A method and system for predicting pellet strength based on an optimized decision neural network

Abstract

The application provides a kind of pellet strength prediction method and system based on optimization decision neural network, the method comprises: S1, collecting PLC real-time data and report historical data in the production process of pellet;S2, the PLC real-time data and / or report historical data are input into the trained multilayer perception machine prediction model, and first strength prediction value is obtained;S3, the time series data of pellet strength in report historical data is input into the trained long short-term memory network prediction model, and second strength prediction value is obtained;S4, first strength prediction value and second strength prediction value are input into optimization decision model, the weight of first strength prediction value and second strength prediction value is dynamically adjusted, and the final pellet strength prediction result is output.The application realizes accurate and stable prediction of pellet compressive strength by double model parallel prediction and intelligent decision fusion, and solves the problem of low prediction accuracy and poor robustness caused by data acquisition failure in the prior art.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the intensity of pellet ore, specifically to a method and system for predicting the intensity of pellet ore based on an optimized decision neural network, belonging to the field of pellet production technology. Background Technology

[0002] Iron ore pellets are an important processed iron ore product. Due to their high iron content, good reducibility, high compressive strength, and low impurities, they have become an indispensable high-quality raw material for blast furnace ironmaking. The production process of iron ore pellets mainly includes raw material preparation, pelletizing, drying, roasting, cooling, and screening. Among these, the strength of the pellets (especially their compressive strength) is the core indicator for measuring their quality, directly determining their performance during subsequent transportation, storage, and blast furnace smelting. Pellets with insufficient strength are prone to producing powder before entering the furnace, affecting the permeability of the blast furnace charge, thereby reducing smelting efficiency and increasing energy consumption.

[0003] The strength formation mechanism of iron ore pellets is complex, influenced by the coupling effects of multiple nonlinear factors, including raw material properties (such as iron grade, basicity, and binder ratio), equipment conditions (such as chain grate speed and rotary kiln rotation speed), and process parameters (such as temperature at each stage and fuel quantity). Traditional mathematical models based on metallurgical mechanisms often have limited predictive accuracy due to the involvement of numerous intermediate variables that are difficult to measure precisely. With the development of industrial internet and big data technologies, data-driven machine learning methods, such as support vector machines (SVM) and backpropagation neural networks, are beginning to be applied to iron ore pellet strength prediction.

[0004] For example, patent document CN104480300A discloses a pellet production method based on predicting the compressive strength of pellets in a rotary kiln. This method relies on a thermal model for prediction, which is highly complex and depends on numerous physical assumptions. Patent document CN116305779A discloses an optimized roasting control method for a belt roaster, which uses linear regression analysis, making it difficult to capture the nonlinear relationship between process parameters and strength. Patent document CN113793308A discloses a neural network-based intelligent pellet quality rating method, focusing on microscopic feature analysis and quality grading in image processing rather than precise numerical prediction of strength. Patent document CN116334382A discloses a data-driven method for improving pellet compressive strength, employing a support vector machine (SVM) model. However, it has limitations in handling large-scale industrial data and complex nonlinear fitting capabilities, and is highly dependent on the completeness of real-time PLC data.

[0005] Furthermore, existing technical solutions generally rely on a single prediction model and have extremely high requirements for the completeness and real-time performance of input data. In actual industrial production sites, PLC data acquisition often faces problems such as signal interference, transmission delays, and storage constraints, while manually entered report data may be delayed or contain errors. These data quality issues directly lead to distortion of model input parameters, thereby seriously affecting the accuracy and stability of prediction results. Therefore, there is an urgent need for a pellet intensity prediction method that can tolerate data acquisition defects, integrate multi-source information, and achieve high prediction accuracy. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a method and system for predicting the compressive strength of pellets based on an optimized decision neural network. This invention employs a dual-algorithm parallel prediction approach, fusing prediction using a multilayer perceptron (MLP) based on real-time process parameters and a long short-term memory (LSTM) network based on historical time-series data, and introducing an optimized decision mechanism. This achieves accurate and stable prediction of the compressive strength of pellets, solving the problems of low prediction accuracy and poor robustness caused by data acquisition failures in existing technologies.

[0007] According to a first embodiment of the present invention, a method for predicting the intensity of pellet ore based on an optimized decision neural network is provided.

[0008] A method for predicting the intensity of pellet ore based on an optimization decision neural network, the method comprising the following steps:

[0009] S1. Data Acquisition Steps: Collect real-time PLC data and historical report data during the pellet production process.

[0010] S2. First prediction step: Input the real-time data from the PLC and / or historical data from reports into the trained first prediction model to obtain the first intensity prediction value. The first prediction model is a model built based on a multilayer perceptron (MLP).

[0011] S3. Second prediction step: Input the time-series data of pellet intensity from the historical data of the report into the trained second prediction model to obtain the second intensity prediction value. The second prediction model is a model built based on a Long Short-Term Memory (LSTM) network.

[0012] S4. Optimization Decision Steps: Input the first intensity prediction value and the second intensity prediction value into the optimization decision model, dynamically adjust the weights of the first intensity prediction value and the second intensity prediction value, and output the final pellet intensity prediction result.

[0013] In this invention, in step S4, the calculation formula of the optimization decision model is:

[0014]

[0015] In the formula: Q represents the final predicted strength of the pellet. MLP This is the first intensity prediction value. Q LSTM K is the weighting coefficient of the first intensity prediction value. J is the weighting coefficient of the second intensity prediction value. L is the variance weighting coefficient of the first and second intensity prediction values. β is the optimization decision correction coefficient. Where K+J=1, and 0≤K≤1, 0≤J≤1. -1≤L≤1. β∈[-500, 500], preferably β∈[-300, 300].

[0016] In this invention, step S4 further includes an exception handling strategy:

[0017] When the input parameters of the first prediction model are detected to exceed the preset threshold, or the first intensity prediction value exceeds the normal range of pellet intensity, K=0, L=0, J=1, β=0 are set.

[0018] When the second intensity prediction value is detected to be outside the normal range of pellet intensity, set J=0, L=0, K=1, β=0.

[0019] In this invention, in step S1, the historical data of the report includes the sampling time and the test value of the pellet strength test.

[0020] Preferably, step S1 also includes a data matching sub-step: using the sampling time of the pellet strength test as the reference point T0, and based on the lag time of the pellet production process, calculating and matching the corresponding PLC real-time data at that time point. Wherein:

[0021] The raw material proportioning node time is T0-A. The green pellet feeding node time is T0-B. The chain grate process node time is T0-C. The rotary kiln process node time is T0-D. And the relationship between the lag times satisfies: A>B>C>D.

[0022] In this invention, the value of A ranges from 180 to 220 min. The value of B ranges from 110 to 130 min. The value of C ranges from 100 to 120 min. The value of D ranges from 70 to 90 min.

