Ore grinding granularity soft measurement method and device, computer equipment and storage medium

Abstract

The invention relates to the technical field of particle size measurement in the ore grinding production process. The ore grinding granularity soft measurement method comprises the steps of analyzing a time sequence characteristic matrix based on a step-by-step regularization characteristic sorting method, forming an optimized input characteristic matrix, obtaining a linear prediction value based on the optimized input characteristic matrix, and calculating the ore grinding granularity according to the linear prediction value. Performing error calculation on the linear predicted value and the actual measurement value to obtain a nonlinear predicted value, combining the linear predicted value and the nonlinear predicted value according to the weight to generate a preliminary soft measurement model, performing global optimization on the combined weight of the preliminary soft measurement model through a multi-objective optimization algorithm to generate an optimized soft measurement model, and performing soft measurement on the optimal soft measurement model. When the statistical property exceeds a set threshold value, a dynamic correction mechanism is triggered, and an updated soft measurement model is obtained according to the dynamic correction mechanism. The method has the effect of meeting the high-precision prediction requirement under the dynamic working condition.

Description

Technical Field

[0001] The present invention relates to the technical field of particle size measurement in the grinding production process, and in particular, to a soft measurement method, device, computer device and storage medium for grinding particle size. Background Technique

[0002] Currently, in the mineral processing process, grinding, as an important link in material preparation, directly affects the efficiency of ore dressing and the product quality. Traditional grinding particle size detection methods rely on physical detection devices, such as particle size detectors, to obtain the particle size information of grinding products in real time, providing data support for process optimization. However, these physical detection devices have problems such as high cost, complex maintenance, and poor stability under harsh working conditions, making it difficult to meet the requirements of modern ore dressing processes for high-precision particle size control.

[0003] Existing soft measurement technologies, as an alternative or supplement to physical detection, establish a mathematical model by inputting characteristic information to predict the particle size value, thereby reducing the dependence on physical detection devices. Although soft measurement technologies have improved the flexibility of grinding particle size prediction to a certain extent, there are still defects in the dynamic adaptability of the model. Summary of the Invention

[0004] In order to meet the high-precision prediction requirements under dynamic working conditions, the present application provides a soft measurement method, device, computer device and storage medium for grinding particle size.

[0005] The first invention object of the present application is achieved by the following technical solutions: A soft measurement method for grinding particle size, the soft measurement method for grinding particle size includes: Obtain the input characteristic information in the grinding production process, perform denoising processing on the input characteristic information to obtain the preprocessed input characteristic information; Expand the preprocessed input characteristic information according to a time window to generate a time series characteristic matrix including a historical length; Analyze the time series characteristic matrix based on the stepwise regularization feature ranking method to form an optimized input feature matrix. Based on the optimized input feature matrix, establish a linear model using the weighted multi-dimensional regression method to obtain a linear prediction value; obtain the actual measurement value of the particle size detection device in the on-site working condition, calculate the error between the linear prediction value and the actual measurement value to obtain an error value, and based on the error value, perform non-linear modeling using a dynamic weighted recursive network to obtain a non-linear prediction value. Combine the linear prediction value and the non-linear prediction value according to weights to generate a preliminary soft measurement model; The combined weights of the preliminary soft-sensing model are globally optimized by a multi-objective optimization algorithm to generate an optimized soft-sensing model, and the statistical characteristics of the input feature information in the grinding production process are monitored in real time. When the statistical characteristics exceed the set threshold, a dynamic correction mechanism is triggered. According to the dynamic correction mechanism, the parameters of the optimized soft-sensing model are updated by an incremental learning method to obtain an updated soft-sensing model; Based on the updated soft-sensing model, the grinding particle size is predicted in real time, and a particle size prediction value is output. According to the particle size prediction value, adjustment suggestions for grinding process parameters are generated.

[0006] By adopting the above technical solution, the statistical characteristics of the input feature information are monitored in real time through the dynamic correction mechanism, and the changes in the mean value, variance and distribution form of the input feature information can be accurately captured. When the change amplitude exceeds the set threshold, the model correction process is triggered, thereby improving the adaptability of the soft-sensing model to dynamic working conditions. By combining the new information to adjust the regression coefficients and bias terms of the linear model, the fitting effect of the linear model on the input feature information can be optimized, and the prediction error of the model when the features change greatly can be reduced, thereby improving the stability and prediction accuracy of the linear model. By combining the new information to update the network parameters of the non-linear model, the expression ability of the non-linear model for complex non-linear relationships can be enhanced, and the prediction performance of the overall soft-sensing model can be optimized. By dynamically adjusting the weight coefficients of the optimized soft-sensing model and reasonably allocating the weight ratio of the linear model and the non-linear model, the comprehensive prediction ability and adaptability of the soft-sensing model can be further enhanced. By combining the updated model to output the particle size prediction value and generating adjustment suggestions for grinding process parameters, the configuration of grinding production process parameters can be optimized on the basis of accurate particle size prediction, thereby improving the efficiency and stability of grinding production.

[0007] In a preferred example of the present application, it can be further configured that: based on the optimized input feature matrix, a weighted multi-dimensional regression method is used to establish a linear model to obtain a linear prediction value, including: The weighted multi-dimensional regression method establishes the linear model based on the following formula to obtain the linear prediction value: Wherein, the Y linear is the linear prediction value, the β 0 is the bias term of the linear model, and the n is the number of input features; the w i is the weight of the i-th feature, the β i is the regression coefficient of the i-th feature, and the x i is the value of the i-th feature.

