Soft measurement method and system for compressive strength of pellets

Abstract

The invention discloses a soft measurement method and system for compressive strength of pellets. The method comprises the following steps: extracting characteristics of time sequence data in a production process through a gating circulation unit; constructing a space-time diagram neural network containing an adaptive learning branch and a physical prior branch, and extracting data association features and physical constraint features conforming to a process mechanism in parallel; and the intensity is predicted through a regression layer after the double-branch features are fused. The model training adopts a composite loss function containing a prediction error and an entropy regularization item, takes an off-line test value as a supervision label, and forces the network structure to be sparse while optimizing the prediction precision. And inputting real-time data into the trained model to output a compressive strength prediction value on line. According to the method, the problems of manual detection lag and poor physical interpretability of the model are solved, and effective online prediction of the compressive strength of the pellets is realized. The problems that manual detection lags behind and a traditional model is poor in physical interpretability are effectively solved, and therefore the effective online prediction degree of the compressive strength of the pellets can be achieved.

Description

Technical Field

[0001] This invention relates to the field of soft measurement of pellet quality indicators, and more particularly to a soft measurement method and system for pellet compressive strength based on a physical prior graph network. Background Technology

[0002] In the iron and steel metallurgical industry, iron ore pellets are a crucial raw material for blast furnace ironmaking, and their quality directly impacts the efficiency and cost of high-speed rail production. Belt roasters are the most important production equipment for iron ore pellets. The quality stability of the produced pellets plays a key role in stable and smooth ironmaking operations, as well as energy conservation and carbon reduction. Among these factors, the compressive strength of the pellets is a critical indicator of product quality, directly affecting the permeability of the blast furnace and smelting efficiency.

[0003] Currently, the testing of pellet compressive strength mainly relies on manual, timed sampling and offline laboratory analysis. This method has significant drawbacks, such as low sampling frequency and strong lag (results typically take 2-4 hours to obtain), failing to reflect product quality fluctuations under current operating conditions and hindering guidance for real-time production process control. By the time quality defects are detected, a large amount of waste has often already been produced, resulting in severe resource waste and economic losses.

[0004] To overcome the lag issues of manual inspection, soft measurement technology has been introduced into industrial processes. Existing soft measurement methods mostly employ traditional machine learning (such as support vector regression and random forests) or standard deep learning (such as multilayer perceptrons). However, these methods have significant limitations in the complex industrial process of pellet roasting.

[0005] 1. Ignoring the spatiotemporal coupling between sensors: Traditional methods typically treat all process parameters (such as temperature, pressure, flow rate, etc.) collected by all sensors as flat, independent input features. However, in reality, the pellet roasting process involves multiple tightly coupled stages such as drying, preheating, roasting, and cooling, involving hundreds of sensors. There are complex spatiotemporal correlations and causal relationships among the data from these sensors.

[0006] 2. Lack of guidance from physical mechanisms: The pellet roasting process has clear physical laws, such as material flow and gas regeneration. Existing data-driven models fail to effectively utilize these known prior knowledge of process mechanisms, resulting in insufficient generalization ability of the models under complex operating conditions, a lack of physical interpretability in the prediction results, and difficulty in gaining the trust of field operators.

[0007] 3. Poor model interpretability: Traditional black box models cannot reveal the intrinsic relationship between process variables and compressive strength. When predictions are biased, they are difficult to provide effective guidance for process adjustments.

[0008] Therefore, there is an urgent need for a soft measurement method that can integrate data-driven and physical mechanisms to predict the compressive strength of pellets in real time and accurately, and has good interpretability, in order to solve the problems of lag in manual detection, poor physical interpretability of traditional models, and difficulty in capturing the coupling relationship between multiple variables under complex working conditions. Summary of the Invention