[0023] In this invention, in step S1, the real-time data from the PLC includes one or more of the following: bentonite ratio in the mixture, Fe3O4 content in the mixture, Fe2O3 content in the mixture, alkalinity of the mixture, fixed carbon content in the mixture, green pellet feed rate, wind box temperature and hood temperature of the chain grate's pH section, chain grate speed, rotary kiln temperature, fuel quantity, fuel consumption per unit, and rotary kiln rotation speed.

[0024] In this invention, in step S2, the first prediction model includes an input layer, at least one hidden layer, and an output layer. The number of nodes in the input layer corresponds to the number of types of real-time PLC data input, and the number of nodes in the output layer is 1, used to output a first intensity prediction value.

[0025] In this invention, in step S3, the second prediction model includes an input layer, a long short-term memory network hidden layer, and an output layer. The input layer is used to input a feature vector composed of the pellet intensity values ​​at consecutive time nodes, and the output layer is used to output the second intensity prediction value at the target time node.

[0026] In this invention, the method further includes a model training step:

[0027] Using historical PLC data and corresponding pellet strength test values ​​as training samples, the multilayer perceptron is trained to obtain the first prediction model.

[0028] The second prediction model is obtained by training a long short-term memory network using a continuous time series of historical pellet strength test values ​​as training samples.

[0029] Using the outputs of the first and second prediction models as inputs and the actual pellet strength test values ​​as targets, the weight coefficients in the optimization decision model are optimized using validation set data.

[0030] According to a second embodiment of the present invention, a pellet intensity prediction system based on an optimized decision neural network is provided.

[0031] A pellet intensity prediction system based on an optimization decision neural network or a pellet intensity prediction system for the method described in the first embodiment, the system comprising:

[0032] Data acquisition module: used to collect PLC process parameters in real time as PLC real-time data, and to collect historical test values ​​of pellet strength as historical data for reports.

[0033] First prediction module: It has a pre-trained first prediction model built in, which is used to receive real-time data from the PLC and / or historical data from reports as input, and output the first intensity prediction value.

[0034] The second prediction module has a built-in trained second prediction model, which is used to receive a time-series feature vector composed of the pellet intensity values ​​of consecutive time nodes in the historical data of the report, and output the second intensity prediction value.

[0035] Optimization Decision Module: It has a built-in optimization decision model, which is used to receive the first intensity prediction value and the second intensity prediction value, and perform anomaly detection and dynamic weighted decision to calculate the final pellet intensity prediction result.

[0036] Result output module: used to output the final pellet strength prediction result.

[0037] According to a third embodiment of the present invention, a computer-readable storage medium is provided.

[0038] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in the first embodiment.

[0039] The present invention constructs a dual-decision collaborative system. The first decision-making process utilizes a multilayer perceptron to deeply mine the "causal logic" between production process parameters and pellet strength, while the second decision-making process utilizes a long short-term memory network to capture the "temporal inertia" of pellet strength evolution over time. The optimized decision model integrates the advantages of both, resulting in a more stable and accurate final prediction. The present invention aims to solve the problems of low prediction accuracy and poor robustness caused by data acquisition constraints in existing technologies by constructing an architecture that integrates parallel prediction with intelligent decision-making using dual algorithms.

[0040] Specifically, this invention proposes a method and system for predicting the intensity of pellet ore based on an optimized decision neural network. The method mainly encompasses a data acquisition step, a first prediction step, a second prediction step, and an optimized decision-making step. Correspondingly, the system includes a data acquisition module (which further comprises a PLC data acquisition module and a report data acquisition module), a first prediction module (i.e., a multilayer perceptron prediction algorithm module), a second prediction module (i.e., a long short-term memory network prediction algorithm module), an optimized decision-making module, and a result output module.

[0041] The following is a detailed description of the technical solution of the present invention:

[0042] 1. Overall Plan Flow:

[0043] like Figure 1 As shown, the pellet intensity prediction method of this invention, during the model (or algorithm) training phase:

[0044] The first step of this method is to collect PLC data and report data.

[0045] The second step involves using PLC data and report data as training data in the multilayer perceptron prediction algorithm module to train the multilayer perceptron prediction algorithm, thereby obtaining the optimal multilayer perceptron prediction model; and using report data as training data in the long short-term memory network prediction module to train the long short-term memory network prediction algorithm, thereby obtaining the optimal long short-term memory network prediction model.

[0046] The third step involves using the outputs of the multilayer perceptron prediction algorithm module and the long short-term memory network prediction algorithm module, along with the report data, to perform decision optimization training and obtain the optimal optimization decision model.

[0047] The fourth step is to optimize the weight output of the decision model and complete the algorithm training.

[0048] like Figure 2 As shown, the pellet intensity prediction method of this invention is in the model (or algorithm) testing phase:

[0049] The first step of this method is to collect PLC data and report data.

[0050] The second step involves using PLC data as input data in the multilayer perceptron prediction algorithm module to obtain the optimal multilayer perceptron prediction result; and using report data as input data in the long short-term memory network prediction module to obtain the optimal long short-term memory network prediction result.

[0051] The third step is to optimize the output results of the multilayer perceptron prediction algorithm module and the long short-term memory network prediction algorithm module.

[0052] The fourth step is to output the final pellet strength prediction result to complete the algorithm prediction.

[0053] Furthermore, in the second step, the input data for the multilayer perceptron prediction algorithm module can be PLC data, report data, or a mixture of both. The multilayer perceptron model (i.e., the first prediction model) is essentially a function mapper; the input is "intensity-related features," and the output is "intensity values." In the model training phase of this invention, the multilayer perceptron is trained using both PLC data and report data (i.e., the report data provides label values, and also contains some features that can be used as input). Therefore, in the model testing phase, as long as feature data that meets the input dimension requirements can be provided, it can run normally and output prediction results. This design offers high flexibility. For example, when some PLC data is missing due to transmission failure, but the corresponding process parameter statistics for that period can be extracted from historical reports, the complete statistical values ​​can be used to replace the missing real-time data input to the model, ensuring that the prediction process is not interrupted. This greatly improves the tolerance of the entire prediction method to defects in on-site data and broadens the application scenarios of the algorithm.

[0054] 2. PLC data acquisition module:

[0055] The PLC data acquisition module is the core data interface unit, responsible for real-time acquisition of analog and digital signals from field sensors, actuators, and other equipment, and uploading them to the PLC host or upper-level database system. The acquired data includes, but is not limited to: bentonite ratio in the mixture, Fe3O4 content in the mixture, Fe2O3 content in the mixture, alkalinity of the mixture, fixed carbon content of the mixture, green pellet feed rate, wind box temperature and fume hood temperature in the pH section of the chain grate, chain grate speed, rotary kiln temperature, fuel quantity, fuel consumption per unit area, rotary kiln speed, and the acquisition time for each data point.