[0008] By adopting the above technical solution, the weighted information of the input features can be effectively utilized to perform linear fitting on the target variable. By introducing the weight coefficient wi Dynamically reflect the importance of each feature to the prediction result, improving the interpretability and flexibility of the model through the regression coefficient β i Further accurately quantify the influence degree of each feature value on the target variable, enabling the model to more precisely capture the relationship between the input features and the target variable, thereby improving the fitting accuracy of the linear model through the bias term β 0 The introduction of can effectively correct the overall prediction bias of the model, ensuring that the model output is consistent with the global trend of the target variable.

[0009] In a preferred example of the present application, it can be further configured that: based on the error value, a dynamic weighted recurrent network is used for non-linear modeling to obtain a non-linear prediction value, including: Taking the error value as the modeling target of the dynamic weighted recurrent network, and taking the optimized input feature matrix as the input of the dynamic weighted recurrent network; In the hidden layer of the dynamic weighted recurrent network, adjust the weights and biases of the hidden nodes according to the dynamic changes of the error value, and the dynamic adjustment of the weights and the biases is based on the following formula: wh,i = wh, i-1 -α·Δw h,i-1 , where, the w h,i and the b h,i are respectively the weight and the bias of the i-th hidden node, the α is the adaptive step coefficient, the Δw h,i-1 and the Δb h,i-1 are respectively the update amplitudes of the weight and the bias caused by the previous step error; In the hidden layer of the dynamic weighted recurrent network, recursively calculate the output of each node through a recursive activation function, and the recursive activation function is represented by the following formula: h i = g(w h,i ·x i +b h,i +γ·h i-1 ), where, the hi is the output of the i-th hidden node, the xi is the input feature value, h i-1 is the recursive output of the previous node, and γ is the recursive gain coefficient; In the output layer of the dynamic weighted recurrent network, obtain the non-linear prediction value through the following calculation formula: where the Y nonlinear is the non-linear prediction value, φi is the output weight of the i-th hidden node, and h i is the output value of the i-th hidden node.

[0010] By adopting the above technical solution, by dynamically adjusting the weights and biases of the hidden nodes, the network can flexibly optimize the parameter configuration according to the error change, so as to quickly converge to the optimal state. The introduction of the recursive activation function enables the hidden layer nodes to capture the deep associations between time series information and features, enhancing the network's processing ability for dynamic working conditions. In the output layer, the non-linear prediction value is calculated by weighted accumulation of the outputs of the hidden nodes, comprehensively utilizing the feature contributions of each hidden node to ensure the accurate fitting of the prediction result to the non-linear relationship.

[0011] In a preferred example of the present application, it can be further configured that: combining the linear prediction value and the non-linear prediction value according to weights to generate a preliminary soft measurement model, including: The preliminary soft measurement model is obtained through the following formula: M initial = f linear (X)·λ 1 + f nonlinear (X,E)·λ 2 , where Minitial is the preliminary soft measurement model, flinear(X) represents the mapping of the linear model to the input feature matrix, f nonlinear (X,E) represents the mapping of the non-linear model based on the input feature matrix and the error value, and the λ 1 and the λ 2 are the weight coefficients of the linear model and the non-linear model respectively, and satisfy λ1 + λ2 = 1.

[0012] By adopting the above technical solution, by performing weighted combination on the linear mapping of the input feature matrix and the non-linear mapping based on the error value, the generated preliminary soft measurement model can not only efficiently capture the linear relationship between the input features and the target variable, but also effectively express complex non-linear characteristics. By introducing the weight coefficient, the model can dynamically adjust the contribution ratio of the linear and non-linear parts, so as to better adapt to the actual needs under different working conditions.

[0013] In a preferred example of the present application, it can be further configured that: globally optimizing the combined weights of the preliminary soft measurement model through a multi-objective optimization algorithm to generate an optimized soft measurement model, including: The optimized soft measurement model is obtained through the following soft measurement model calculation formula: where M optimized is the optimized soft measurement model, M k represents the output of the k-th sub-model, and λ k opt is the optimized weight of the k-th sub-model; The λk opt The solution is based on the following objective function: where J(λ) is the objective function value, Z l is the actual particle size measurement value of the sample, q is the number of samples used in the optimization process, β is the adjustment coefficient, and λ k is the k-th initial weight. By iteratively optimizing J(λ), λ is adjusted k to obtain λ k opt Then, substituting λ k opt into the calculation formula of the soft sensor model, the optimized soft sensor model is obtained.

[0014] By adopting the above technical solution, it is possible to effectively balance the contributions of multiple sub-models in the soft sensor model through the dynamic adjustment of the optimized weights, comprehensively utilize the prediction characteristics of each sub-model, and significantly improve the prediction accuracy and robustness of the soft sensor model. Through the construction of the objective function, not only can the error between the model prediction value and the actual particle size measurement value be minimized, but also a regularization term is introduced to control the sparsity of the weight distribution, thereby avoiding the problems of over-concentration or dispersion of weight allocation and improving the stability and rationality of the optimization process. The most important technical benefit is that through the iterative optimization of the weights, the model can adapt to the complex and variable working conditions in the grinding production process, and the generated optimized soft sensor model has higher prediction accuracy and adaptability, providing a reliable technical guarantee for the accurate prediction of grinding particle size and process optimization.

