Ore grinding overflow particle size soft measurement method based on automatic feature selection of time-delay causal model

Abstract

The invention discloses an automatic feature selection ore grinding overflow granularity soft measurement method based on a time-delay causal model, and belongs to the technical field of automatic measurement of an ore grinding system. The method comprises the following steps: collecting ore grinding process data and forming a historical sample data set; using a method based on a time-delay causal model to find an optimal delay time point influencing the overflow granularity for each process variable, and selecting features having a strong time-delay causal relationship with the ore grinding overflow granularity to form a causal feature set; and by taking the causal feature set as an input variable and the ore grinding overflow granularity as an output, constructing and training a CH-SBiLSTM soft measurement model, and selecting a model with the lowest mini-batch RMSE as a final model. According to the method, deep fusion of feature selection and model training is realized, dynamic characteristics of the industrial process can be accurately captured, resistance to abnormal values of different degrees is achieved, the accuracy, robustness and practicability of ore grinding overflow particle size soft measurement are remarkably improved, and certain data support is provided for subsequent optimization of the ore grinding control process.

Description

Technical Field

[0001] This invention relates to the field of automated measurement technology for grinding systems, and specifically to a soft measurement method for automatically selecting grinding overflow particle size based on a time-delay causal model. Background Technology

[0002] The grinding process is the most energy-intensive and costly stage in a mineral processing plant, accounting for 45-70% of total energy consumption and 40-60% of total production costs. Clearly, grinding plays a crucial role in the entire mineral processing production process. Its function is to grind the mined ore to a suitable particle size so that subsequent beneficiation processes can efficiently separate valuable minerals from gangue minerals. Grinding overflow particle size is a core indicator for measuring the quality of grinding products, and it has a profound impact on the production indicators of subsequent operations and even the overall economic and technical indicators of the mineral processing plant, as well as the production efficiency, concentrate grade, and resource recovery rate of the entire process. Therefore, real-time measurement of grinding overflow particle size is of great significance for guiding the grinding production process.

[0003] Traditional methods for detecting grinding overflow particle size mainly include offline analysis and online detection. Offline manual sampling and detection relies on manual operation, which suffers from problems such as long detection time intervals, strong subjectivity, and delayed result feedback, making it difficult to meet the needs of real-time monitoring and precise control in modern mineral processing production. While online particle size analyzer detection can provide relatively accurate and timely measurement results, the instruments are expensive, maintenance costs are high, and the requirements for the on-site operating environment are harsh, making it unaffordable for most mineral processing plants. Based on this, soft measurement technology has emerged. It constructs a mathematical model and uses easily obtainable auxiliary variables from the production process to infer the grinding overflow particle size, which is difficult to measure directly. This enables real-time monitoring of grinding overflow particle size, providing a new approach and effective method for solving the problem of grinding overflow particle size measurement.

[0004] In recent years, deep learning has attracted attention for data-driven soft sensor modeling due to its ability to learn abstract latent features from data and its applicability in soft sensors. However, in feature selection, most methods neglect the lag characteristics of industrial processes. Grinding is a typical dynamic process with large lag; changes in process variables often take a certain amount of time to be reflected in the overflow granularity. Existing methods mostly use instantaneous values ​​or simple time windows for feature selection, failing to deeply explore the time-lag causal relationships between variables. This results in the selected feature set failing to accurately reflect the true dynamic mechanism of the process, leading to a large amount of invalid or redundant information in the model input. In terms of model construction, the dynamic information of the quality variables themselves is not fully utilized. Most soft sensor methods based on deep learning models such as LSTM and CNN focus on extracting features from process variables, lacking effective utilization of the dynamic information contained in the historical quality variable sequence, thus limiting the completeness of the description of the process's dynamic behavior. In terms of model robustness, the ability to handle outliers is insufficient. Industrial field data generally contains noise, fluctuations and outliers. Existing models mostly use loss functions such as mean squared error (MSE), which are very sensitive to outliers, leading to instability in the training process and a decline in model performance. Although improvement methods such as Huber loss are effective to some extent, their hyperparameters (such as segmented thresholds) usually need to be set statically based on experience, which cannot adapt to changes in the distribution of outliers in the data. Their generalization ability is limited in complex and ever-changing industrial scenarios.