[0009] This invention provides a soft measurement method and system for pellet compressive strength based on physical prior graph networks. The method aims to solve the technical problems in the prior art, such as the long detection cycle of pellets, the lag in laboratory tests, and the difficulty of traditional data-driven models in accurately capturing the coupling relationship between temperature fluctuations and compressive strength under complex environments, which leads to the inability to effectively predict compressive strength in real time.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] In a first aspect, the present invention provides a soft measurement method for the compressive strength of pellets, comprising: S1: Obtain multiple sets of historical production data during the production process of the belt roaster. Each set of production data includes time-series data of multiple process parameters and their corresponding offline test values ​​of pellet compressive strength. S2: Preprocess the time series data and use the gated loop unit to extract the time series features of each process parameter node to obtain the node time series feature matrix; S3: Construct a dual-branch parallel physical perception spatiotemporal graph neural network model, which includes an adaptive learning branch and a physical prior branch. Input the node temporal feature matrix into both the adaptive learning branch and the physical prior branch to extract the adaptive features and physical prior features respectively. S4: Fuse the adaptive node features with the physical prior features to obtain the fused feature vector; S5: Input the fused feature vector into the regression prediction layer to obtain the predicted value of the pellet compressive strength; S6: Using multiple sets of production data from S1 as input and the corresponding offline test values ​​as supervision labels, train the physical perception spatiotemporal graph neural network model so that the predicted values ​​output by the model approximate the offline test values; wherein, during the model training process, a composite loss function including prediction error and entropy regularization is introduced as a sparsity constraint term to optimize the graph structure generated by the adaptive learning branch. S7: Input the real-time collected production process parameters into the trained model and output the real-time predicted value of the pellet compressive strength.

[0012] Furthermore, in S2, the specific process of extracting the temporal features of each process parameter node using a gated recurrent unit (GRU) to obtain the node temporal feature matrix is ​​as follows: ; ; in, Indicates the first element in the input data. One sensor in Normalized measurement value at time; Indicates the first One sensor in The hidden state vector at time t; Indicates the first The sensor at the previous moment The hidden state vector at time t; Indicates the first A sensor passed The final output node temporal feature vector after processing at each time step has a dimension of . .

[0013] Furthermore, S2 also includes, for each sensor node Randomly initialize a learnable static node embedding vector. This leads to the static embedding vector matrix. : ; in, The dimension in which the node is embedded.

[0014] Furthermore, in S3, the specific process of extracting adaptive features is as follows: S311: Based on node embedding vector matrix Calculate the similarity between nodes: ; in, Representing nodes respectively and nodes Learnable embedding vectors; Represents a node and nodes The original similarity scores between them; S312: Introducing a dual sparsity mechanism to generate an adaptive sparse adjacency matrix: ; ; in, Represents the linear rectification activation function; Represents a node With nodes The original similarity between them; Represents a node With nodes The non-negative intermediate similarity score between the two sides after hard truncation and temperature scaling; Indicates the temperature coefficient; This represents the nodes in the generated adaptive sparse adjacency matrix. For nodes Normalized connection weights; S313: Based on the generated adaptive sparse adjacency matrix Using graph convolutional networks to analyze the temporal features of nodes Perform spatial aggregation and extract adaptive features : ; in, This is the node temporal feature matrix output by S2, with a matrix dimension of... ; The learnable weight matrix has a dimension of . ; It is a non-linear activation function; This represents the final output adaptive node feature matrix.

[0015] Furthermore, in S3, the specific process of extracting physical prior features is as follows: S321: Based on the physical process flow of pellet roasting, divide the N sensor nodes into K process regions; S322: Computation Node Original importance score within its respective process area: ; in, For nodes Adaptive features; These are the learnable weight matrix and bias term of the attention mechanism, respectively. Assign an attention scoring vector; For nodes The original importance score; S323: Softmax normalize the original importance scores of all nodes within the process region to obtain normalized attention weights: ; in, For nodes Normalized attention weights within its respective process region; For the first One process area; S324: Weighted aggregation yields the region feature vector: ; in, For the aggregated first Prior feature vectors of each process region; S325: Constructing a physical mask matrix based on the physical laws of the pellet roasting process. ; S326: Update region features using a graph attention network combined with a physical mask matrix: ; ; ; ; in, The linear transformation weight matrix of the GAT layer; This represents the parameter vector for the attention mechanism. This is a vector concatenation operation; Process area based on data calculation For process area The original attention coefficient; For the process area after injecting physical mask For process area Corrected attention coefficient; For process area arrive The mask matrix; These are the final attention weights after Softmax normalization; These are the final attention weights after Softmax normalization; For the updated number The prior physical characteristics of each region; It is a non-linear activation function.

[0016] Furthermore, the constructed physical mask matrix satisfies: if the process region arrive If there is a physical flow of materials or air, then ,otherwise .