[0056] Because production takes a long time and the data collection times are inconsistent, data matching and processing are required after collection.

[0057] Node 1 (Raw Material Proportioning Node): Includes bentonite proportion in the mixture, Fe3O4 content in the mixture, Fe2O3 content in the mixture, alkalinity of the mixture, and fixed carbon content in the mixture, etc.

[0058] Node 2 (Green Ball Feeding Node): Includes green ball feeding amount, etc.

[0059] Node 3 (Chain grate machine process node): including the air box temperature of the pH section of the chain grate machine, the fume hood temperature of the pH section of the chain grate machine, and the speed of the chain grate machine, etc.

[0060] Node 4 (Rotary Kiln Process Node): Includes rotary kiln temperature, fuel quantity, fuel consumption per unit area, rotary kiln speed, etc.

[0061] Based on the on-site production situation, taking the sampling time of the last pellet strength test as the baseline T0, the node times of nodes 4, 3, 2, and 1 are calculated backwards (in minutes):

[0062] Node 4 = Sampling time for pellet strength testing - D = T0 - D;

[0063] Node 3 = Sampling time for pellet strength testing - C = T0 - C;

[0064] Node 2 = Sampling time for pellet strength testing - B = T0 - B;

[0065] Node 1 = Sampling time for pellet strength testing - A = T0 - A;

[0066] Furthermore, the relationship between the lag times satisfies: A > B > C > D.

[0067] Depending on the specific production conditions, the time parameter D can be set within a range of 70 to 90 minutes. Correspondingly, C ranges from 100 to 120 minutes, B from 110 to 130 minutes, and A from 180 to 220 minutes. For example, D, C, B, and A can be set to 80, 110, 120, and 200 minutes respectively to match the production data.

[0068] 3. Report Data Acquisition Module:

[0069] The report data acquisition module is a core component, used to automatically collect and integrate information from sampling and testing data sources for pellet strength (e.g., cold-pressing strength of pellets), providing fundamental support for report generation and data-driven decision-making. Report data, as the actual values ​​used for algorithm training, possesses core attributes such as accuracy, completeness, consistency, timeliness, and traceability.

[0070] The report data includes the sampling time and test values ​​of the pellet strength test samples.

[0071] 4. Multilayer Perceptron Prediction Algorithm Module:

[0072] The multilayer perceptron prediction algorithm module includes a multilayer perceptron prediction algorithm. The algorithm's input parameters include: bentonite ratio in the mixture, Fe3O4 content in the mixture, Fe2O3 content in the mixture, basicity of the mixture, fixed carbon content in the mixture, green pellet feed rate, windbox temperature and fume hood temperature in the pH section of the chain grate, chain grate speed, rotary kiln temperature, fuel quantity, fuel consumption per unit area, rotary kiln speed, and data acquisition time, etc. The algorithm's output parameter is the multilayer perceptron pellet strength prediction result (i.e., the first strength prediction value). The hidden layer can be customized according to the target characteristics.

[0073] like Figure 3 As shown, in this example, the hidden layers are two: the first layer has four nodes, a11, a12, a13, and a14; the second layer has three nodes, a21, a22, and a23. The entire network is connected using fully connected layers. The input layer is mapped to the first hidden layer after passing through a fully connected layer; the first hidden layer is then mapped to the second hidden layer after passing through another fully connected layer; and finally, the second hidden layer is mapped to the output layer after passing through a fully connected layer. Ultimately, the trained model will fit the non-linear relationship between the input and output very well, thus achieving the goal of intensity prediction.

[0074] 5. Long Short-Term Memory Network Prediction Algorithm Module:

[0075] The Long Short-Term Memory (LSTM) network prediction algorithm uses the intensity values ​​of pellet ore from the previous two time points as input data for training. By constructing intensity sequences at two consecutive time points, it mines the patterns and dependencies in the temporal dimension. Based on the continuous characteristic of pellet ore intensity changes—that is, the current intensity state evolves from recent historical states—the intensity data at consecutive time points contains key information such as trend changes and fluctuation amplitudes, providing effective input features for model learning. The training label value is explicitly set to the intensity value of the next time point, forming a supervised learning mode of "historical temporal input - future target output," ensuring that the model focuses on learning the temporal evolution patterns rather than fitting single data points.

[0076] Network structure:

[0077] Input layer: Dimension is set to 2, corresponding to the intensity values ​​of the previous time step and the previous time step in the input features;

[0078] Hidden layer: Multiple LSTM layers are set up, containing 64 neurons. This layer includes three gate control structures: forget gate, input gate, and output gate.

[0079] Forget gate: Uses the Sigmoid activation function to filter out valid information from historical data and discard invalid or interfering information;

[0080] Input gate: Combines the Sigmoid activation function and the Tanh activation function to fuse the current input features with the filtered historical information and update the state;

[0081] Output gate: Based on the updated cell state, the hidden layer output is generated using the Sigmoid activation function;

[0082] Output layer: The dimension is set to 1, and a linear activation function is used to output the predicted value of the pellet intensity at the next time step.

[0083] train:

[0084] Optimizer: The Adam optimizer is used, with an initial learning rate of 0.001, which adaptively decreases with the number of iterations.

[0085] Loss function: The mean squared error (MSE) is used as the loss function, and the formula is as follows:

[0086] MSE=1 / n Σ(y_true-y_pred)²;

[0087] Where y_true is the true label value, y_pred is the predicted value, and n is the number of samples;

[0088] Training strategy: The batch size is set to 32, the number of epochs is set to 100, and early stopping is used to prevent overfitting. When the validation set MSE shows no decreasing trend for several consecutive iterations, training is stopped, the current model parameters are saved, and the optimal LSTM prediction model is obtained.

[0089] predict:

[0090] Input reception: Receive the target prediction time node T as input;

[0091] Feature matching: Extract the intensity values ​​of multiple consecutive time nodes before the target time node T from the historical intensity time series database, and use them as the input features to be predicted;

[0092] Model inference: Input the input features to be predicted into the optimal LSTM prediction model, and generate standardized prediction values ​​through the temporal inference operation of the model gating structure;

[0093] Output result: The predicted value of pellet intensity at the target time node T is obtained.

[0094] 6. Optimize the decision-making module:

[0095] The optimized decision-making module focuses on algorithmic fusion and innovation, deeply integrating the output results of the multilayer perceptron prediction algorithm module and the long short-term memory network prediction algorithm module (Q). MLP Q LSTM By constructing an intelligent optimization decision-making mechanism, the module accurately weighs and integrates the outputs of the two algorithms, achieving a significant improvement in prediction efficiency. Addressing common issues in industrial scenarios such as data acquisition overhead, transmission delays, and data gaps in PLC and report data, this module effectively mitigates the interference of these problems on the input parameters of the prediction network through dynamic correction and fault-tolerant processing logic. This ensures the input quality of the prediction model from the source, thereby significantly improving prediction accuracy and result stability.