[0005] In summary, existing soft measurement techniques for grinding overflow particle size cannot simultaneously meet the requirements of scientific feature selection, full utilization of dynamic information, and robustness against outliers. Therefore, there is an urgent need to develop an intelligent soft measurement method that can automatically identify time-delay causal features, fully mine dynamic information of process and quality variables, and has adaptive outlier handling capabilities, in order to improve the accuracy, reliability, and practicality of grinding overflow particle size prediction. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a soft measurement method for automatically selecting grinding overflow particle size based on a time-delay causal model.

[0007] To achieve the above technology, the steps include: S1. Collect grinding process data to form a historical sample dataset.

[0008] S2. Using an automatic feature selection method based on a time-delay causal model, find the optimal lag time point for each process data variable to affect the grinding overflow particle size, and select features that have a strong time-delay causal relationship with the grinding overflow particle size to form a causal feature set S. The steps are as follows: S2.1 Divide the grinding historical sample dataset into two parts: the first part is the grinding overflow particle size, and the second part is the remaining process data; set the remaining process data as the input feature set. F Set the grinding overflow particle size as the output feature. Y ; S2.2, Use a histogram-based method to process the input feature set. F and output features Y Calculate the optimal number of bins, perform discretization, and obtain discretized data; The expression for calculating the optimal number of bins is as follows: In the formula, Indicates the optimal number of boxes; Indicates the entire range of data; Indicates the interquartile range; Indicates the number of samples; This represents the function that takes the maximum value. S2.3 Calculate the input feature set based on the discretized data. F Each input feature With output features Y The mutual information value, and based on all input features and output features Y The mutual information values ​​are sorted from largest to smallest to obtain a new set of candidate input features. ; The expression for calculating mutual information value is as follows: In the formula, Represents mutual information value; Represents the input feature set F Any process variable in the process; This indicates the output characteristic, namely the grinding overflow particle size; Represents the input feature set after discretization F The "box" or interval to which any process variable belongs; Represents the output features after discretization The "box" to which (grind overflow particle size) belongs; express X and Y The joint probability density; express X The marginal probability density; express Y The marginal probability density; S2.4 Set the maximum lag time step to obtain the lagged variables of each variable in the new candidate input feature set; The maximum lag time step is set as follows: Based on knowledge of grinding processes, there is a lag of 10-30 minutes during the grinding process, therefore, the maximum lag time step in this invention is... Set to 30 minutes; The lagged variable representation of each variable in the new candidate input feature set is as follows: Among them, the lag step size , Indicates time; S2.5. Create an empty feature set and calculate the lagged variables for each variable in descending order. Compared with the current output Y ( t ) causal effect ; Calculating causal effects The expression is as follows: In the formula, Indicates causal effect; Indicates a lagged variable; Indicates the current output; Represents the new candidate input feature set In One variable; Represents Shannon entropy; S2.6, Finding the causal effect Maximum lag step And record the corresponding causal effect value. ; S2.7, Corresponding causal effect value obtained based on records If settings Then the features are considered Lag Place and Y If a causal relationship exists, this feature and its optimal hysteresis step size should be added as a new feature to the feature set. S middle, The final set of causal features S It is a set containing variables and their optimal lag time points, for example... .