[0015] In a preferred example of the present application, it can be further configured that: during the real-time monitoring of the grinding production process, the statistical characteristics of the input feature information are monitored, and when the statistical characteristics exceed the set threshold, a dynamic correction mechanism is triggered, including: The statistical characteristics of the input feature information are monitored in real time, and the statistical characteristics include the mean value, variance, and distribution form; The mean value, variance, and distribution form are respectively compared with the historical reference benchmark. When the change range of any one of the mean value, variance, or distribution form exceeds the set threshold, a distribution change warning message is generated. After generating the distribution change warning message, the dynamic correction mechanism is started.

[0016] By adopting the above technical solutions, real-time monitoring and dynamic analysis of input feature information can be achieved. Through comprehensive monitoring of the mean, variance, and distribution pattern of statistical characteristics, the change trend of feature information under dynamic working conditions can be captured in a timely manner. By comparing the current statistical characteristics with historical reference benchmarks, abnormal fluctuations in feature data or significant changes in working conditions can be accurately identified, ensuring that the model always reflects the real production state. When it is detected that the statistical characteristics exceed the set threshold, distribution change warning information is generated in a timely manner and the dynamic correction mechanism is activated to ensure that the model can quickly adjust parameters to adapt to the current working conditions.

[0017] In a preferred example of the present application, it can be further configured as follows: According to the dynamic correction mechanism, the parameters of the optimized soft-sensor model are updated by an incremental learning method to obtain an updated soft-sensor model, including: According to the dynamic correction mechanism, obtain the newly added information at the current moment; Based on the newly added information at the current moment, adjust the regression coefficients and bias terms of the linear model through the following formula: where, the β updated，i is the updated regression coefficient, the β previous，i is the regression coefficient before update, the L linear is the loss function of the linear model calculated based on the newly added information, and η is the learning rate. The β updated，0 is the updated bias term, the β previous,0 is the bias term before update; Based on the newly added information at the current moment, adjust the network parameters of the non-linear model through the following formula: where, the φ updated is the updated network parameter, the φ previous is the network parameter before update, the L nonlinear is the loss function of the non-linear model calculated based on the newly added information, and η is the learning rate; Based on the newly added information at the current moment, update the weight coefficients of the optimized soft-sensor model through the following formula: where, λ updated,k is the updated weight coefficient, λ previous,k is the k-th weight coefficient before update, L comb is the combined error loss function of the newly added information in the optimized soft-sensor model, and η is the learning rate; Based on the updated regression coefficients, the updated bias terms, the updated network parameters, and the updated weight coefficients, the updated soft-sensor model is generated through the following formula: The formula for the updated linear prediction value is, The updated formula for the non - linear predicted value is Y nonlinear,updated = f(X new , φ updated ), where Xi is the input feature matrix collected in real - time, and X new is the original value of the i - th feature in the current feature set. The combined formula for the updated soft - sensor model is M updated = λ updated,1 ·Y linear,updated + λ updated,2 ·Y nonlinear,updated , where M updated is the updated soft - sensor model, and λupdated,1 and λupdated,2 are the weights of the updated linear model and non - linear model respectively.

[0018] By adopting the above technical solutions, the soft - sensor model can be optimized layer by layer using the dynamic correction mechanism, significantly improving the real - time performance and adaptability of the model. By obtaining the new information at the current moment and dynamically adjusting the regression coefficients and bias terms of the linear model based on the new information, the deviation of the linear model can be quickly corrected to ensure accurate fitting of the latest working conditions. By updating the network parameters of the non - linear model, the expression ability of the non - linear model for complex non - linear relationships can be enhanced, improving the prediction accuracy of the model under dynamic working conditions. By dynamically optimizing the weight coefficients of the soft - sensor model, a reasonable contribution distribution of the linear model and non - linear model under different dynamic working conditions is achieved, further optimizing the overall prediction performance of the model. Finally, based on the updated parameters, a soft - sensor model is generated, which can output accurate particle size prediction values in real - time, providing a reliable basis for the optimization and adjustment of the grinding process.

[0019] The second above - mentioned invention object of this application is achieved through the following technical solutions: A dynamic grinding particle size soft - sensor device, which includes: An input feature pre - processing module, which is used to obtain the input feature information in the grinding production process, perform denoising processing on the input feature information, and obtain the pre - processed input feature information; A feature optimization and linear modeling module, which is used to expand the pre - processed input feature information according to a time window to generate a time - series feature matrix including a historical length; A non - linear modeling module, which is used to analyze the time - series feature matrix based on the step - by - step regularization feature sorting method to form an optimized input feature matrix, and based on the optimized input feature matrix, establish a linear model using the weighted multi - dimensional regression method to obtain a linear predicted value; The preliminary soft sensor modeling module is used to obtain the actual measurement value of the particle size detection device in the on-site working condition, calculate the error between the linear prediction value and the actual measurement value to obtain an error value, and based on the error value, perform non-linear modeling using a dynamic weighted recursive network to obtain a non-linear prediction value. The linear prediction value and the non-linear prediction value are combined according to weights to generate a preliminary soft sensor model; The dynamic correction and incremental learning module is used to globally optimize the combined weights of the preliminary soft sensor model through a multi-objective optimization algorithm to generate an optimized soft sensor model, and to monitor the statistical characteristics of the input feature information in real time during the grinding production process. When the statistical characteristics exceed the set threshold, a dynamic correction mechanism is triggered. According to the dynamic correction mechanism, the parameters of the optimized soft sensor model are updated through an incremental learning method to obtain an updated soft sensor model; The particle size prediction and process adjustment suggestion module is used to perform real-time prediction of the grinding particle size based on the updated soft sensor model, output a particle size prediction value, and generate a grinding process parameter adjustment suggestion according to the particle size prediction value.