[0096] The formula for calculating the output Q of the optimization decision module is as follows:

[0097]

[0098] In the formula: Q represents the final predicted strength of the pellet; Q MLP Q is the first intensity prediction value; LSTM K is the weighting coefficient of the first intensity prediction value; J is the weighting coefficient of the second intensity prediction value; L is the variance weighting coefficient of the first intensity prediction value and the second intensity prediction value; β is the optimization decision correction coefficient; where K+J=1, and 0≤K≤1, 0≤J≤1; -1≤L≤1; β∈[-500, 500], preferably β∈[-300, 300].

[0099] This invention introduces a variance term into the optimized decision model, enabling the model to automatically evaluate the consistency between the MLP and LSTM branches. When the prediction results of the two branches deviate significantly, the variance term automatically adjusts the final prediction result to shift towards the more reliable branch, preventing a single algorithm error from directly affecting the global prediction result. Conversely, when the prediction results of the two branches deviate slightly, the variance term's correction to the result is minimal, maintaining the stability of the original fusion result and avoiding the introduction of additional prediction bias.

[0100] During the model training phase, the weight coefficients K, J, L, and correction coefficient β are continuously updated iteratively through the backpropagation algorithm. The optimal parameter combination is finally obtained when the validation set error converges, ensuring that the optimized decision model can adapt to the production data characteristics of different sites without requiring manual parameter adjustment, significantly lowering the threshold for on-site implementation. Compared to simple linear weighted fusion, the optimized decision formula proposed in this invention considers both the basic bias and dispersion of the two prediction results, enabling a more accurate fit between the true intensity value and the nonlinear relationship between the two single-model predictions. This further amplifies the advantages of each single model: during periods of significant fluctuation in production process parameters, the causal logic mined by the multilayer perceptron has a higher proportion, and the model automatically increases the K value, allowing the influence of process parameters on the results to dominate; during periods of stable production process parameters and continuous intensity changes, the temporal inertia captured by the long short-term memory network has a higher proportion, and the model automatically increases the J value, allowing the influence of temporal evolution patterns on the results to dominate. This achieves truly dynamic intelligent decision-making, rather than static fusion with fixed weights.

[0101] In practical industrial applications, when there are partial gaps in the data collected by the PLC, although the prediction results of the multilayer perceptron will have some deviation, the prediction results of the long short-term memory network can still remain relatively stable based on historical time series data. The optimization decision module will automatically identify the deviation between the two results and adjust the weight allocation through the variance term, ultimately obtaining a prediction result with a deviation much smaller than that of the single model, which fully demonstrates the robustness advantage of the dual-decision fusion architecture.

[0102] Furthermore, when the production parameters input to the multilayer perceptron prediction algorithm module exceed the threshold for any reason, or when the multilayer perceptron prediction result exceeds the normal range, setting K=0 (i.e., J=1), L=0, and β=0 can reduce algorithm fluctuations caused by production parameters exceeding the limits. Similarly, when the long short-term memory network prediction algorithm result exceeds the normal range, setting J=0 (i.e., K=1), L=0, and β=0 can reduce algorithm fluctuations caused by time-series prediction exceeding the limits. This invention designs this anomaly handling strategy so that when the MLP model input is abnormal or the output is unreliable due to problems with on-site data acquisition, it can automatically switch to a prediction mode relying solely on the LSTM model; and vice versa. This mechanism effectively avoids the shortcomings of a single model, solves the interference of problems such as data occupation, untimely data acquisition, and parameter exceeding limits in industrial on-site data collection on the prediction results, and greatly improves the availability and stability of the model in actual production environments.

[0103] During normal production, the optimization decision will be jointly led by the results of the multilayer perceptron prediction algorithm module and the long short-term memory network prediction algorithm module, so that the prediction model of this application can combine production parameters, test values ​​and time series information to make optimization predictions.

[0104] 7. Result Output Module:

[0105] The final pellet intensity prediction output module takes the optimized decision output from the decision module and uses it as the final pellet intensity prediction result of this method. This completes the pellet intensity prediction based on an optimized decision neural network.

[0106] Compared with the prior art, the present invention has the following beneficial technical effects:

[0107] (1) High precision and strong robustness: This invention adopts a dual-algorithm parallel prediction. By integrating multilayer perceptron (MLP) prediction based on real-time process parameters and long short-term memory network (LSTM) prediction based on historical time series data, and introducing an optimization decision mechanism, it achieves accurate and stable prediction of the compressive strength of pellets, and solves the problem of low prediction accuracy and poor robustness caused by data acquisition failure in the prior art.

[0108] (2) Flexible prediction mode: This invention can not only complete high-precision prediction based on various real-time parameters in the production process, but also break through the limitations of traditional prediction modes and directly call historical pellet strength data to build prediction models without relying on complex real-time parameter acquisition links. This flexible prediction mode reduces the cost of relying on on-site data acquisition equipment and broadens the applicable scenarios of the algorithm. It can provide efficient and reliable technical support for pellet production quality control, process optimization and other links, and help enterprises achieve refined production and intelligent decision-making upgrades.

[0109] (3) Adaptive fusion and variance control: The optimization decision model proposed in this invention considers both the basic bias and dispersion of the dual prediction results, which can more accurately fit the nonlinear relationship between the true intensity value and the prediction results of the two single models, further amplifying the advantages of each single model. The optimization decision model also introduces a variance term, which can automatically evaluate the consistency between the two branches of MLP and LSTM: when the prediction results of the two branches have a large deviation, the variance term will automatically adjust the final prediction result to shift towards the more reliable branch, avoiding the direct impact of a single algorithm error on the global prediction result; while when the prediction results of the two branches have a small deviation, the correction magnitude of the variance term is very small, which can maintain the stability of the original fusion result and will not introduce additional prediction bias.

[0110] (4) High fault tolerance and scenario adaptability: This invention designs an anomaly handling strategy. When the input of the MLP model is abnormal or the output is unreliable due to problems with on-site data acquisition, it can automatically switch to the prediction mode that relies solely on the LSTM model; and vice versa. This mechanism effectively avoids the shortcomings of a single model and solves the interference of problems such as data acquisition occupation, untimely acquisition, and parameter out-of-bounds in industrial field on the prediction results, which greatly improves the availability and stability of the model in the actual production environment.