[0009] S3, the set of causal features constituted S The variables in the model are used as input variables for the CH-SBiLSTM soft sensor model, and the grinding overflow particle size is used as the output variable. The CH-SBiLSTM soft sensor model is trained, and each training mini-batch is selected. mini- batch The model with the lowest RMSE is used as the final model; The steps are as follows: S3.1 Construct the CH-SBiLSTM soft measurement model; The data passes through the sequential input layer, SBiLSTM layer, dropout optimization layer, fully connected layer and output layer in sequence, and the loss function of the model is improved. The sequential input layer feeds the given time series data into the neural network in chronological order. The input to the SBiLSTM layer includes: the current features Optimal hysteresis eigenvector and past quality variable window Using a bidirectional LSTM, the result is output by summing elements through the forward and backward hidden states. The dropout optimization layer randomly removes the output of each SBiLSTM unit with a specific probability P to prevent model overfitting. Fully connected layers use weights and biases to output the quality of features and predictions. Connected; The output layer outputs the final prediction result of the model; The improved model loss function is achieved by using the improved Huber loss function, namely the Capped Huber loss function, as the model's loss function, as shown in the following formula: In the formula, This represents the Capped Huber loss function; Represents the true value, i.e., the first... Each node corresponds to the actual measured value of the grinding overflow particle size; This represents the predicted value, i.e., the first... The predicted value of the grinding overflow particle size corresponding to each node; This represents the first hyperparameter, which determines the emphasis on squared loss and linear loss in the piecewise function; This represents the second hyperparameter, indicating the tolerance for outliers; The first hyperparameter is dynamically generated based on the quantiles of the error distribution in each training mini-batch. Second hyperparameter For each training mini-batch sample, calculate its absolute prediction error sequence, expressed as: In the formula, This represents the absolute prediction error sequence; This represents the first true value in each training mini-batch sample; This represents the first predicted value in each training mini-batch sample; In each training mini-batch sample, the first... One true value; In each training mini-batch sample, the first... One predicted value; Calculate the median Q50 and the 95th percentile Q95 of this sequence: In the formula, Indicates the 50% quantile; Indicates the 95th percentile; Calculate the quantile ratio R. The larger this ratio, the "thicker" the tail of the data, indicating more and more outliers and more extreme values. The expression is as follows: Then we have: In the formula, Indicates the global scaling factor; Indicates the global scaling factor; This represents the outlier tolerance threshold dynamically calculated for the current small batch. This represents the loss switching threshold dynamically calculated for the current small batch; the calculated threshold will be used to... and Substitute into the Capped Huber loss function for loss calculation of the current batch; S3.2 Training the CH-SBiLSTM soft measurement model; S3.2.1. Construct a dataset for model training using the variables in the causal feature set obtained in S2 and the corresponding grinding overflow granularity values. ,in The values ​​of all variables in the causal feature set representing the T-th time step of the training process. Indicates the time step from the first time step to the second time step. The output value at each time step, let , Indicates the first step at time T. The values ​​of the causal variables are determined and divided into training and test sets according to a certain ratio. S3.2.2 Determining the dynamic order of the quality variable through trial and error k The hyperparameters of the model were determined using a grid search method, including the mini-batch size, maximum number of epochs, and initial learning rate. Number of hidden neurons, global scaling factor Global scaling factor ; S3.2.3 Initialize the input weights using the Glorot initializer. The cyclic weights are initialized using the orthogonal matrix obtained from QR decomposition. ; S3.2.4. Divide the training set into mini-batch subsets using mini-batch gradient descent. For all integers... ,use Optimal hysteresis eigenvector and past quality variable window The CH-SBiLSTM network was trained using the Adam optimizer; and selected... mini-batch The model with the lowest RMSE; S3.2.5, Use the test set to test the trained dataset. mini-batch RMSE The lowest-level CH-SBiLSTM soft sensor model was tested and passed the testing phase. To evaluate the model's performance, if If the actual working conditions are met, save the model; otherwise, proceed to S3.2 for retraining.

[0010] S4. Use the trained CH-SBiLSTM soft measurement model to predict the particle size of grinding overflow.

[0011] Beneficial effects of the present invention This invention can automatically select features and determine their optimal lag time based on the time-delay causal relationship between process variables and key quality variables, ensuring the scientific validity and effectiveness of input features from the source. The constructed CH-SBiLSTM model not only utilizes a bidirectional structure and historical quality information to mine nonlinear dynamic features, but also employs a Capped Huber loss function with adaptively adjustable hyperparameters, enabling the model to automatically optimize its outlier handling strategy based on the real-time distribution of training data. Through the synergistic effect of the above technologies, high-precision and robust online prediction of grinding overflow particle size is achieved.

[0012] This invention overcomes the shortcomings of traditional methods that ignore process lag by introducing time-delay causal analysis. It can quantify and locate the optimal time point at which each process variable affects the quality variable, thereby constructing a feature subset with clear causality and the largest amount of information. This improves the quality of the model input data from the source and lays a solid foundation for high-precision prediction.