[0020] By adopting the above technical solutions, the statistical characteristics of the input feature information are monitored in real time through the dynamic correction mechanism, which can accurately capture the changes in the mean value, variance and distribution form of the input feature information. When the change amplitude exceeds the set threshold, the model correction process is triggered, thereby improving the adaptability of the soft sensor model to dynamic working conditions. By combining the new information to adjust the regression coefficients and bias terms of the linear model, the fitting effect of the linear model on the input feature information can be optimized, and the prediction error of the model when the features change greatly can be reduced, thereby improving the stability and prediction accuracy of the linear model. By combining the new information to update the network parameters of the non-linear model, the expression ability of the non-linear model for complex non-linear relationships can be improved, and the prediction performance of the overall soft sensor model can be optimized. By dynamically adjusting the weight coefficients of the optimized soft sensor model and reasonably allocating the weight ratio of the linear model and the non-linear model, the comprehensive prediction ability and adaptability of the soft sensor model can be further enhanced. By combining the updated model to output the particle size prediction value and generating a grinding process parameter adjustment suggestion, the configuration of the grinding production process parameters can be optimized on the basis of the accuracy of the particle size prediction, thereby improving the efficiency and stability of the grinding production.

[0021] The above object three of the present application is achieved by the following technical solutions: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned grinding particle size soft sensor method are implemented.

[0022] The above object four of the present application is achieved by the following technical solutions: A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned soft measurement method for grinding particle size are implemented.

[0023] In summary, the present application includes at least one of the following beneficial technical effects: 1. Through the dynamic correction mechanism, the statistical characteristics of the input feature information are monitored in real time, and the changes in the mean value, variance, and distribution form of the input feature information can be accurately captured. When the change amplitude exceeds the set threshold, the model correction process is triggered, thereby improving the adaptability of the soft measurement model to dynamic working conditions. By combining the newly added information to adjust the regression coefficients and bias terms of the linear model, the fitting effect of the linear model on the input feature information can be optimized, and the prediction error of the model when the features change greatly can be reduced, thereby improving the stability and prediction accuracy of the linear model. By combining the newly added information to update the network parameters of the non-linear model, the expression ability of the non-linear model for complex non-linear relationships can be enhanced, and the prediction performance of the overall soft measurement model can be optimized. By dynamically adjusting the weight coefficients of the optimized soft measurement model and reasonably allocating the weight ratio of the linear model and the non-linear model, the comprehensive prediction ability and adaptability of the soft measurement model can be further enhanced. By combining the updated model output particle size prediction value and generating grinding process parameter adjustment suggestions, the grinding production process parameter configuration can be optimized on the basis of the accuracy of particle size prediction, thereby improving the efficiency and stability of grinding production; 2. It can realize the real-time monitoring and dynamic analysis of the input feature information. By comprehensively monitoring the mean value, variance, and distribution form of the statistical characteristics, the change trend of the feature information under dynamic working conditions can be captured in time. By comparing the current statistical characteristics with the historical reference benchmark, the abnormal fluctuations of the feature data or the significant changes in the working conditions can be accurately identified to ensure that the model always reflects the real production state. When it is detected that the statistical characteristics exceed the set threshold, the distribution change warning information is generated in time and the dynamic correction mechanism is started to ensure that the model can quickly adjust the parameters to adapt to the current working conditions; 3. It can use the dynamic correction mechanism to optimize the soft measurement model layer by layer, significantly improving the real-time performance and adaptability of the model. By obtaining the newly added information at the current moment and dynamically adjusting the regression coefficients and bias terms of the linear model based on the newly added information, the deviation of the linear model can be quickly corrected to ensure accurate fitting of the latest working conditions; by updating the network parameters of the non-linear model, the expression ability of the non-linear model for complex non-linear relationships can be enhanced, and the prediction accuracy of the model under dynamic working conditions can be improved; by dynamically optimizing the weight coefficients of the soft measurement model, a reasonable contribution allocation of the linear model and the non-linear model under different dynamic working conditions is realized, further optimizing the overall prediction performance of the model; finally, based on the updated parameters, a soft measurement model is generated, which can output accurate particle size prediction values in real time, providing a reliable basis for the optimization and adjustment of the grinding process. Brief Description of the Drawings

[0024] Figure 1 is a flowchart of a soft measurement method for grinding particle size in an embodiment of the present application; Figure 2 is a flowchart of the implementation in step S30 of the soft measurement method for grinding particle size in an embodiment of the present application; Figure 3 is a flowchart of the implementation in step S40 of the soft measurement method for grinding particle size in an embodiment of the present application; Figure 4 is another flowchart of the implementation in step S40 of the soft measurement method for grinding particle size in an embodiment of the present application; Figure 5 is a flowchart of the implementation in step S50 of the soft measurement method for grinding particle size in an embodiment of the present application; Figure 6 is a flowchart of the implementation in step S50 of the soft measurement method for grinding particle size in an embodiment of the present application; Figure 7 is another flowchart of the implementation in step S50 of the soft measurement method for grinding particle size in an embodiment of the present application; Figure 8 is a principle block diagram of a dynamic soft measurement device for grinding particle size in an embodiment of the present application; Figure 9 is a schematic diagram of the equipment in an embodiment of the present application. Detailed Description of the Embodiment

[0025] The present application will be further described in detail below with reference to the accompanying drawings.