[0111] (5) Assisted intelligent decision-making: By providing high-precision intensity prediction, this invention can provide data support for process engineers to optimize batching schemes and adjust thermal systems, thereby reducing fuel consumption, reducing the risk of ring formation in rotary kilns, and improving production efficiency while ensuring the quality of pellets. Attached Figure Description

[0112] Figure 1 This is a flowchart of the pellet intensity prediction method in Example 1 of the present invention during the model training phase;

[0113] Figure 2 This is a flowchart of the pellet intensity prediction method in the model testing stage of Example 1 of the present invention;

[0114] Figure 3 This is a schematic diagram of the network structure of the multilayer perceptron prediction algorithm in Application Embodiment 1 of the present invention;

[0115] Figure 4 This is a schematic diagram of the feature extraction module structure of the Long Short-Term Memory Network prediction algorithm in Application Embodiment 1 of the present invention. Detailed Implementation

[0116] The technical solution of the present invention will be illustrated below with examples. The scope of protection sought by the present invention includes, but is not limited to, the following embodiments.

[0117] Example 1

[0118] A method for predicting the intensity of pellet ore based on an optimization decision neural network, the method comprising the following steps:

[0119] S1. Data Acquisition Steps: Collect real-time PLC data and historical report data during the pellet production process.

[0120] S2. First prediction step: Input the real-time data from the PLC into the trained first prediction model to obtain the first intensity prediction value. The first prediction model is a model built based on a multilayer perceptron (MLP).

[0121] S3. Second prediction step: Input the time-series data of pellet intensity from the historical data of the report into the trained second prediction model to obtain the second intensity prediction value. The second prediction model is a model built based on a Long Short-Term Memory (LSTM) network.

[0122] S4. Optimization Decision Steps: Input the first intensity prediction value and the second intensity prediction value into the optimization decision model, dynamically adjust the weights of the first intensity prediction value and the second intensity prediction value, and output the final pellet intensity prediction result.

[0123] Example 2

[0124] A method for predicting the intensity of pellet ore based on an optimization decision neural network, the method comprising the following steps:

[0125] S1. Data Acquisition Steps: Collect real-time PLC data and historical report data during the pellet production process.

[0126] S2. First prediction step: Input the real-time data from the PLC and the historical data from the reports into the trained first prediction model to obtain the first intensity prediction value. The first prediction model is a model built based on a multilayer perceptron (MLP).

[0127] S3. Second prediction step: Input the time-series data of pellet intensity from the historical data of the report into the trained second prediction model to obtain the second intensity prediction value. The second prediction model is a model built based on a Long Short-Term Memory (LSTM) network.

[0128] S4. Optimization Decision Steps: Input the first intensity prediction value and the second intensity prediction value into the optimization decision model, dynamically adjust the weights of the first intensity prediction value and the second intensity prediction value, and output the final pellet intensity prediction result.

[0129] Example 3

[0130] A method for predicting the intensity of pellet ore based on an optimization decision neural network, the method comprising the following steps:

[0131] S1. Data Acquisition Steps: Collect real-time PLC data and historical report data during the pellet production process.

[0132] S2. First prediction step: Input the historical data from the report into the trained first prediction model to obtain the first intensity prediction value. The first prediction model is a model built based on a multilayer perceptron (MLP).

[0133] S3. Second prediction step: Input the time-series data of pellet intensity from the historical data of the report into the trained second prediction model to obtain the second intensity prediction value. The second prediction model is a model built based on a Long Short-Term Memory (LSTM) network.

[0134] S4. Optimization Decision Steps: Input the first intensity prediction value and the second intensity prediction value into the optimization decision model, dynamically adjust the weights of the first intensity prediction value and the second intensity prediction value, and output the final pellet intensity prediction result.

[0135] Example 4

[0136] Repeat Example 1, except that in step S4, the calculation formula for the optimization decision model is:

[0137]

[0138] In the formula: Q represents the final predicted strength of the pellet. MLP This is the first intensity prediction value. Q LSTM K is the weighting coefficient for the first intensity prediction. J is the weighting coefficient for the second intensity prediction. L is the variance weighting coefficient between the first and second intensity predictions. β is the optimization decision correction coefficient. Where K=0.5, J=0.5, L=0.00005, β=180.

[0139] Example 5

[0140] Repeat Example 4, except that step S4 also includes an exception handling strategy:

[0141] When the input parameters of the first prediction model are detected to exceed the preset threshold, or the first intensity prediction value exceeds the normal range of pellet intensity, K=0, L=0, J=1, β=0 are set.

[0142] When the second intensity prediction value is detected to be outside the normal range of pellet intensity, set J=0, L=0, K=1, β=0.

[0143] Example 6

[0144] Repeat Example 5, except that in step S1, the historical data in the report includes the sampling time and the test value of the pellet strength test.

[0145] Step S1 also includes a data matching sub-step: using the sampling time of the pellet strength test as the reference point T0, and based on the lag time of the pellet production process, calculating and matching the corresponding PLC real-time data at that time point. Wherein:

[0146] The raw material proportioning node time is T0-A. The green pellet feeding node time is T0-B. The chain grate process node time is T0-C. The rotary kiln process node time is T0-D. And the relationship between the lag times satisfies: A>B>C>D.

[0147] Example 7

[0148] Repeat Example 6, except that A is 200 min, B is 120 min, C is 110 min, and D is 80 min.

[0149] Example 8

[0150] Repeat Example 7, except that in step S1, the real-time data of the PLC includes the bentonite ratio in the mixture, the Fe3O4 content in the mixture, the Fe2O3 content in the mixture, the alkalinity of the mixture, the fixed carbon content of the mixture, the green pellet feed rate, the wind box temperature and fume hood temperature of the pH section of the chain grate, the chain grate speed, the temperature in the rotary kiln, the fuel quantity, the fuel consumption per unit, and the rotary kiln speed.

[0151] Example 9

[0152] Repeat Example 8, except that in step S2, the first prediction model includes an input layer, two hidden layers, and an output layer. The number of nodes in the input layer corresponds to the number of types of real-time PLC data input, and the number of nodes in the output layer is 1, used to output the first intensity prediction value.

[0153] Example 10

[0154] Repeat Example 9, except that in step S3, the second prediction model includes an input layer, a long short-term memory network hidden layer, and an output layer. The input layer is used to input a feature vector composed of pellet intensity values ​​at consecutive time nodes, and the output layer is used to output the second intensity prediction value at the target time node.

[0155] Example 11

[0156] Repeat Example 10, except that the method further includes a model training step:

[0157] Using historical PLC data and corresponding pellet strength test values ​​as training samples, the multilayer perceptron is trained to obtain the first prediction model.

[0158] The second prediction model is obtained by training a long short-term memory network using a continuous time series of historical pellet strength test values ​​as training samples.

[0159] Using the outputs of the first and second prediction models as inputs and the actual pellet strength test values ​​as targets, the weight coefficients in the optimization decision model are optimized using validation set data.