[0013] This invention proposes a CH-SBiLSTM network for data-driven dynamic soft measurement modeling. By combining hysteresis information with extended quality information utilization and a quality information window spanning k time steps, and leveraging a bidirectional architecture and dynamic quality supervision structure, the CH-SBiLSTM network can extract and utilize nonlinear dynamic latent information from process and quality variables in industrial process data. By employing a Capped Huber loss function and innovatively enabling its hyperparameters to dynamically and adaptively adjust according to the error distribution of each training batch, the model exhibits strong resistance to complex and variable noise and outliers in industrial data. This mechanism eliminates the need for tedious manual parameter tuning, ensuring fitting accuracy to normal data while intelligently suppressing outlier interference, significantly improving model stability and generalization ability. From the selection of time-delay causal features to the adaptive adjustment of loss function hyperparameters, the entire method greatly reduces reliance on domain prior knowledge and manual intervention, forming an end-to-end, highly automated soft measurement solution that is more suitable for complex real-world industrial environments. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a structural diagram of the SBiLSTM unit of the present invention; Figure 3 This is a diagram of the bidirectional structure of the SBiLSTM network of this invention; Figure 4 This is a schematic diagram of the CH-SBiLSTM network soft measurement model of the present invention; Figure 5 This is a graph showing the predicted particle size of the grinding overflow in this invention. Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Please see Figure 1-5 A soft measurement method for automatic feature selection of grinding overflow particle size based on a time-delay causal model includes the following steps: S1. Collect grinding process data to form a historical sample dataset; In this invention, based on the data that can be collected during the grinding process of the ore dressing plant, production data from the past month is selected from the historical database as a historical sample dataset, including process data such as mill power, feed water volume, and grinding overflow particle size; In this embodiment, the constructed historical sample dataset contains more than 50 variables and a total of 100,000 data entries.

[0016] S2. Using an automatic feature selection method based on a time-delay causal model, find the optimal lag time point for each process data variable to affect the grinding overflow particle size, and select features that have a strong time-delay causal relationship with the grinding overflow particle size to form a causal feature set S. The steps are as follows: S2.1 Divide the grinding historical sample dataset into two parts: the first part is the grinding overflow particle size, and the second part is the remaining process data; set the remaining process data as the input feature set. F Set the grinding overflow particle size as the output feature. Y ; S2.2, Use a histogram-based method to process the input feature set. F and output features Y Calculate the optimal number of bins, perform discretization, and obtain discretized data; The expression for calculating the optimal number of bins is as follows: In the formula, This represents the optimal number of bins, i.e., the recommended number of histogram intervals; Indicates the entire range of data; Indicates the interquartile range; Indicates the number of samples; This represents the function that takes the maximum value. S2.3 Calculate the input feature set based on the discretized data. F Each input feature With output features Y The mutual information value, and based on all input features and output features Y The mutual information values ​​are sorted from largest to smallest to obtain a new set of candidate input features. ; The expression for calculating mutual information value is as follows: In the formula, Mutual information value represents a certain input feature. (e.g., mill power) and output characteristics Y The correlation between (grinding overflow particle size) values ​​indicates that the process variable contains more information about the overflow particle size. Represents the input feature set F Any process variable in the process, such as mill power; This indicates the output characteristic, namely the grinding overflow particle size; Represents the input feature set after discretization FThe "box" or interval to which any process variable belongs; Represents the output features after discretization The "box" to which (grind overflow particle size) belongs; express X and Y The joint probability density; express X The marginal probability density; express Y The marginal probability density; S2.4 Set the maximum lag time step to obtain the lagged variables of each variable in the new candidate input feature set; The maximum lag time step is set as follows: Based on knowledge of grinding processes, there is a lag of 10-30 minutes during the grinding process, therefore, the maximum lag time step in this invention is... Set to 30 minutes; The lagged variable representation of each variable in the new candidate input feature set is as follows: Among them, the lag step size , Indicates time; S2.5. Create an empty feature set and calculate the lagged variables for each variable in descending order. Compared with the current output Y ( t ) causal effect ; Calculating causal effects The expression is as follows: ) in, ; ; In the formula, Indicates causal effect; Indicates a lagged variable; Indicates the current output; Represents the new candidate input feature set In One variable; Represents Shannon entropy; S2.6, Finding the causal effect Maximum lag step And record the corresponding causal effect value. ; Maximum lag step The expression is as follows: ; Corresponding causal effect value The expression is as follows: ; S2.7, Corresponding causal effect value obtained based on records If settings Then the features are considered Lag Place and Y If a causal relationship exists, this feature and its optimal hysteresis step size should be added as a new feature to the feature set S. The final causal feature set S is a set containing variables and their optimal lag time points, for example... ; In this invention, the final causal feature set S includes six variables: mill power, feed water volume, discharge water volume, feed rate, pump pool level, and classifier power, with optimal lag times of 22, 24, 2, 25, 4, and 3, respectively.