[0017] Furthermore, in step S6, a composite loss function including prediction error and entropy regularization is included. Specifically: ; ; ; in This is the mean square error loss; This refers to the number of sample batches. For the first The true compressive strength test values ​​of each batch of samples; This represents the predicted compressive strength of the pellets output by the model. The regularization coefficient is used. This represents the nodes in the generated adaptive sparse adjacency matrix. For nodes Normalized connection weights; This represents the number of sensor nodes; This is the information entropy regularization term; It is a constant.

[0018] Furthermore, the process parameters include at least one of the following: wind box temperature, wind box pressure, burner temperature, burner flow rate, fan speed, machine speed, and material layer thickness in the drying section, preheating section, calcination section, and cooling section.

[0019] Secondly, a soft measurement system for pellet compressive strength, the system being used to perform the method described above, comprising: The historical production data acquisition module is used to acquire multiple sets of historical production data during the production process of the belt roaster. Each set of production data includes time-series data of multiple process parameters and their corresponding offline test values ​​of pellet compressive strength. The time series feature extraction module is used to preprocess the time series data and use the gated loop unit (GRU) to extract the time series features of each process parameter node to obtain the node time series feature matrix. The model building module constructs a dual-branch parallel physical perception spatiotemporal graph neural network model, which includes an adaptive learning branch and a physical prior branch. The node temporal feature matrix is ​​simultaneously input into the adaptive learning branch and the physical prior branch to extract adaptive features and physical prior features respectively. The feature fusion and prediction module is used to fuse adaptive node features with physical prior features to obtain a fused feature vector; the fused feature vector is then input into the regression prediction layer to obtain the predicted value of the pellet compressive strength. The model training module is used to train the physical perception spatiotemporal graph neural network model with multiple sets of historical production data as input and corresponding offline test values ​​as supervision labels, so that the predicted values ​​output by the model approximate the offline test values. In the model training process, a composite loss function including prediction error and entropy regularization is introduced as a sparsity constraint term to optimize the graph structure generated by the adaptive learning branch. The online prediction module is used to collect production process parameters in real time, input them into the trained model, and output real-time predicted values ​​of pellet compressive strength.

[0020] Furthermore, the system also includes a visualization module for displaying the compressive strength prediction curve in real time, and visually displaying the process variable correlation topology diagram generated by the adaptive learning branch and the attention weight heatmap of each process segment.

[0021] The present invention provides a soft measurement method and system for the compressive strength of pellets, which has the following advantages compared with the prior art:

[0022] 1. Real-time online prediction of pellet compressive strength was achieved: Minute-level prediction of compressive strength was achieved through a data-driven model, completely overcoming the 2-4 hour lag problem of manual detection, and timely reflecting changes in production conditions, providing the possibility for real-time production control.

[0023] 2. Significantly enhanced physical interpretability of the model: By designing physical prior branches and using a physical mask matrix to force information to be transmitted only along paths that conform to the process mechanism (e.g., materials can only flow from the drying section to the preheating section), the model's learning process follows the physical laws of pellet roasting. This design gives the model's decision-making process clear physical meaning, greatly improving the model's credibility and acceptability.

[0024] 3. Automatically uncovering sparse causal relationships and reducing model complexity: By combining an adaptive graph learner with a dual sparsification mechanism of ReLU truncation and entropy regularization, sparse and critical causal relationship topologies can be automatically learned and filtered from high-dimensional, strongly coupled sensor data. This not only reduces model complexity and the risk of overfitting but also reveals key process variables affecting compressive strength and their correlations, providing direct guidance for process optimization.

[0025] 4. Dual-branch parallel architecture enables deep integration of data and mechanisms: A parallel dual-branch design is adopted, with the adaptive learning branch extracting statistical patterns from the data, and the physical prior branch introducing domain knowledge as constraints, finally performing feature-level fusion. This architecture maintains the flexibility of the data-driven model while incorporating the reliability of physical mechanisms, enabling the model to remain robust and accurate under complex and variable conditions. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the soft measurement method for pellet compressive strength provided in an embodiment of the present invention; Figure 2 This is a flowchart of the soft measurement method for pellet compressive strength provided in an embodiment of the present invention; Figure 3This is a graph showing the model prediction results of the soft measurement method for compressive strength of pellets provided in this embodiment of the invention compared with random forest, multilayer perceptron, and support vector machine models; Figure 4 This is a weighted heatmap of each process segment provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0029] Example 1