[0160] Example 12

[0161] This embodiment provides a pellet intensity prediction system based on an optimized decision neural network. This system is used to implement the method described in Embodiment 11, including:

[0162] Data acquisition module: This module is used to collect PLC process parameters in real time as PLC real-time data, and to collect historical test values ​​of pellet strength as historical data for reports. This module is also used to calculate and match the corresponding PLC real-time data based on the pellet strength testing sampling time and the preset lag time for each process segment, achieving time alignment of multi-source heterogeneous data.

[0163] First prediction module: It has a pre-trained first prediction model built in, which is used to receive real-time data from the PLC as input and output the first intensity prediction value.

[0164] In practical applications, real-time PLC data is preferred as input; when PLC data is missing or unreliable for any reason, process parameters recorded in historical reports can also be used as alternative input to ensure the fault tolerance of the system.

[0165] The second prediction module has a built-in trained second prediction model, which is used to receive a time-series feature vector composed of the pellet intensity values ​​of consecutive time nodes in the historical data of the report, and output the second intensity prediction value.

[0166] Optimization Decision Module: This module receives the first intensity prediction value and the second intensity prediction value, performs anomaly detection and dynamic weighted decision-making, and calculates the final pellet intensity prediction result.

[0167] Specifically, the module has a pre-set decision function and anomaly detection logic. When the input parameters of the first prediction module exceed a reasonable threshold (e.g., the temperature in the rotary kiln >1600℃ or <0℃), or when the first intensity prediction value of its output exceeds the normal intensity range (e.g., >5000N or <1000N), it is determined that the MLP prediction is abnormal, and the module will automatically switch to the prediction mode that relies solely on the LSTM model. When the second intensity prediction value exceeds the normal intensity range, it is determined that the LSTM prediction is abnormal, and the module will switch to the prediction mode that relies solely on the MLP model. Under normal circumstances, the module calculates the final pellet intensity prediction result by substituting the optimized decision parameters into formula (1).

[0168] The result output module is used to output the final pellet strength prediction results, providing a basis for decision-making on process parameter adjustments.

[0169] The modules described above can be deployed on industrial control computers, edge computing gateways, or cloud servers. In practical applications, the first and second prediction modules can run in parallel, and the optimization decision module can output prediction results at a high frequency to achieve real-time, accurate, and robust prediction of pellet intensity.

[0170] Example 13

[0171] This embodiment provides a computer-readable storage medium on which a computer program is stored. When executed by a processor, the program can implement the steps of the pellet intensity prediction method based on an optimized decision neural network described in any of the above embodiments. The storage medium can be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or it can be a virtual storage space in cloud storage.

[0172] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium.

[0173] Application Example 1

[0174] This application example provides a method for predicting the compressive strength of pellets based on an optimized decision neural network. This method integrates predictions from a multilayer perceptron (MLP) based on real-time process parameters and predictions from a long short-term memory (LSTM) network based on historical time-series data, and introduces an optimized decision mechanism to achieve accurate and stable prediction of the compressive strength of pellets. The method includes the following steps:

[0175] Step 1: Data Acquisition and Matching: Collect real-time PLC data and historical report data during the pellet production process.

[0176] Specifically, this step is completed jointly by the PLC data acquisition module and the report data acquisition module.

[0177] The PLC data acquisition module is responsible for collecting signals from field sensors, actuators, and other equipment in real time and uploading them to the PLC host or upper-level database. The collected parameters include node 1 data characterizing raw material properties: bentonite ratio (%), Fe3O4 content (%), Fe2O3 content (%), basicity (R), and fixed carbon content (%); node 2 data characterizing green pellet supply: green pellet feed rate (t / h); and node 3 and node 4 data characterizing the critical states of the chain grate-rotary kiln system: wind box temperature (°C) and hood temperature (°C) of the chain grate's pH section, chain grate speed (m / min), rotary kiln temperature (°C), fuel quantity (m³ / h or t / h), fuel consumption (m³ / t or kg / t), rotary kiln rotation speed (r / min), and the timestamps for each data acquisition.

[0178] The report data acquisition module is used to automatically collect and integrate the quality data of finished pellets from the Laboratory Information Management System (LIMS) or other databases. The core data includes the sampling time (T0) for cold pressing strength testing of pellets and the corresponding sample test value for cold pressing strength testing of pellets (unit: N / piece or daN / piece), which serves as the ground truth (label) for model training.

[0179] Considering the significant time lag between raw material entry into the system and finished product output, data time alignment is necessary. This application example employs a lag-time matching method based on the process mechanism. Using the pellet cold-pressing strength testing sampling time T0 as the baseline, the lag time of data for each process segment relative to T0 is pre-set based on the actual production line layout and process experience. Wherein:

[0180] The lag time for node 4 (rotary kiln data) is D = 80 minutes, that is, the rotary kiln data at time T0 - 80 minutes is taken;

[0181] The lag time C for node 3 (chain grate data) is 110 minutes, that is, the chain grate data at time T0 - 110 minutes is taken;

[0182] The lag time B for node 2 (green pellet feed data) is 120 minutes, which means taking the green pellet feed rate at time T0 - 120 minutes.

[0183] The lag time A for node 1 (raw material data) is 200 minutes, that is, the raw material ratio data at time T0 - 200 minutes is taken.

[0184] Based on the above rules, for each intensity test value (T0, Q) realA set of corresponding, time-aligned PLC process parameter vectors X = [x1, x2, ..., xn] is generated by matching these vectors. This constructs a standardized dataset for model training and prediction.

[0185] Step 2, Model Training: Train the MLP model, LSTM model, and optimization decision model using historical data. This step should be completed offline before system deployment and can be updated periodically or incrementally. The specific process is as follows: Figure 1 As shown.

[0186] (1) Training the multilayer perceptron prediction model (first prediction model):

[0187] The dataset matched in step one is used as training material. The PLC process parameter vector X is used as the input feature, and the corresponding pellet strength test value Q is used as the input feature. real As output labels, construct training sample pairs (X, Q) real ).

[0188] Build as Figure 3 The diagram shows a Multilayer Perceptron (MLP) network. In this application example, the number of nodes in the input layer is equal to the dimension n of the input features (where n=13). There are two hidden layers: the first hidden layer contains 4 neurons (a11, a12, a13, a14), and the second hidden layer contains 3 neurons (a21, a22, a23). All layers are fully connected, and the activation function can be either ReLU or Tanh. The output layer has one node, uses a linear activation function, and outputs the first intensity prediction value Q. MLP .

[0189] The MLP network is trained using a training sample set, with Adam as the optimizer and mean squared error (MSE) as the loss function. The network weights and biases are continuously adjusted through backpropagation until the model converges. The model parameters are then saved to obtain the optimal first prediction model.

[0190] (2) Training the Long Short-Term Memory Network Prediction Model (Second Prediction Model):

[0191] Historical time-series data on pellet intensity were extracted from historical report data to construct a dataset for LSTM training. Specifically, for any time node T', an input feature vector was constructed, which consisted of the intensity values ​​at times T'-2 and T'-1, i.e., [Q T’-2 Q T’-1 The corresponding label (i.e., the value to be predicted) is the intensity value Q at time T'. T’ .