[0017] S3. Use the variables in the causal feature set S as the input variables of the CH-SBiLSTM soft sensor model, and the grinding overflow particle size as the output variable. Train the CH-SBiLSTM soft sensor model and select the model with the lowest mini-batch RMSE as the final model. The steps are as follows: S3.1 Constructing the CH-SBiLSTM soft measurement model, the steps include: S3.1.1 Constructing a sequential input layer: The given time series data (multivariate data samples related to grinding overflow particle size prediction arranged in chronological order) are fed into the neural network in chronological order; S3.1.2 Constructing the SBiLSTM layer: The SBiLSTM layer consists of multiple SBiLSTM neural units, which are used to construct the optimal lag feature vector. and historical quality variable window As an input gate Forgotten Gate Output gate and cell renewal status As part of the activation input, in each SBiLSTM unit, the sequence input data is transformed into nonlinear dynamic features through cell update states and three nonlinear gates. By learning these nonlinear dynamic features, the memory cells in the SBiLSTM unit... Nonlinear dynamic information is stored as long-term memory, while hiding state variables. This indicates the time of each sample. tThe input gate of the short-term memory of the SBiLSTM unit Forgotten Gate Output gate Cell renewal status Cell state and hidden state The calculation formula is as follows: In the formula, Indicates the input gate; This represents the sigmoid activation function; Indicates the current feature, i.e. t Input characteristics at any given time; express t The optimal lag feature vector at time 1; express The hidden state at any given moment; A window representing historical quality variables; This represents the bias term of the input gate; Indicates the first weight of the input gate; This represents the second weight of the input gate; This represents the third weight of the input gate; This represents the weight corresponding to the historical quality variable window of the input gate; In the formula, Indicates the Gate of Oblivion; Indicates the first weight of the forget gate; Indicates the second weight of the forget gate; This represents the third weight of the forget gate; The bias term representing the forget gate; The weights corresponding to the historical quality variable window of the forget gate; In the formula, Indicates the cell renewal status; Represents the hyperbolic tangent activation function; Bias terms representing the cell renewal state; The first weight representing the cell renewal state; The second weight representing the cell renewal state; The third weight representing the cell renewal state; The weights corresponding to the historical quality variable window representing cell renewal status; In the formula, Indicates the output gate; This represents the bias term of the output gate; Indicates the first weight of the output gate; This represents the second weight of the output gate; This represents the third weight of the output gate; This represents the weight corresponding to the historical quality variable window of the output gate; In the formula, Represents memory cells; This represents point-by-point multiplication; In the formula, Indicates a hidden state; The final output of the SBiLSTM unit is: In the formula, This indicates summation by elements. This represents the hidden state of the SBiLSTM cell during forward propagation. This represents the hidden state during backpropagation; S3.1.3 Constructing the dropout optimization layer: The output generated by the SBiLSTM network This is then passed to the dropout optimization layer, which randomly removes the output of each SBiLSTM unit with a specific probability P to prevent overfitting. Therefore, the quality output becomes: In the formula, This indicates that after processing by the dropout layer, the first... The output of each node; Indicates connection weight; Indicates the dropout mask; Indicates the first generation generated by SBiLSTM One feature; This represents the bias term of the dropout optimization layer; where, ; S3.1.4 Constructing a fully connected layer: Using weights and biases, the features and predictions are combined to form a high-quality output. Connected; S3.1.5 Constructing the output layer: Outputting the final prediction results of the model; S3.1.6 Improved model loss function: The improved Huber loss function, namely the Capped Huber loss function, is used as the model's loss function, and its formula is as follows: In the formula, This represents the Capped Huber loss function; Represents