[0030] like Figure 1-2 As shown, this embodiment provides a soft measurement method for the compressive strength of pellets based on a physical prior graph network, including: S1: Data Acquisition and Preprocessing: Acquire multiple sets of historical production data from the belt calciner during the production process. Each set of production data includes time-series data of multiple process parameters and their corresponding offline test values ​​for pellet compressive strength. The specific process is as follows: Specifically, the distributed control system (DCS) of the belt calciner collects time-series data of process parameters for multiple batches during continuous production periods. These process parameters cover key process variables in the drying, preheating, calcining, and cooling sections, including but not limited to the temperature of the air box, air box pressure, burner temperature, burner gas flow rate, fan speed, trolley speed (machine speed), and material layer thickness in each section. At the same time, the offline test values ​​of the pellet compressive strength recorded by production batch for the corresponding time period are extracted.

[0031] S2: Preprocess the time series data and use the gated loop unit (GRU) to extract the time series features of each process parameter node to obtain the node time series feature matrix.

[0032] Specifically, time-series data preprocessing includes data cleaning, anomaly handling, and data normalization: Detect and process missing values ​​in the process data: For short-term missing values, use linear interpolation or the average of the preceding and following time points to fill in the missing values; for long-term missing data, mark the batch of samples as invalid and remove them.

[0033] Identify and remove outliers: Set reasonable physical upper and lower limits for each parameter based on process knowledge, and filter out obviously abnormal data that exceed the range; at the same time, statistical methods (such as those based on standard deviation) can be used to detect and process outliers.

[0034] Significantly abnormal records in the test data (such as those exceeding the normal intensity range or recording errors) are checked and removed to ensure the reliability of the label data.

[0035] Data normalization: The time series data of process parameters after cleaning are standardized to eliminate the impact of differences in the units and numerical ranges of different sensors on model training.

[0036] A three-dimensional tensor structure is constructed from the normalized process parameter time series data. ,in, This refers to the number of sample batches. The time step is determined based on the process window and is not limited thereto; in this embodiment, it is specifically implemented as 60 minutes. This represents the number of sensor nodes. A vector is constructed from the corresponding standardized pellet compressive strength test values. As supervisory labels, the entire dataset is then divided into training, validation, and test sets for subsequent model training, hyperparameter tuning, and independent performance evaluation.

[0037] The input data is encoded using a gated recurrent unit with shared weights. For the first... Time series of sensors The calculation process is as follows: ; ; in, Indicates the first One sensor in Normalized measurement value at time; Indicates the first One sensor in The hidden state vector at time t; Indicates the first The sensor at the previous moment The hidden state vector at time t; Indicates the first A sensor passed The final output node temporal feature vector after processing at each time step has a dimension of . .

[0038] In addition to temporal characteristics, for each sensor node Randomly initialize a learnable static node embedding vector. This leads to the static embedding vector matrix. : ; in, This represents the dimension of node embedding. Static embedding vectors are independent of temporal data and are used in the adaptive learning branch to capture the inherent semantic or topological relationships between nodes, complementing dynamic temporal features.

[0039] S3: Construct a dual-branch parallel physical perception spatiotemporal graph neural network model, which includes an adaptive learning branch and a physical prior branch. Input the node temporal feature matrix into both the adaptive learning branch and the physical prior branch simultaneously to extract the adaptive features and physical prior features respectively.

[0040] Specifically, the process of extracting adaptive features is as follows: S311: Based on node embedding vector matrix Calculate the similarity between nodes: ; in, Representing nodes respectively and nodes Learnable embedding vectors; Represents a node and nodes The original similarity scores between them; S312: Introducing a dual sparsity mechanism to generate an adaptive sparse adjacency matrix: Applying the ReLU function to the similarity scores forces negative values ​​to zero, achieving the first level of sparsity; inputting the truncated similarity scores into the Softmax function for normalization; and controlling the sharpness of the distribution through a temperature coefficient to generate the final adaptive sparse adjacency matrix, achieving the second level of sparsity. ; ; in, This represents the linear rectification activation function, and its calculation formula is: This is used to force negative similarity to 0, perform hard truncation, remove noisy connections, ensure the physical connectivity of the graph structure, and achieve the first layer of sparsity. Represents a node With nodes The original similarity between them; Represents a node With nodes The non-negative intermediate similarity score between the two sides after hard truncation and temperature scaling; It represents the temperature coefficient, used to adjust the smoothness of the distribution; This represents the nodes in the generated adaptive sparse adjacency matrix. For nodes Normalized connection weights.