[0192] Build an LSTM network. For example... Figure 4As shown, the input layer dimension is set to 2. Next, an LSTM hidden layer with 64 neurons is set up. The core of the LSTM unit lies in its gating structure:

[0193] Forget gate: The sigmoid function is used to determine the cell state C from the previous time step. t-1 Which information is discarded?

[0194] Input gate: Through the cooperation of the Sigmoid and Tanh functions, it determines which new information is stored in the current cell state C. t middle.

[0195] Output gate: based on the updated cell state C t The sigmoid function determines which information to output, thus obtaining the hidden layer output h. t .

[0196] The output layer dimension is set to 1, connected after the LSTM layer, and uses a linear activation function to output the second intensity prediction value Q. LSTM (i.e., the predicted Q) T’ ).

[0197] The LSTM network was trained using a pre-constructed time-series dataset. The optimizer was Adam, with an initial learning rate of 0.001; the loss function was MSE; the training strategy included a batch size of 32, 100 epochs, and early stopping, where training was stopped if the loss function value on the validation set did not decrease for several consecutive epochs (e.g., 10) to prevent overfitting and preserve the optimal LSTM prediction model.

[0198] (3) Training and optimizing the decision-making model:

[0199] Prepare an independent validation set. For each sample in the validation set, input its real-time PLC data into the first prediction model to obtain the first intensity prediction value Q. MLP The historical intensity time series data is input into the second prediction model to obtain the second intensity prediction value Q. LSTM At this point, the true intensity value Q of the sample is known. real .

[0200] The core of the optimized decision-making model is the following decision function:

[0201]

[0202] Where: K is the weight coefficient of the first intensity prediction value; J is the weight coefficient of the second intensity prediction value; L is the variance weight coefficient of the first intensity prediction value and the second intensity prediction value; β is the optimization decision correction coefficient; that is, K, J, L and β are the decision parameters to be optimized, and the constraints are K+J=1, 0≤K≤1, 0≤J≤1, -1≤L≤1, β∈[-500, 500], preferably β∈[-300, 300].

[0203] Optimization algorithms such as grid search or genetic algorithms are used to optimize K, J, L, and β with the objective of minimizing the prediction error on the validation set (e.g., minimizing the root mean square error RMSE), thus determining a set of optimal decision parameters (K). opt J opt L opt ,β opt This allows us to construct the final optimized decision-making model.

[0204] Step 3: Online Prediction: Utilize the trained model to perform online predictions on the new data. This step is the core application scenario of this invention, and the specific process is as follows: Figure 2 As shown.

[0205] S1. Data Acquisition: During the online application phase, the PLC process parameters at the current moment are collected in real time, and the latest historical intensity report data is obtained from the report database.

[0206] S2, First Prediction: The collected current PLC process parameters (after necessary data cleaning and normalization) are input into the trained first prediction model. After forward propagation calculation, the model outputs the first intensity prediction value Q for the current batch. MLP It should be noted that in some alternative implementations, when real-time PLC data is missing for any reason, the process parameters recorded in the historical data of the report can also be used as input to obtain effective prediction results.

[0207] S3. Second Prediction: Determine the target time node T to be predicted. Extract the intensity values ​​at times T-2 and T-1 from the historical intensity time series database to construct the feature vector [Q]. T-2 Q T-1 The vector is input into the trained second prediction model. The model performs temporal inference through its gating structure and outputs the second intensity prediction value Q for the current batch. LSTM .

[0208] S4. Optimization Decision-Making and Anomaly Handling: Q... MLP and Q LSTM The input is fed into the trained optimization decision model. Anomaly detection logic is executed before the final calculation:

[0209] Scenario 1 (Data Anomaly): If any key parameter in the input PLC process parameters exceeds a reasonable physical threshold (e.g., the temperature in the rotary kiln exceeds 1600°C, or is 0), or the calculated Q... MLP If the value significantly exceeds the normal range of pellet strength (e.g., greater than 5000N or less than 1000N), the MLP model input or output is deemed abnormal. In this case, an emergency strategy is activated, forcibly setting K=0, L=0, J=1, and β=0. Substituting these values ​​into the decision function, Q = Q0 LSTM That is, to completely trust the time series prediction results of LSTM.

[0210] Scenario 2 (Time Series Prediction Anomaly): If Q is detected LSTM If the value exceeds the normal range, the LSTM model prediction is deemed abnormal. In this case, an emergency strategy is activated, forcibly setting J=0, L=0, K=1, and β=0. Substituting these values ​​into the decision function, Q = Q0. MLP This means that the process parameter prediction results of the MLP are completely trusted.

[0211] Scenario 3 (Normal Case): If no anomaly is detected, the optimal decision parameter (K) obtained in step 2 (3) is used. opt J opt L opt ,β opt Substitute into formula (1) to calculate:

[0212]

[0213] Output results: The Q calculated by the above formula will be output as the final pellet strength prediction result of this method, for reference by production operators or for use in subsequent closed-loop control systems.

[0214] Specifically, to further illustrate the beneficial effects of the present invention, this application embodiment provides comparative results based on historical data from a steel company's chain grate-rotary kiln production line.

[0215] Data collection and segmentation:

[0216] The experiment collected 1000 sets of valid data from the production line. Each set of data was recorded in chronological order of production time. The first 800 sets of data were selected as the base dataset (used for model training and parameter optimization), and the last 200 sets of data were used as the test set (used for final prediction performance evaluation). The test set was completely isolated during the model training and parameter optimization process and did not participate in any learning or selection steps.

[0217] Validation set construction and decision parameter optimization:

[0218] From the first 800 sets of basic data, a random sampling method was used to divide the dataset into 600 sets as a training set and 200 sets as an independent validation set. This validation set is only used to compare the relative effects of different decision parameters and does not involve evaluating time generalization ability.

[0219] The genetic algorithm parameters are set as follows: population size is 50, maximum number of iterations is 100, crossover probability is 0.8, mutation probability is 0.1, and the fitness function is the root mean square error (RMSE) on the validation set. The genetic algorithm is used to optimize the decision parameters K, J, L, and β, with the objective of minimizing the validation set RMSE, resulting in a set of optimal decision parameters: K opt =0.6, J opt =0.4, L opt =0.0002, β opt =50.

[0220] It should be noted that the above-mentioned genetic algorithm parameters are a specific implementation of this application embodiment. Those skilled in the art can adaptively adjust parameters such as population size, number of iterations, crossover probability, and mutation probability according to the complexity of the actual problem and computing resources, which does not depart from the protection scope of this invention.

