram of the confusion matrix results for the KNN model;

[0061] Figure 9 This is a schematic diagram of the KNN_ROC curve results;

[0062] Figure 10 This is a performance comparison chart for each model. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] like Figures 1 to 10 As shown, the dynamic adaptive control method for a new energy hybrid energy storage system based on machine learning of this invention includes the following steps:

[0065] S101. Acquire sensor data during the anaerobic fermentation treatment of organic solid waste, wherein the sensor data includes temperature data, pH value data, and gas composition data;

[0066] S102. The sensor data is preprocessed using wavelet transform to obtain preprocessed sensor data;

[0067] S103. Input the preprocessed sensor data into a pre-trained ensemble classifier, which includes an XGBoost classifier, a K-nearest neighbor classifier, and a BP neural network classifier;

[0068] S104. The preprocessed sensor data is classified using the integrated classifier to obtain the fermentation state classification result;

[0069] S105. Determine the control parameters of the anaerobic fermentation process based on the fermentation state classification results, wherein the fermentation state classification results include normal fermentation state, acid accumulation state, low methane production state, abnormal temperature state, and abnormal pH state.

[0070] Specifically, in step S102, the preprocessing of the sensor data using wavelet transform is as follows:

[0071] S1021. The sensor data is decomposed into multiple levels using the Daubechies wavelet basis function.

[0072] S1022. Remove noise interference from the sensor data through multi-level decomposition;

[0073] S1023. Extract useful features from the sensor data based on the results of the multi-level decomposition;

[0074] S1024. Generate preprocessed sensor data based on useful features. The preprocessed sensor data has a higher signal-to-noise ratio and clearer feature representation.

[0075] In step S103, the specific steps for training the sensor data using the XGBoost classifier are as follows:

[0076] S1031. Construct multiple regression trees as weak learners;

[0077] S1032. Use the gradient boosting method to integrate multiple regression trees;

[0078] S1033. Set the target loss function and add a regularization penalty term;

[0079] S1034. Train the XGBoost classifier by iteratively optimizing the target loss function;

[0080] S1035. Determine the fermentation state category corresponding to the preprocessed sensor data based on the output of the XGBoost classifier. The fermentation state category is used for subsequent control parameter adjustment.

[0081] Specifically, in step S103, the specific steps for training the sensor data using the K-nearest neighbor classifier are as follows:

[0082] S1036. Calculate the distance between the sample to be classified and the training samples;

[0083] S1037. Determine the K nearest neighbor samples based on distance;

[0084] S1038. Calculate the category distribution of K neighbor samples;

[0085] S1039. Use a voting mechanism to determine the category of the sample to be classified;

[0086] S10310 Output the classification results of the sample to be classified. The classification results reflect the fermentation process status corresponding to the preprocessed sensor data.

[0087] Specifically, in step S103, the specific steps for training the sensor data using the BP neural network classifier are as follows:

[0088] S10311. Construct a multi-layer feedforward neural network structure, which includes an input layer, hidden layers, and an output layer.

[0089] S10312. Set Dropout layers between hidden layers to prevent overfitting;

[0090] S10313. Introduce nonlinear transformation using the ReLU activation function;

[0091] S10314. Use the softmax activation function in the output layer for multi-class classification;

[0092] S10315. The BP neural network classifier is trained by the backpropagation algorithm. The BP neural network classifier is used to identify the fermentation state pattern of the preprocessed sensor data.

[0093] Specifically, in step S105, determining the control parameters for the anaerobic fermentation process based on the fermentation state classification results is as follows:

[0094] S1051. If the integrated classifier outputs a normal fermentation state, then the current control parameters remain unchanged.

[0095] S1052. If the integrated classifier outputs an acid accumulation status, then adjust the pH control parameter.

[0096] S1053. If the integrated classifier outputs a low methane production status, adjust the temperature and stirring control parameters.

[0097] S1054. If the output temperature of the integrated classifier is abnormal, adjust the heating or cooling control parameters.

[0098] S1055. If the integrated classifier outputs an abnormal pH value, adjust the buffer dosing control parameters.