[0041] S313: Based on the generated adaptive sparse adjacency matrix Using graph convolutional networks to analyze the temporal features of nodes Perform spatial aggregation and extract adaptive features : ; in, This is the node temporal feature matrix output by S2, with a matrix dimension of... ; The learnable weight matrix has a dimension of . ; It is a non-linear activation function; This represents the final output adaptive node feature matrix, which contains comprehensive information in both spatiotemporal dimensions.

[0042] Specifically, the process of extracting physical prior features is as follows: S321: Based on the physical process flow of pellet roasting, divide the N sensor nodes into K process areas (e.g., blast zone, exhaust zone, preheating zone, roasting zone, first cooling zone, second cooling zone, global zone). S322: Computation Node Original importance score within its respective process area: ; in, For nodes Adaptive features; For the learnable weight matrix and bias terms of the attention mechanism; Assign an attention scoring vector; For nodes The original importance score; S323: Softmax normalize the original importance scores of all nodes within the process region to obtain normalized attention weights: ; in, For nodes Normalized attention weights within its respective process region; S324: Weighted aggregation yields the region feature vector: ; in, For the aggregated first Prior feature vectors of each process region; For the first One process area; S325: Constructing a physical mask matrix based on the physical laws of the pellet roasting process. ; where the physical mask matrix satisfies: if the process region arrive If there is a physical flow of materials or air, then ,otherwise ; In practice, the material flow direction is constrained: the material can only flow from the upstream area to the downstream area (blowing drying → exhaust drying → preheating → roasting → first cooling → second cooling).

[0043] Gas reheat constraint: Hot gas from the cooling section can be partially recirculated to the drying section for preheating;

[0044] Example: There is a material flow from the roasting zone (designated as zone 4) to a cooling zone (designated as zone 5), therefore... There is no direct material flow from the cooling zone to the roasting zone, therefore .

[0045] S326: Update region features using a graph attention network combined with a physical mask matrix: ; ; ; ; in, The linear transformation weight matrix of the GAT layer; This represents the parameter vector for the attention mechanism. This is a vector concatenation operation; Process area based on data calculation For process area The original attention coefficient; The corrected attention coefficient after injecting the physical mask; For process area arrive The mask matrix. Because The coefficients of non-physical connections will be forced to become negative infinity; These are the final attention weights after Softmax normalization; the weights of non-physical connections become 0 here. For the updated number The prior physical characteristics of each region; It is a non-linear activation function.

[0046] The adaptive branch mines potential correlations between sensors using an adaptive graph learner and removes negative correlation noise using a ReLU truncation mechanism; the physical prior branch uses a physical mask to constrain the graph attention network, allowing information to be transmitted only between process regions where there is a physical material flow or gas regeneration relationship, thus constructing a feature evolution model that conforms to the industrial mechanism of belt roasters.

[0047] S4: Fuse the adaptive node features with the physical prior features to obtain the fused feature vector.

[0048] Specifically, adaptive features With prior physical characteristics splicing after dimensional transformation: ; ; ; in, , These are the adaptive features after dimensional transformation and the prior physical features, respectively. This is a feature of fusion.

[0049] S5: Input the fused feature vector into the regression prediction layer to obtain the predicted value of the pellet compressive strength; Specifically, the fused feature vector is then input into the fully connected layer, which outputs the final predicted value of the pellet compressive strength. : ; in, All of these are weights of the multilayer perceptron; These are all multilayer perceptron bias parameters; This is the predicted compressive strength of the pellets output by the model.

[0050] S6: Using multiple sets of production data from S1 as input and the corresponding offline test values ​​as supervision labels, train the physical perception spatiotemporal graph neural network model so that the predicted values ​​output by the model approximate the offline test values; in the model training process, a composite loss function including prediction error and entropy regularization is introduced as a sparsity constraint term to optimize the graph structure generated by the adaptive learning branch.