[0221] Comparative experimental setup:

[0222] The prediction results of the method of this invention are obtained by using this set of parameters on the test set. The method of this invention is compared with a single MLP model and a single LSTM model. To ensure the fairness of the comparison, all comparison models use the exact same data partitioning (the same sub-training set, validation set, and test set) and are trained using the default hyperparameter configurations under their respective frameworks, without using the decision parameters optimized by this invention (these parameters are unique to this invention). The evaluation metrics are: root mean square error (RMSE), mean absolute percentage error (MAPE), and hit rate.

[0223] The hit rate is defined as the proportion of samples in the test set whose absolute difference between the model's predicted value and the actual value is less than or equal to a preset tolerance threshold. In this application embodiment, the preset tolerance threshold is set to 50N, i.e., |predicted value - actual value| ≤ 50N, which is considered a hit. It should be noted that the preset tolerance threshold can be set according to the actual application scenario and accuracy requirements, such as, but not limited to, 50N, 100N, etc.

[0224] The results are shown in the table below:

[0225]

[0226] It should be noted that the above decision parameters (K=0.6, J=0.4, L=0.0002, β=50) are merely a set of example values ​​obtained based on the validation set data of this application embodiment (production is continuous and needs to be re-optimized after a certain period of time). In practical applications, those skilled in the art can obtain the appropriate parameters using the same optimization method based on their own production data. As long as the parameters meet the range defined by the corresponding constraints, they all fall within the protection scope of this invention and can achieve the technical effects of this invention.

[0227] The experimental data above show that the optimized decision fusion method proposed in this invention has a significantly higher prediction accuracy than a single MLP or LSTM model, demonstrating the superiority of this method in improving the accuracy of pellet intensity prediction.

[0228] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. For those skilled in the art, several improvements and modifications can be made without departing from the spirit and principles of this application, such as replacing the MLP model with other regression models capable of handling nonlinear relationships (such as support vector regression SVR), or replacing the variance term in the optimization decision model with the standard deviation term. These improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A method for predicting the intensity of pellet ore based on an optimization decision neural network, characterized in that: The method includes the following steps: S1. Data Acquisition Steps: Collect real-time PLC data and historical report data during the pellet production process; S2. First prediction step: Input the real-time data from the PLC and / or historical data from the reports into the trained first prediction model to obtain the first intensity prediction value; the first prediction model is a model built based on a multilayer perceptron (MLP); S3. Second prediction step: Input the time series data of pellet intensity from the historical data of the report into the trained second prediction model to obtain the second intensity prediction value; the second prediction model is a model built based on Long Short-Term Memory Network (LSTM); S4. Optimization Decision Steps: Input the first intensity prediction value and the second intensity prediction value into the optimization decision model, dynamically adjust the weights of the first intensity prediction value and the second intensity prediction value, and output the final pellet intensity prediction result.

2. The method according to claim 1, characterized in that: In step S4, the calculation formula for the optimization decision model is as follows: In the formula: Q represents the final predicted strength of the pellet; Q MLP Q is the first intensity prediction value; LSTM K is the weighting coefficient for the first intensity prediction value; J is the weighting coefficient for the second intensity prediction value; L is the weighting coefficient for the variance between the first and second intensity prediction values. β is the optimization decision correction coefficient; where K+J=1, and 0≤K≤1, 0≤J≤1; -1≤L≤1; β∈[-500, 500], preferably β∈[-300, 300].

3. The method according to claim 2, characterized in that: Step S4 also includes exception handling strategies: When the input parameters of the first prediction model are detected to exceed the preset threshold, or the first intensity prediction value exceeds the normal range of pellet intensity, K=0, L=0, J=1, β=0 are set. When the second intensity prediction value is detected to be outside the normal range of pellet intensity, set J=0, L=0, K=1, β=0.

4. The method according to any one of claims 1-3, characterized in that: In step S1, the historical data in the report includes the sampling time and the test value of the pellet strength test. Preferably, step S1 also includes a data matching sub-step: taking the sampling time of pellet strength testing as the reference point T0, and based on the lag time of the pellet production process, calculating and matching the corresponding PLC real-time data at that time point; wherein: The raw material proportioning node time is T0-A; the green pellet feeding node time is T0-B; the chain grate process node time is T0-C; the rotary kiln process node time is T0-D; and the relationship between the lag times satisfies: A>B>C>D.

5. The method according to claim 4, characterized in that: The value range for A is 180~220min; the value range for B is 110~130min; the value range for C is 100~120min; and the value range for D is 70~90min.

6. The method according to any one of claims 1-5, characterized in that: In step S1, the real-time data from the PLC includes one or more of the following: bentonite ratio in the mixture, Fe3O4 content in the mixture, Fe2O3 content in the mixture, alkalinity of the mixture, fixed carbon content in the mixture, green pellet feed rate, wind box temperature and hood temperature of the chain grate's pH section, chain grate speed, rotary kiln temperature, fuel quantity, fuel consumption per unit, and rotary kiln rotation speed.

7. The method according to any one of claims 1-6, characterized in that: In step S2, the first prediction model includes an input layer, at least one hidden layer, and an output layer; the number of nodes in the input layer corresponds to the number of types of real-time PLC data input, and the number of nodes in the output layer is 1, used to output the first intensity prediction value; and / or In step S3, the second prediction model includes an input layer, a long short-term memory network hidden layer, and an output layer; the input layer is used to input a feature vector composed of the pellet intensity values ​​at consecutive time nodes, and the output layer is used to output the second intensity prediction value at the target time node.

8. The method according to any one of claims 1-7, characterized in that: The method also includes a model training step: Using historical PLC data and corresponding pellet strength test values ​​as training samples, the multilayer perceptron is trained to obtain the first prediction model; The second prediction model is obtained by training a long short-term memory network using a continuous time series of historical pellet strength test values ​​as training samples. Using the outputs of the first and second prediction models as inputs and the actual pellet strength test values ​​as targets, the weight coefficients in the optimization decision model are optimized using validation set data.

9. A pellet intensity prediction system based on an optimization decision neural network or a pellet intensity prediction system for use in any one of claims 1-8, characterized in that: The system includes: Data acquisition module: used to collect PLC process parameters in real time as PLC real-time data, and to collect historical test values ​​of pellet strength as historical data for reports; First prediction module: It has a pre-trained first prediction model built in, which is used to receive real-time data from the PLC and / or historical data from reports as input, and output a first intensity prediction value; The second prediction module has a built-in trained second prediction model, which is used to receive a time-series feature vector composed of the pellet intensity values ​​of consecutive time nodes in the historical data of the report, and output the second intensity prediction value. Optimization Decision Module: It has a built-in optimization decision model, which is used to receive the first intensity prediction value and the second intensity prediction value, and perform anomaly detection and dynamic weighted decision to calculate the final pellet intensity prediction result. Result output module: used to output the final pellet strength prediction result.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.

Images