[0099] The adjustment and determination of the control parameters are specifically as follows:

[0100] S1056. Query the preset control strategy database based on the fermentation state classification results;

[0101] S1057. Obtain the corresponding parameter adjustment scheme from the control strategy database;

[0102] S1058, Calculate the execution priority of the parameter adjustment scheme;

[0103] S1059. Adjust control parameters according to execution priority order;

[0104] S10510, Monitor the system response after the control parameters are adjusted. The system response is used to verify the control effect and provide feedback information for the next round of control decisions.

[0105] Specifically, in step S101, acquiring sensor data during the anaerobic fermentation treatment of organic solid waste involves:

[0106] S1011. Real-time monitoring of temperature changes inside the anaerobic fermentation reactor using a temperature sensor;

[0107] S1012. Real-time monitoring of pH changes within the anaerobic fermentation reactor using a pH sensor;

[0108] S1013. Real-time monitoring of the gas composition generated during the anaerobic fermentation process using a gas composition sensor;

[0109] S1014. Collect data on temperature changes, pH changes, and gas composition according to a preset sampling frequency;

[0110] S1015. The collected data is time-series labeled to form sensor data for subsequent machine learning analysis and processing.

[0111] In this invention, data is sampled once per hour to record sensor data at different fermentation stages. The data is divided into five categories: normal fermentation (N), acid accumulation (A), low methane production (V), abnormal temperature (L), and abnormal pH (R).

[0112] Many factors affect the automatic classification of ECG signals during data preprocessing, the most important of which is noise interference, such as baseline drift and motion artifacts. The purpose of preprocessing is to remove outliers from the data, obtain a clean signal, and improve the reliability of subsequent experiments.

[0113] Wavelet transform is a time-frequency analysis method that decomposes a signal into a linear combination of wavelet functions. Wavelet functions possess excellent time-frequency localization properties, effectively capturing useful information in the signal while filtering out noise. In this study, the Daubechies wavelet basis function (DB5) was selected to perform a 9-level decomposition on the original ECG signal. For each heartbeat, data from the 99 points before and 201 points after the R-wave were extracted, implemented using the PyWavelets (pywt) library. The preprocessing algorithm's effect is shown below. Figure 2 Show.

[0114] The specific steps for data classification are as follows.

[0115] XGBoost (Extreme Gradient Boosting) is a highly efficient gradient boosting decision tree algorithm. It improves upon the original GBDT algorithm, significantly enhancing model performance. Its core principle is ensemble learning—the boosting concept—which integrates multiple weak learners into a single strong learner. This involves multiple trees making decisions collaboratively, with each tree's result representing the difference between the target value and the predictions of all previous trees. The final result is obtained by summing all the results, thus improving the overall model performance.

[0116] The main component of the XGBoost algorithm is the regression tree. XGBoost combines multiple regression trees using an ensemble method for classification. The initial tree can be represented as...

[0117]

[0118] The model for round t can be represented as:

[0119]

[0120] in, For the model predictions in the first t-1 rounds, f t (x i Let f(t) be the function representation of the t-th tree. Adding a new tree will decrease the objective loss function value. However, if the number of leaf nodes is too large, it will lead to overfitting of the model and affect the final classification result. Therefore, it is necessary to limit the number of leaf nodes and add a penalty term Ω(f) to the objective function. t ).

[0121]

[0122] Where T is the number of leaf nodes, γ and λ are the penalty strengths, and λ is the weight of the leaf node. The L2 penalty term is used, and the final objective function is...

[0123]

[0124] The parameters of the XGBoost ensemble classifier are shown in Table 1.

[0125] Table 1. Parameter settings for the XGBoost ensemble classifier

[0126] Parameter name meaning Setting value n_estimators Number of decision trees 200 max_depth Maximum depth of tree 5 learning_rate Learning rate 0.1 random_state random seed parameters 42 num_class Number of categories 5

[0127] A backpropagation (BP) neural network is a multi-layer feedforward neural network, mainly composed of an input layer, hidden layers, and an output layer. Its key feature is that the signal propagates forward while the error propagates backward. "BP" stands for "Back Propagation," which is also its primary learning algorithm. It minimizes the error between the predicted output and the actual value using gradient descent. The topology of this model is as follows: Figure 3 As shown.