the true value, i.e., the first... Each node corresponds to the actual measured value of the grinding overflow particle size; This represents the predicted value, i.e., the first... The predicted value of the grinding overflow particle size corresponding to each node; This represents the first hyperparameter, which determines the emphasis on squared loss and linear loss in the piecewise function; This represents the second hyperparameter, indicating the tolerance for outliers; S3.1.7 Dynamically generate the first hyperparameter based on the quantiles of the error distribution in each training mini-batch. Second hyperparameter For each training mini-batch sample, calculate its absolute prediction error sequence, expressed as: In the formula, This represents the absolute prediction error sequence; This represents the first true value in each training mini-batch sample; This represents the first predicted value in each training mini-batch sample; In each training mini-batch sample, the first... One true value; In each training mini-batch sample, the first... One predicted value; Calculate the median Q50 and the 95th percentile Q95 of this sequence: In the formula, Indicates the 50% quantile; Indicates the 95th percentile; Calculate the quantile ratio R. The larger this ratio, the "thicker" the tail of the data, indicating more and more outliers and more extreme values. The expression is as follows: Then we have: In the formula, Indicates the global scaling factor; Indicates the global scaling factor; This represents the outlier tolerance threshold dynamically calculated for the current small batch. This represents the loss switching threshold dynamically calculated for the current small batch; the calculated threshold will be used to... and Substitute into the Capped Huber loss function for loss calculation of the current batch; S3.2 Training the CH-SBiLSTM soft measurement model, the steps are as follows: S3.2.1. Construct a dataset for model training using the variables in the causal feature set obtained in S2 and the corresponding grinding overflow granularity values. , ,in The values ​​of all variables in the causal feature set representing the T-th time step of the training process. Indicates the time step from the first time step to the second time step. The output value at each time step, let , Indicates the first step at time T. The values ​​of the causal variables are determined and divided into training and test sets according to a certain ratio. In this embodiment, 70% is divided into the training set and 30% into the test set; S3.2.2 Determining the dynamic order of the quality variable through trial and error k The hyperparameters of the model were determined using a grid search method, including the mini-batch size, maximum number of epochs, and initial learning rate. Number of hidden neurons, global scaling factor Global scaling factor ; The final determined dynamic order k The initial learning rate is 3, the mini-batch size is 20, the maximum number of rounds is 100, and the initial learning rate is... The value is 0.01, the number of hidden neurons is 15, and the global scaling factor is... The global scaling factor is 1. It is 0.5; S3.2.3 Initialize the input weights using the Glorot initializer. The cyclic weights are initialized using the orthogonal matrix obtained from QR decomposition. ; S3.2.4. Divide the training set into mini-batch subsets using mini-batch gradient descent. For all integers... ,use Optimal hysteresis eigenvector and past quality variable window The CH-SBiLSTM network is trained using the Adam optimizer, where: In the formula, Indicates the starting offset of the valid sample; Indicates the range of valid time points; Indicates the first The value of each variable at the optimal lag step; And select mini-batch The model with the lowest RMSE mini-batch The formula for calculating RMSE is as follows: In the formula, This indicates the root mean square error of a small batch. Indicates the number of samples in a mini-batch; S3.2.5, Use the test set to test the trained dataset. mini-batch RMSE The lowest-level CH-SBiLSTM soft sensor model was tested and passed the testing phase. To evaluate the model's performance, if If the model meets the actual working conditions, save it; otherwise, proceed to S3.2 for retraining. The calculation formula is as follows: in, Indicates the number of samples in the test set; Indicates the number of samples in the training set; Indicates the time range of the test samples; Indicates the first test set The true value of a moment; Indicates the first test set The predicted value at any given time.