[0051] Specifically, it includes a composite loss function that incorporates prediction error and entropy regularization. Specifically: ; ; ; in Mean squared error loss is used to measure prediction accuracy. This refers to the number of sample batches. For the first The true compressive strength test values ​​of each batch of samples; This represents the predicted compressive strength of the pellets output by the model. is the regularization coefficient, used to balance prediction accuracy and graph sparsity; This represents the number of sensor nodes; It is a constant, a small constant used to ensure numerical stability; As an information entropy regularization term, minimizing this term forces The probability distribution in the model tends towards 0 or 1 (i.e., deterministic connections), thereby reducing fuzzy and redundant connections. By optimizing this loss function, the model can automatically filter out the connections between key process parameter variables while maintaining prediction accuracy.

[0052] Entropy regularization constraints are introduced for offline model training and parameter optimization. Preprocessed features are input into the network to construct a composite loss function consisting of the mean squared error of prediction (MSE) and an information entropy regularization term. Backpropagation training is performed using the Adam optimizer, which forces the adaptive graph structure distribution to become sparser and sharper by minimizing the entropy loss, automatically selecting key causal paths. At the same time, an early stopping mechanism is introduced to save the optimal model parameters when the loss on the validation set no longer decreases.

[0053] S7: Input the real-time collected production process parameters into the trained model and output the real-time predicted value of the pellet compressive strength.

[0054] Specifically, in the prediction phase, the model receives minute-level process variables from the production site in real time through a data acquisition interface, and performs forward inference calculations via a trained graph network model. The model integrates adaptive features and physical prior features through a multi-scale fusion module, thereby outputting high-precision predicted values ​​of pellet compressive strength. Furthermore, it performs a visual analysis of the attention weights of each process section in the trained model, providing real-time guidance for on-site operators to adjust thermal operating procedures.

[0055] More specifically, the process parameters include at least one of the following: air box temperature, air box pressure, burner temperature, burner flow rate, fan speed, machine speed, and material layer thickness in the drying section, preheating section, calcination section, and cooling section.

[0056] Through the above steps, the soft measurement method for pellet compressive strength based on physical prior graph network can accurately predict the compressive strength of pellets based on the input production data, providing support for quality optimization and control in the production process.

[0057] To further verify the technical solution of this application, the proposed method was combined with random forest, multilayer perceptron, and support vector machine models to predict the compressive strength of pellets. The model prediction results are as follows: Figure 3 As shown, the error R² of the proposed method compared to the true value is 0.8082, significantly higher than traditional methods such as Support Vector Regression (R²=0.5608), Random Forest (R²=0.4532), and Multilayer Perceptron (R²=0.5805). The proposed method's prediction curve shows the highest degree of fit with the true value curve, maintaining a high degree of consistency with the overall trend. The figure shows that the resilience of the real sample data fluctuates frequently and dramatically, and the proposed method (blue dashed line) accurately captures these fluctuations, demonstrating strong predictive ability in both local peaks and troughs. In contrast, the prediction curves of Support Vector Regression, Random Forest, and Multilayer Perceptron appear relatively flat and conservative, failing to effectively predict extreme high and low values ​​in the real data, exhibiting significant deviations from the true values.

[0058] Example 2

[0059] This embodiment provides a soft measurement system for the compressive strength of pellets, including: The historical production data acquisition module is used to acquire multiple sets of historical production data during the production process of the belt roaster. Each set of production data includes time-series data of multiple process parameters and their corresponding offline test values ​​of pellet compressive strength. The time series feature extraction module is used to preprocess the time series data and use the gated loop unit (GRU) to extract the time series features of each process parameter node to obtain the node time series feature matrix. The model building module constructs a dual-branch parallel physical perception spatiotemporal graph neural network model, which includes an adaptive learning branch and a physical prior branch. The node temporal feature matrix is ​​simultaneously input into the adaptive learning branch and the physical prior branch to extract adaptive features and physical prior features respectively. The feature fusion and prediction module is used to fuse adaptive node features with physical prior features to obtain a fused feature vector; the fused feature vector is then input into the regression prediction layer to obtain the predicted value of the pellet compressive strength. The model training module is used to train the physical perception spatiotemporal graph neural network model with multiple sets of historical production data as input and corresponding offline test values ​​as supervision labels, so that the predicted values ​​output by the model approximate the offline test values. In the model training process, a composite loss function including prediction error and entropy regularization is introduced as a sparsity constraint term to optimize the graph structure generated by the adaptive learning branch. The online prediction module is used to collect production process parameters in real time, input them into the trained model, and output real-time predicted values ​​of pellet compressive strength.