[0128] The specific steps are as follows:

[0129] Step 1: Construct a multi-layer feedforward neural network with an input dimension of 300, corresponding to the feature dimensions of the ECG data. The network consists of four layers, with ReLU activation function used between each layer to introduce non-linearity. The output layer uses softmax activation function for multi-class classification. To prevent overfitting, Dropout layers are added between the fully connected layers, with dropout rates of 0.3, 0.3, and 0.2, respectively.

[0130] Step 2: During the compilation phase, select the Adam optimizer, use classification cross-entropy as the loss function, and select accuracy as the evaluation metric.

[0131] Step 3: Training configuration. In the training settings, the model is trained for 10 epochs, with 128 samples processed in each batch. To monitor the training process and prevent overfitting, a 20% validation set split is set.

[0132] The parameters of the BP neural network classifier are shown in Table 2.

[0133] Table 2 Parameter settings for the BP neural network classifier

[0134]

[0135]

[0136] The k-Nearest Neighbor (KNN) algorithm is an instance-based supervised learning method first proposed by Cover and Hart in 1967. This algorithm is based on the fundamental assumption that similar instances in the feature space have similar class attributes.

[0137] The theoretical basis of the K-Nearest Neighbors algorithm can be traced back to Bayesian decision theory. Given an instance x, the optimal prediction should be the class with the highest posterior probability P(y|x).

[0138] y = argc j -maxP(c j |x)

[0139] The KNN algorithm estimates posterior probabilities using a non-parametric method. Specifically, for any point x in the feature space, the posterior probability of its class j can be estimated using the neighborhood N. k Estimate using sample statistics from (x):

[0140]

[0141] Where I(·) is the indicator function, N k(x) This represents the k nearest neighbors of instance x.

[0142] The parameter settings for the K-nearest neighbor classifier are shown in Table 3.

[0143] Table 3. Parameter settings for the K-nearest neighbor classifier

[0144] Parameter name meaning Setting value n_neighbors Number of recent neighbors 5 leaf_size Sample size 30 n_jobs CPU core count -1

[0145] Experiments showed that all three models (BP Network, KNN, and XGBoost) demonstrated good performance on classification tasks.

[0146] The XGBoost model exhibits relatively balanced performance, such as Figure 4 , Figure 5 As shown, the average AUC is 0.9965, and it performs exceptionally well in class V and class L classification, with AUC values ​​of 0.9995 and 0.9998, respectively.

[0147] like Figure 6 , Figure 7 As shown, the average AUC of the BP Network model reaches 0.9845, with the best performance in the classification of L and R classes, with AUC values ​​of 0.9998 and 0.9998, respectively.

[0148] The overall performance of the KNN model is slightly lower, such as Figure 8 , Figure 9 As shown, the average AUC is 0.9845, but it shows high accuracy when dealing with N-class samples, with the confusion matrix showing an accuracy of 1.00.

[0149] Table 4 shows a comparison of the performance evaluation indicators of each model.

[0150] Table 4 Comparison of Performance Evaluation Indicators for Each Model

[0151]

[0152] As shown in Table 4 and Figure 10 As shown, performance evaluations of three different classifier models on the anaerobic fermentation system revealed that all classifiers achieved excellent classification results. The BP neural network classifier performed best, achieving an accuracy of 98.89%. The XGBoost classifier followed closely, with an ROC_AUC of 99.65%, demonstrating the strongest class discrimination ability. While the K-nearest neighbor classifier had relatively lower overall metrics, its accuracy still reached a high level of 98.63%. Overall, all three models exhibited extremely high reliability in the anaerobic fermentation system performance evaluation, with their minor performance differences mainly reflected in their ability to handle a small number of complex samples.

[0153] Finally, it should be emphasized that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic adaptive control method for a new energy hybrid energy storage system based on machine learning, characterized in that, The method includes the following steps: S101. Acquire sensor data during the anaerobic fermentation treatment of organic solid waste, wherein the sensor data includes temperature data, pH value data, and gas composition data; S102. The sensor data is preprocessed using wavelet transform to obtain preprocessed sensor data; S103. Input the preprocessed sensor data into a pre-trained ensemble classifier, which includes an XGBoost classifier, a K-nearest neighbor classifier, and a BP neural network classifier; S104. The preprocessed sensor data is classified using the integrated classifier to obtain the fermentation state classification result; S105. Determine the control parameters of the anaerobic fermentation process based on the fermentation state classification results, wherein the fermentation state classification results include normal fermentation state, acid accumulation state, low methane production state, abnormal temperature state, and abnormal pH state.