[0018] S4. Predict grinding overflow particle size using the trained CH-SBiLSTM soft measurement model. The prediction method is as follows: The new input variable vector... Optimal hysteresis eigenvector and historical quality variable window As input to the CH-SBiLSTM network and used to train the network, the computation time is... tx Two-way hidden state at time Cell state and network output By This information is passed to the dropout optimization layer, the fully connected layer, and the output layer to predict the grinding overflow granularity. .

[0019] Figure 5The graph shows the predicted particle size of the grinding overflow. As can be clearly seen, the predicted curve and the actual curve show a high degree of overlap in their overall trends. This indicates that the method used in this paper can accurately capture the dynamic changes in the grinding overflow particle size. Whether the particle size is rising, falling, or in a stable phase, the predicted results closely follow the fluctuations of the actual value, overcoming the large time lag characteristic of the grinding process and demonstrating the model's strong ability to capture the dynamic characteristics of the process. Observing the curve spacing, the deviation between the predicted and actual values ​​remains within a small range, without significant large deviations. The low deviation characteristic of the data in the graph confirms that the model effectively controls the prediction error by optimizing the loss function, especially exhibiting good resistance to potential outliers and ensuring the stability of the error.

[0020] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A soft measurement method for automatic feature selection of grinding overflow particle size based on a time-delay causal model, characterized in that, Includes the following steps: S1. Collect grinding process data to form a historical sample dataset; S2. Using an automatic feature selection method based on a time-delay causal model, find the optimal lag time point that affects the grinding overflow particle size for each process data variable, and select features that have a strong time-delay causal relationship with the grinding overflow particle size to form a causal feature set. The causal feature set includes six variables: mill power, feed water volume, discharge water volume, feed volume, pump pool level, and classifier power. The optimal lag times are 22, 24, 2, 25, 4, and 3, respectively. S3. Use the causal feature set as the input of the CH-SBiLSTM soft measurement model and the grinding overflow particle size as the output to train the CH-SBiLSTM soft measurement model, and select the model with the lowest RMSE in each training mini-batch as the final model. S4. Use the trained CH-SBiLSTM soft sensor model to predict the grinding overflow particle size, thus completing the construction of the soft sensor method.

2. The method for automatic feature selection of grinding overflow particle size based on a time-delay causal model according to claim 1, characterized in that: In step S2, the steps for constructing the causal feature set include: S2.1 Divide the grinding history sample dataset into two parts: the first part is the grinding overflow particle size, and the second part is the remaining process data. Set the remaining process data as the input feature set. F Set the grinding overflow particle size as the output feature. Y ; S2.2, Use a histogram-based method to process the input feature set. F and output features Y Calculate the optimal number of bins, perform discretization, and obtain discretized data; The expression for calculating the optimal number of bins is as follows: In the formula, This represents the optimal number of bins, i.e., the recommended number of histogram intervals; Indicates the entire range of data; Indicates the interquartile range; Indicates the number of samples; This represents the function that takes the maximum value. S2.3 Calculate the input feature set based on the discretized data. F Each input feature With output features Y The mutual information value, and based on all input features and output features Y The mutual information values ​​are sorted from largest to smallest to obtain a new set of candidate input features. ; S2.4 Set the maximum lag time step to obtain the lagged variables of each variable in the new candidate input feature set; Maximum time lag step Set to 30 minutes; The lagged variable representation of each variable in the new candidate input feature set is as follows: Among them, the lag step size , Indicates time; S2.

5. Create an empty feature set and calculate the lagged variables for each variable in descending order. Compared with the current output Y ( t ) causal effect ; Calculating causal effects The expression is as follows: In the formula, Indicates causal effect; Indicates a lagged variable; Indicates the current output; Represents the new candidate input feature set In One variable; Represents Shannon entropy; S2.6, Finding the causal effect Maximum lag step And record the corresponding causal effect value. ; S2.7, Corresponding causal effect value obtained based on records If settings Then the features are considered Lag Place and Y If a causal relationship exists, this feature and its optimal lag step size should be added as a new feature to the causal feature set. S middle, .

3. The method for automatic feature selection of grinding overflow particle size based on a time-delay causal model according to claim 1, characterized in that: The steps in S3 include: S3.1 Construct the CH-SBiLSTM soft measurement model; In the CH-SBiLSTM soft measurement model, the data passes through the sequential input layer, SBiLSTM layer, dropout optimization layer, fully connected layer and output layer in sequence, and the loss function of the model is improved. The input to the SBiLSTM layer includes: current features Optimal hysteresis eigenvector and past quality variable window Using a bidirectional LSTM, the result is output by summing elements through the forward and backward hidden states. The dropout optimization layer randomly removes the output of each SBiLSTM unit with a preset probability P to prevent model overfitting. S3.2 Training the CH-SBiLSTM soft measurement model.