[0026] In an embodiment, as Figure 1 shown, the present application discloses a soft measurement method for grinding particle size, which specifically includes the following steps: S10: Obtain the input feature information in the grinding production process, perform noise reduction processing on the input feature information, and obtain the preprocessed input feature information.

[0027] Specifically, multi-dimensional input feature information related to the grinding process is collected in real time from sensors arranged at the core nodes of the grinding equipment. These features include, but are not limited to, grinding medium filling rate, mill speed, ore hardness, feed particle size, grinding current, and discharge concentration, etc. The collected information is transmitted to the signal processing unit at a fixed frequency through the data acquisition module. High-frequency noise in the collected signal is filtered out by the Kalman filter algorithm. The Kalman filter first predicts the instantaneous value of the signal using the state estimation equation, and then updates the prediction result by combining the observed value, achieving noise suppression by minimizing the difference between the predicted value and the observed value. After denoising, the eigenvalue is adjusted to a unified dimension range through the mean normalization method. Finally, the preprocessed input feature information is obtained.

[0028] S20: Unfold the preprocessed input feature information according to a time window to generate a time series feature matrix containing the historical length.

[0029] Specifically, the preprocessed input feature information is divided into time windows of a fixed length. For example, the time window is set to 10 seconds. Successive eigenvalue within 10 seconds are selected in chronological order to construct a feature subset. The data within each subset are arranged in sequence according to the time step to ensure the complete retention of time dimension information. To enhance the representativeness of the matrix, a sliding window process is performed on each time window, with a sliding step set to 1 second to generate overlapping subsets to make full use of the feature information. After all subsets are combined, a time series feature matrix is formed, where each column of the matrix represents the values of a feature at multiple time steps. To handle possible feature missing situations, a linear interpolation algorithm is used to complete the filling based on adjacent time step values, and finally a time series feature matrix containing the historical length is generated, obtaining the time series feature matrix.

[0030] S30: Analyze the time series feature matrix based on the stepwise regularization feature ranking method to form an optimized input feature matrix. Based on the optimized input feature matrix, a weighted multi-dimensional regression method is used to establish a linear model to obtain a linear prediction value.

[0031] Specifically, first, statistical analysis is performed on each feature in the time series feature matrix to calculate the mean, variance, and Pearson correlation coefficient of the feature with the target variable. The stepwise regularization feature ranking method is used to gradually eliminate features starting from those with low correlation to reduce the computational complexity of the model. At the same time, weights are assigned to the remaining features, and the weights are dynamically allocated according to the influence degree of the features on the target variable to form an optimized input feature matrix. Subsequently, a weighted multi-dimensional regression model is constructed using the optimized matrix, and the relationship between the feature weights and the target variable is fitted by the least squares method to calculate the weighted sum of the target variable. In each iteration, the feature weights and bias terms are updated to reduce the sum of squared residuals. After the iteration converges, the linear prediction value is output, obtaining the linear prediction value.

[0032] S40: Obtain the actual measurement value of the particle size detection device in the on-site working condition, calculate the error between the linear prediction value and the actual measurement value to obtain an error value. Based on the error value, use a dynamic weighted recurrent network for non-linear modeling to obtain a non-linear prediction value, and combine the linear prediction value and the non-linear prediction value according to weights to generate a preliminary soft sensor model.

[0033] Specifically, read the real-time output of the on-site particle size detection device as the actual measurement value, correspond it with the linear prediction value one by one according to the samples, calculate the difference between the two as the error value. The absolute value function is used for error value calculation to avoid positive and negative cancellation. At the same time, record the time step information of each error. Input the error value into the dynamic weighted recurrent network. The recurrent network consists of multiple hidden layers. Each hidden layer performs non-linear mapping on the error value through an activation function and calculates the state output at the current moment in combination with the output value at the previous moment. The weights of the recurrent network are dynamically optimized by the gradient descent method to reduce the output error, and finally a non-linear prediction value is obtained. Combine the linear prediction value and the non-linear prediction value linearly according to a preset weight formula. The weights are set based on feature importance and historical effects to obtain a preliminary soft sensor model.

[0034] S50: Globally optimize the combined weights of the preliminary soft sensor model through a multi-objective optimization algorithm to generate an optimized soft sensor model. Real-time monitor the statistical characteristics of the input feature information in the grinding production process. When the statistical characteristics exceed the set threshold, trigger the dynamic correction mechanism. According to the dynamic correction mechanism, update the parameters of the optimized soft sensor model through an incremental learning method to obtain an updated soft sensor model.

[0035] Specifically, construct a multi-objective optimization function. The first objective is to minimize the prediction error, defined as the mean square error between the predicted value and the actual value. The second objective is to control the weight sparsity, defined as the two-norm of the weight vector. Optimize the weight combination through a genetic algorithm. The genetic algorithm randomly generates an initial weight combination, calculates the fitness value of each combination based on the optimization objective function, selects the weight combination with a higher fitness value for crossover and mutation operations, and repeats the iteration until the objective function value converges. Substitute the optimized weights into the preliminary soft sensor model to generate an optimized soft sensor model. At the same time, analyze the mean and variance of the input features in real time. When the change in statistical characteristics exceeds the threshold range, trigger the dynamic correction mechanism, and gradually adjust the parameters of the optimized soft sensor model based on the incremental learning method. Each adjustment updates the feature weights and model parameters of the newly added data at the current moment, so as to obtain an updated soft sensor model.