2. The method as described in claim 1, characterized in that, In step S102, the preprocessing of the sensor data using wavelet transform specifically involves: S1021. The sensor data is decomposed into multiple levels using the Daubechies wavelet basis function. S1022. Remove noise interference from the sensor data through multi-level decomposition; S1023. Extract useful features from the sensor data based on the results of the multi-level decomposition; S1024. Generate preprocessed sensor data based on useful features. The preprocessed sensor data has a higher signal-to-noise ratio and clearer feature representation.

3. The method as described in claim 1, characterized in that, In step S103, the specific steps for training the sensor data using the XGBoost classifier are as follows: S1031. Construct multiple regression trees as weak learners; S1032. Use the gradient boosting method to integrate multiple regression trees; S1033. Set the target loss function and add a regularization penalty term; S1034. Train the XGBoost classifier by iteratively optimizing the target loss function; S1035. Determine the fermentation state category corresponding to the preprocessed sensor data based on the output of the XGBoost classifier. The fermentation state category is used for subsequent control parameter adjustment.

4. The method as described in claim 1, characterized in that, In step S103, the specific steps for training the sensor data using the K-nearest neighbor classifier are as follows: S1036. Calculate the distance between the sample to be classified and the training samples; S1037. Determine the K nearest neighbor samples based on distance; S1038. Calculate the category distribution of K neighbor samples; S1039. Use a voting mechanism to determine the category of the sample to be classified; S10310 Output the classification results of the sample to be classified. The classification results reflect the fermentation process status corresponding to the preprocessed sensor data.

5. The method as described in claim 1, characterized in that, In step S103, the specific steps for training the sensor data using the BP neural network classifier are as follows: S10311. Construct a multi-layer feedforward neural network structure, which includes an input layer, hidden layers, and an output layer. S10312. Set Dropout layers between hidden layers to prevent overfitting; S10313. Introduce nonlinear transformation using the ReLU activation function; S10314. Use the softmax activation function in the output layer for multi-class classification; S10315. The BP neural network classifier is trained by the backpropagation algorithm. The BP neural network classifier is used to identify the fermentation state pattern of the preprocessed sensor data.

6. The method as described in claim 1, characterized in that, In step S105, the control parameters for the anaerobic fermentation process are determined based on the fermentation state classification results as follows: S1051. If the integrated classifier outputs a normal fermentation state, then the current control parameters remain unchanged. S1052. If the integrated classifier outputs an acid accumulation status, then adjust the pH control parameter. S1053. If the integrated classifier outputs a low methane production status, adjust the temperature and stirring control parameters. S1054. If the output temperature of the integrated classifier is abnormal, adjust the heating or cooling control parameters. S1055. If the integrated classifier outputs an abnormal pH value, adjust the buffer dosing control parameters.

7. The method as described in claim 1, characterized in that, In step S101, acquiring sensor data during the anaerobic fermentation treatment of organic solid waste specifically involves: S1011. Real-time monitoring of temperature changes inside the anaerobic fermentation reactor using a temperature sensor; S1012. Real-time monitoring of pH changes within the anaerobic fermentation reactor using a pH sensor; S1013. Real-time monitoring of the gas composition generated during the anaerobic fermentation process using a gas composition sensor; S1014. Collect data on temperature changes, pH changes, and gas composition according to a preset sampling frequency; S1015. The collected data is time-series labeled to form sensor data for subsequent machine learning analysis and processing.

8. The method as described in claim 6, characterized in that, The adjustment and determination of the control parameters are specifically as follows: S1056. Query the preset control strategy database based on the fermentation state classification results; S1057. Obtain the corresponding parameter adjustment scheme from the control strategy database; S1058, Calculate the execution priority of the parameter adjustment scheme; S1059. Adjust control parameters according to execution priority order; S10510, Monitor the system response after the control parameters are adjusted. The system response is used to verify the control effect and provide feedback information for the next round of control decisions.

Images