4. The method for automatic feature selection of grinding overflow particle size based on a time-delay causal model according to claim 3, characterized in that: In the CH-SBiLSTM soft measurement model, the improved model's loss function is as follows: the improved Huber loss function, i.e., the Capped Huber loss function, is used as the model's loss function, and its formula is as follows: In the formula, This represents the Capped Huber loss function; Represents the true value, i.e., the first... Each node corresponds to the actual measured value of the grinding overflow particle size; This represents the predicted value, i.e., the first... The predicted value of the grinding overflow particle size corresponding to each node; This represents the first hyperparameter, which determines the emphasis on squared loss and linear loss in the piecewise function; This represents the second hyperparameter, indicating the tolerance for outliers; The first hyperparameter is dynamically generated based on the quantiles of the error distribution in each training mini-batch. Second hyperparameter For each training mini-batch sample, calculate its absolute prediction error sequence, expressed as: In the formula, This represents the absolute prediction error sequence; This represents the first true value in each training mini-batch sample; This represents the first predicted value in each training mini-batch sample; In each training mini-batch sample, the first... One true value; In each training mini-batch sample, the first... One predicted value; Calculate the median Q50 and the 95th percentile Q95 of this sequence: In the formula, Indicates the 50% quantile; Indicates the 95th percentile; Calculate the quantile ratio R. The larger this ratio, the "thicker" the tails of the data, indicating more and more outliers and more extreme values. The expression is as follows: Then we have: In the formula, Indicates the global scaling factor; Indicates the global scaling factor; This represents the outlier tolerance threshold dynamically calculated for the current small batch. This represents the loss switching threshold dynamically calculated for the current small batch; the calculated threshold will be used to... and Substitute into the Capped Huber loss function for loss calculation of the current batch.

5. The method for automatic feature selection of grinding overflow particle size based on a time-delay causal model according to claim 3, characterized in that: The steps for training the CH-SBiLSTM soft sensor model include: S3.2.

1. Construct a dataset for model training using the variables in the causal feature set obtained in S2 and the corresponding grinding overflow granularity values. ,in The values ​​of all variables in the causal feature set representing the T-th time step of the training process. Indicates the time step from the first time step to the second time step. The output value at each time step, let ， Indicates the first step at time T. The values ​​of each causal variable are determined and the data are divided into training and test sets according to a predefined ratio. S3.2.2 Determining the dynamic order of the quality variable through trial and error k The hyperparameters of the model were determined using a grid search method, including: mini-batch size, maximum number of epochs, and initial learning rate. Number of hidden neurons, global scaling factor Global scaling factor ; The final determined dynamic order k The initial learning rate is 3, the mini-batch size is 20, the maximum number of rounds is 100, and the initial learning rate is... The value is 0.01, the number of hidden neurons is 15, and the global scaling factor is... The global scaling factor is 1. It is 0.5; S3.2.3 Initialize the input weights using the Glorot initializer. The cyclic weights are initialized using the orthogonal matrix obtained from QR decomposition. ; in, Indicates the first weight of the input gate; Indicates the first weight of the forget gate; The first weight representing the cell renewal state; Indicates the first weight of the output gate; This represents the third weight of the input gate; This represents the third weight of the forget gate; The third weight representing the cell renewal state; This represents the third weight of the output gate; S3.2.

4. Divide the training set into mini-batch subsets using mini-batch gradient descent. For all integers... ,use Optimal hysteresis eigenvector and past quality variable window The CH-SBiLSTM network is trained using the Adam optimizer, where: In the formula, Indicates the starting offset of the valid sample; Indicates the range of valid time points; Indicates the first The value of each variable at the optimal lag step; S3.2.5, Use the test set to test the trained dataset. mini-batch RMSE The lowest-level CH-SBiLSTM soft sensor model was tested and passed the testing phase. To evaluate the model's performance, if If the actual working conditions are met, save the model; otherwise, proceed to S3.2 for retraining.

Images