[0036] S60: Based on the updated soft sensor model, perform real-time prediction on the grinding particle size, output the particle size prediction value, and generate suggestions for adjusting the grinding process parameters according to the particle size prediction value.

[0037] Specifically, the input feature information collected in real time is input into the updated soft sensor model. The corresponding predicted values are calculated using the linear part and the nonlinear part of the model respectively, and integrated into a particle size predicted value according to the combined weights. After comparing the particle size predicted value with the target particle size range, grinding process parameter adjustment suggestions are generated based on the deviation. The adjustment suggestions include the optimized range of the mill speed, the adjustment ratio of the medium filling rate, and the optimized range of the feed particle size. The adjustment range is calculated by a discrete optimization algorithm combined with the deviation value, and finally the adjustment suggestions are visually presented on the operation panel.

[0038] In one embodiment, as Figure 2 shown, in step S30, that is, based on the optimized input feature matrix, a linear model is established using the weighted multi-dimensional regression method to obtain a linear predicted value, including: S301: The weighted multi-dimensional regression method establishes a linear model based on the following formula to obtain a linear predicted value: Where, Y linear is the linear predicted value, β 0 is the bias term of the linear model, n is the number of input features; w i is the weight of the i-th feature, β i is the regression coefficient of the i-th feature, x i is the value of the i-th feature.

[0039] Specifically, when establishing a linear model using the weighted multi-dimensional regression method, first, through a preliminary analysis of the input features, the importance coefficient w i of each feature is determined. The calculation of the importance coefficient is based on the correlation between the feature and the target variable, and the Pearson correlation coefficient is used as a measure. Features with high correlation are assigned larger weights w i , and at the same time, features with low correlation or redundancy are removed; subsequently, according to the optimized input feature matrix, the least squares method is used to determine the regression coefficient β i , specifically, a target function is constructed to minimize the sum of the squared errors between the predicted value Y linear and the actual value, and the regression coefficient β i of each feature is solved by taking the partial derivative of the target function. During the calculation process, the bias term β 0 is used to balance the overall deviation of the model. Finally, through iterative optimization, the weights w i and the regression coefficient β i are comprehensively applied to the regression formula to calculate the linear predicted value Y linear .

[0040] In one embodiment, as Figure 3 shown, in step S40, that is, based on the error value, a dynamic weighted recursive network is used for non-linear modeling to obtain a non-linear predicted value, including: S401: Take the error value as the modeling objective of the dynamic weighted recurrent network, and take the optimized input feature matrix as the input of the dynamic weighted recurrent network.

[0041] Specifically, define the error value as the optimization objective of the dynamic weighted recurrent network, and obtain the optimized input feature matrix. Assume that the number of features of the matrix is 5, and each column of features represents the mill speed, medium filling rate, ore hardness, feed particle size, and discharge concentration in sequence. Input these feature values into the input layer of the recurrent network in sequence. At the same time, initialize the weight of the input layer to [0.1, 0.2, -0.1, 0.3, 0.05], and the bias to zero. The initial signal calculates the weighted value of each column of features through matrix calculation. For example, for the first column of features (mill speed), when the input value is 150, the weighted signal is 150×0.1 = 15. This weighted value will be transmitted to the hidden layer for further processing, and finally complete the input initialization of the recurrent network.

[0042] S402: In the hidden layer of the dynamic weighted recurrent network, adjust the weights and biases of the hidden nodes according to the dynamic change of the error value. The dynamic adjustment of the weights and biases is based on the following formula: wh,i = wh, i-1 -α·Δw h,i-1 , where, w h,i and b h,i are the weights and biases of the i-th hidden node respectively, α is the adaptive step coefficient, Δw h,i-1 and Δb h,i-1 are the update amplitudes of the weights and biases caused by the error in the previous step respectively.

[0043] Specifically, the weights and biases of the hidden nodes are dynamically updated through the error backpropagation algorithm. First, obtain the current error value, and calculate the update amounts of the weights and biases through the chain rule. The size of the update amounts is determined by the gradient of the loss function and the learning rate. To avoid gradient explosion or disappearance, the learning rate is usually set to a small value, such as 0.01. Gradually reduce the error of the network through iterative updates. For example, the error value at the current moment is 2.5, the learning rate is set to α = 0.01, the initial weight is w h,1 = 0.2, the bias is b h,1 = 0.1. Calculate the adjustment amount of the weight as Δw h,1 = 2.5×0.5 = 1.25 according to the influence of the error on the node output. The updated weight is wh,1 = 0.2 - 0.01×1.25 = 0.1875. The bias is adjusted in a similar way, and the new bias value is calculated as b h,1 = 0.1 - 0.01×0.8 = 0.092. The updated weights and biases are used for the next iterative calculation, gradually reducing the error and optimizing the node output of the hidden layer.