[0060] Preferably, the system also includes a visualization module for real-time display of the compressive strength prediction curve and for visually displaying the process variable correlation topology diagram generated by the adaptive learning branch and the attention weight heatmap of each process segment, such as... Figure 4 As shown, the proposed method has an R² of 0.8082, support vector regression has an R² of 0.5608, random forest has an R² of 0.4532, and multilayer perceptron has an R² of 0.5805. This demonstrates that the attention weight heatmap of the proposed method further validates its interpretability. The model clearly quantifies the energy transfer dependencies between different process regions; for example, "UDD" is highly correlated with "Cooling 2" (weight 0.5), and "Cooling 1" is tightly coupled to the upstream "Roasting" (weight 0.35). This indicates that the model not only achieves high-precision prediction but also successfully analyzes the macroscopic physical energy flow laws that conform to industrial norms, providing a physical basis for the prediction results. Experiments show that the prediction accuracy of this invention (R² = 0.8082) is significantly higher than traditional methods such as support vector regression (R² = 0.5608), random forest (R² = 0.4532), and multilayer perceptron (R² = 0.5805).

[0061] Example 3

[0062] This embodiment provides a readable storage medium storing a computer program that, when invoked by a processor, performs the steps of the method described above.

[0063] It should be understood that, in the embodiments of the present invention, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

[0064] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the controller described in any of the foregoing embodiments, such as the controller's hard drive or memory. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both the controller's internal storage unit and external storage devices. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0065] Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0067] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A soft measurement method for the compressive strength of pellets, characterized in that, include: S1: Obtain multiple sets of historical production data during the production process of the belt roaster. Each set of production data includes time-series data of multiple process parameters and their corresponding offline test values ​​of pellet compressive strength. S2: Preprocess the time series data and use the gated loop unit to extract the time series features of each process parameter node to obtain the node time series feature matrix; S3: Construct a dual-branch parallel physical perception spatiotemporal graph neural network model, which includes an adaptive learning branch and a physical prior branch. Input the node temporal feature matrix into both the adaptive learning branch and the physical prior branch to extract the adaptive features and physical prior features respectively. S4: Fuse the adaptive node features with the physical prior features to obtain the fused feature vector; S5: Input the fused feature vector into the regression prediction layer to obtain the predicted value of the pellet compressive strength; S6: Using multiple sets of production data from S1 as input and the corresponding offline test values ​​as supervision labels, train the physical perception spatiotemporal graph neural network model so that the predicted values ​​output by the model approximate the offline test values; wherein, during the model training process, a composite loss function including prediction error and entropy regularization is introduced as a sparsity constraint term to optimize the graph structure generated by the adaptive learning branch. S7: Input the real-time collected production process parameters into the trained model and output the real-time predicted value of the pellet compressive strength.

2. The method according to claim 1, characterized in that, In step S2, the specific process of extracting the temporal features of each process parameter node using a gated recurrent unit (GRU) to obtain the node temporal feature matrix is ​​as follows: ； ； in, Indicates the first element in the input data. One sensor in Normalized measurement value at time; Indicates the first One sensor in The hidden state vector at time t; Indicates the first The sensor at the previous moment The hidden state vector at time t; Indicates the first A sensor passed The final output node temporal feature vector after processing at each time step has a dimension of . .

3. The method according to claim 1, characterized in that, S2 also includes, for each sensor node Randomly initialize a learnable static node embedding vector. This leads to the static embedding vector matrix. : ； in, The dimension in which the node is embedded.

4. The method according to claim 3, characterized in that, In S3, the specific process of extracting adaptive features is as follows: S311: Based on node embedding vector matrix Calculate the similarity between nodes: ； in, Representing nodes respectively and nodes Learnable embedding vectors; Represents a node and nodes The original similarity scores between them; S312: Introducing a dual sparsity mechanism to generate an adaptive sparse adjacency matrix: ； ； in, Represents the linear rectification activation function; Represents a node With nodes The original similarity between them; Represents a node With nodes The non-negative intermediate similarity score between the two sides after hard truncation and temperature scaling; Indicates the temperature coefficient; This represents the nodes in the generated adaptive sparse adjacency matrix. For nodes Normalized connection weights; S313: Based on the generated adaptive sparse adjacency matrix Using graph convolutional networks to analyze the temporal features of nodes Perform spatial aggregation and extract adaptive features : ； in, This is the node temporal feature matrix output by S2, with a matrix dimension of... ; The learnable weight matrix has a dimension of . ; It is a non-linear activation function; This represents the final output adaptive node feature matrix.