[0044] S403: In the hidden layer of the dynamic weighted recurrent network, the output of each node is recursively calculated through a recurrent activation function, and the recurrent activation function is represented by the following formula: h i = g(w h,i ·x i + b h,i + γ·h i-1 ), where hi is the output of the i-th hidden node, xi is the input feature value, hi-1 is the recurrent output of the previous node, and γ is the recurrent gain coefficient.

[0045] Specifically, the recurrent activation function combines the weighted signal of the input feature with the recurrent output of the previous moment to generate the output of the current node. To achieve this process, the recurrent gain coefficient is used to adjust the influence intensity of the recurrent signal. For example, the Sigmoid activation function can be selected to normalize the output value between [0, 1], making the gradient distribution of the recurrent signal smoother during backpropagation. The output of the current moment depends on the superposition relationship of the recurrent output of the previous moment. Combining the non-linear mapping of the activation function, the recurrent process can capture the time series characteristics of the input feature, thereby enhancing the network's adaptability to dynamic working conditions. For example, assume the input feature value is x 1 = 200, the recurrent output of the previous moment is h 0 = 0.3, the recurrent gain coefficient is γ = 0.8, the activation function uses the Sigmoid function, the input signal of the node is z = (200×0.1 + 0.8×0.3 + 0.05) = 20.29, the output of the activation function is g(z)≈1, and the recurrent output value is 1. This value is passed to the next layer of nodes for continued recursive calculation.

[0046] S404: In the output layer of the dynamic weighted recurrent network, the non-linear prediction value is obtained through the following calculation formula: where Y nonlinear is the non-linear prediction value, φi is the output weight of the i-th hidden node, and h i is the output value of the i-th hidden node.

[0047] Specifically, all the node signals output by the hidden layer are weighted and accumulated to form the final non-linear prediction value. To improve the accuracy of the prediction value, the weights of the output layer are optimized by the gradient descent method. The optimization process adjusts the weights in each iteration to gradually reduce the sum of squared errors between the prediction value and the target value. For example, at the initial stage of the network, the output layer weights are randomly initialized to enhance diversity, and then the weight parameters are optimized in combination with the error value and the strength of the activation signal. After multiple iterations, the finally output prediction value can effectively capture the non-linear relationship between the input features and the target variable. For example, at the output layer, the node outputs of the hidden layer are aggregated and weighted to generate the non-linear prediction value. For example, assume that the hidden layer has 3 nodes, and their output values are h 1 = 0.8, h 2 = 0.5, h 3 = 0.7, and the corresponding output weights are φ 1 = 0.3, φ 2 = 0.4, φ 3 = 0.3. The non-linear prediction value is calculated as Y nonlinear = (0.8 × 0.3) + (0.5 × 0.4) + (0.7 × 0.3) = 0.24 + 0.2 + 0.21 = 0.65. The finally output non-linear prediction value is 0.65, which can be used in combination with the linear prediction value to generate a soft sensor model.

[0048] In one embodiment, as Figure 4 shown, in step S40, the linear prediction value and the non-linear prediction value are combined by weights to generate a preliminary soft sensor model, including: S405: Obtain the preliminary soft sensor model through the following formula: M initial = f linear (X)·λ 1 + f nonlinear (X,E)·λ 2 , where Minitial is the preliminary soft sensor model, flinear(X) represents the mapping of the linear model to the input feature matrix, fnonlinear(X,E) represents the mapping of the non-linear model based on the input feature matrix and the error value, and λ1 and λ2 are the weight coefficients of the linear model and the non-linear model respectively, and satisfy λ1 + λ2 = 1.

[0049] Specifically, the input feature matrix X and the error value E are mapped through the linear model and the non-linear model respectively. First, f linear (X) is calculated using the linear model. Specifically, each input feature value is weighted by w i and the regression coefficient β iPerform weighted summation. For example, when the input feature matrix X contains three eigenvalues [100, 200, 150], and the corresponding weights are [0.1, 0.2, 0.3], [0.5, 0.4, 0.6], then f linear (X) = (100 × 0.1 × 0.5) + (200 × 0.2 × 0.4) + (150 × 0.3 × 0.6) = 50.4. Then, use the non-linear model f nonlinear (X, E) to generate a non-linear mapping result by combining the input feature matrix and the error value. The non-linear model is mapped through a recursive network. Assume the error value E = 10, and the output result after the non-linear transformation of the input feature through the recursive network is 40, then f nonlinear (X, E) = 40. Subsequently, according to the formula M initial = f linear (X)·λ 1 + f nonlinear (X,E)·λ 2 Set the weight coefficients λ 1 = 0.6 and λ 2 = 0.4, substitute the above linear and non-linear results, and calculate M initial = (0.6 × 50.4) + (0.4 × 40) = 30.24 + 16 = 46.24. Through this weighted combination, the prediction results of the linear model and the non-linear model are integrated into a preliminary soft sensor model, and the finally generated M initial is 46.24, which serves as the basis for subsequent optimization.

[0050] In one embodiment, as Figure 5 shown, in step S50, that is, through a multi-objective optimization algorithm, the combined weights of the preliminary soft sensor model are globally optimized to generate an optimized soft sensor model, including: S501: Obtain the optimized soft sensor model through the following soft sensor model calculation formula: Where, M optimized is the optimized soft sensor model, M k represents the output of the k-th sub-model, and λ k opt is the optimized weight of the k-th sub-model.