5. The method according to claim 1, characterized in that, In S3, the specific process of extracting physical prior features is as follows: S321: Based on the physical process flow of pellet roasting, divide the N sensor nodes into K process regions; S322: Computation Node Original importance score within its respective process area: ； in, For nodes Adaptive features; These are the learnable weight matrix and bias term of the attention mechanism, respectively. Assign an attention scoring vector; For nodes The original importance score; S323: Softmax normalize the original importance scores of all nodes within the process region to obtain normalized attention weights: ； in, For nodes Normalized attention weights within its respective process region; For the first One process area; S324: Weighted aggregation yields the region feature vector: ； in, For the aggregated first Prior feature vectors of each process region; S325: Constructing a physical mask matrix based on the physical laws of the pellet roasting process. ; S326: Update region features using a graph attention network combined with a physical mask matrix: ； ； ； ； in, The linear transformation weight matrix of the GAT layer; This represents the parameter vector for the attention mechanism. This is a vector concatenation operation; Process area based on data calculation For process area The original attention coefficient; For the process area after injecting physical mask For process area Corrected attention coefficient; For process area arrive The mask matrix; These are the final attention weights after Softmax normalization; For the updated number The prior physical characteristics of each region; It is a non-linear activation function.

6. The method according to claim 5, characterized in that, The constructed physical mask matrix satisfies: if the process region arrive If there is a physical flow of materials or air, then ,otherwise .

7. The method according to claim 1, characterized in that, S6 includes a composite loss function of prediction error and entropy regularization. Specifically: ； ； ； in This is the mean square error loss; This refers to the number of sample batches. For the first The true compressive strength test values ​​of each batch of samples; This represents the predicted compressive strength of the pellets output by the model. The regularization coefficient is used. This represents the nodes in the generated adaptive sparse adjacency matrix. For nodes Normalized connection weights; This represents the number of sensor nodes; This is the information entropy regularization term; It is a constant.

8. The method according to claim 1, characterized in that, The process parameters include at least one of the following: air box temperature, air box pressure, burner temperature, burner flow rate, fan speed, machine speed, and material layer thickness in the drying section, preheating section, calcination section, and cooling section.

9. A soft measurement system for the compressive strength of pellets, characterized in that, The steps for performing the method according to any one of claims 1-8 include: The historical production data acquisition module is used to acquire multiple sets of historical production data during the production process of the belt roaster. Each set of production data includes time-series data of multiple process parameters and their corresponding offline test values ​​of pellet compressive strength. The time series feature extraction module is used to preprocess the time series data and use the gated loop unit (GRU) to extract the time series features of each process parameter node to obtain the node time series feature matrix. The model building module constructs a dual-branch parallel physical perception spatiotemporal graph neural network model, which includes an adaptive learning branch and a physical prior branch. The node temporal feature matrix is ​​simultaneously input into the adaptive learning branch and the physical prior branch to extract adaptive features and physical prior features respectively. The feature fusion and prediction module is used to fuse adaptive node features with physical prior features to obtain a fused feature vector; the fused feature vector is then input into the regression prediction layer to obtain the predicted value of the pellet compressive strength. The model training module is used to train the physical perception spatiotemporal graph neural network model with multiple sets of historical production data as input and corresponding offline test values ​​as supervision labels, so that the predicted values ​​output by the model approximate the offline test values. In the model training process, a composite loss function including prediction error and entropy regularization is introduced as a sparsity constraint term to optimize the graph structure generated by the adaptive learning branch. The online prediction module is used to collect production process parameters in real time, input them into the trained model, and output real-time predicted values ​​of pellet compressive strength.

10. The system according to claim 9, characterized in that, The system also includes a visualization module for displaying the compressive strength prediction curve in real time, and visually displaying the process variable correlation topology diagram generated by the adaptive learning branch and the attention weight heatmap of each process segment.

Images