[0051] Specifically, first, obtain the output values of each sub-model, perform normalization processing on these output values to ensure that all sub-model outputs are within the same numerical range, so as to eliminate the influence caused by the dimensional differences of the output values. Then, perform a linear weighted combination of the outputs of each sub-model according to the initial weights to calculate the preliminary soft sensor model. Evaluate the effect of the current weights through the objective function, and gradually adjust the weights according to the evaluation results. The adjusted weights need to meet the constraint condition, that is, the sum of all weights is equal to 1. Finally, substitute the optimized weights into the weighted combination formula to recalculate the optimized soft sensor model.

[0052] S502: λ k opt The solution of is based on the following objective function: Where, J(λ) is the objective function value, Z l is the actual particle size measurement value of the sample, q is the number of samples used in the optimization process, β is the adjustment coefficient, and λ k is the k-th initial weight. By iteratively optimizing J(λ), adjust λ k , to obtain λ k opt , substitute λ k opt into the soft sensor model calculation formula to obtain the optimized soft sensor model.

[0053] Specifically, construct an objective function. The objective function includes two parts. The first part is the sum of the squared mean errors between the model prediction value and the actual particle size measurement value, representing the prediction accuracy of the current weight combination. The second part is the regularization term for weight sparsification, which is used to balance the complexity of the weight distribution. Adjust the relative importance of the two parts through the adjustment coefficient. When optimizing the objective function, use an iterative optimization algorithm. Start from the initial weights and gradually adjust. Determine the adjustment direction and amplitude by calculating the gradient of the objective function value with respect to each weight. After each iteration, recalculate the objective function value according to the new weight combination and judge whether the optimization convergence condition is met. Stop the iteration when the objective function value no longer decreases significantly. Take the finally optimized weights as λ k opt , substitute it into the calculation formula of the soft sensor model to generate the optimized soft sensor model.

[0054] In one embodiment, as Figure 6 shown, in step S50, that is, real-time monitor the statistical characteristics of the input feature information during the grinding production process. When the statistical characteristics exceed the set threshold, trigger the dynamic correction mechanism, including: S503: Real-time monitor the statistical characteristics of the input feature information. The statistical characteristics include the mean value, variance, and distribution form.

[0055] Specifically, by collecting input feature information in real time, statistical characteristics of each feature value are calculated, including the mean, variance, and distribution pattern. The mean is calculated using a sliding window method, where all feature values within each time window are summed and then averaged to reflect the central tendency of the feature information in real time. The variance is obtained by calculating the sum of the squares of the deviations of each feature value from its mean and then taking the average to evaluate the fluctuation of the feature information. The distribution pattern is presented by constructing a frequency distribution diagram of the feature values or generating a distribution curve using the kernel density estimation method to show the distribution law of the feature information. To ensure the real-time calculation of statistical characteristics, each time the mean, variance, and distribution pattern are updated, only the latest information within the current time window is used, and expired information is ignored, thereby accurately capturing the dynamic change characteristics of the feature information.

[0056] S504: Compare the mean, variance, and distribution pattern with the historical reference benchmarks respectively. When the change amplitude of any one of the mean, variance, or distribution pattern exceeds the set threshold, generate a distribution change warning message. After generating the distribution change warning message, activate the dynamic correction mechanism.

[0057] Specifically, after real-time monitoring of the statistical characteristics, the mean, variance, and distribution pattern calculated within the current time window are compared with the historical reference benchmarks one by one. The historical reference benchmarks can be calculated from information samples with long-term stable operation, including the mean benchmark, variance benchmark, and stable distribution curve of the feature values. The change amplitude is calculated based on the deviation percentage formula. For example, for the change in the mean, the change amplitude is determined by calculating the ratio of the difference between the current mean and the benchmark mean to the benchmark mean. For the change in the variance, a similar method is used. For the change in the distribution pattern, a distribution distance metric method is adopted, such as calculating the similarity between the current distribution and the benchmark distribution using the Kullback-Leibler (KL) divergence. When the change amplitude of any one item exceeds the preset threshold, immediately trigger a distribution change warning message. The warning message includes the feature name, change type, and change amplitude that trigger the change. After the warning is generated, activate the dynamic correction mechanism to adapt to the distribution change of the current input features.

[0058] In one embodiment, as Figure 7 shown, in step S50, according to the dynamic correction mechanism, the parameters of the optimized soft sensor model are updated by the incremental learning method to obtain the updated soft sensor model, including: S505: According to the dynamic correction mechanism, obtain the new information at the current moment.

[0059] Specifically, after the dynamic correction mechanism is triggered, by real-time monitoring the input feature data collected by various sensors in the grinding production process, the new information at the current moment is obtained. The new information includes the latest data in the input feature matrix within the current time window and the actual measurement value of the particle size detection device. For the newly added input feature data, first, data integrity checks are performed on it to identify whether there are missing values or outliers. The missing values are filled using the interpolation method of adjacent time steps, and the outliers are replaced by setting reasonable upper and lower threshold values to ensure the validity of the data. At the same time, preprocessing is performed on the newly added data, including normalization operations, which map all feature values to a unified numerical range to avoid the influence caused by differences in different feature dimensions. After completing the data checks and preprocessing, the new information is used as the input data for the dynamic correction mechanism, ready to be used for subsequent parameter adjustment and optimization of the model.

[0060] S506: Based on the new information at the current moment, adjust the regression coefficients and bias terms of the linear model through the following formula: where, β updated，i is the updated regression coefficient, β previous，i is the regression coefficient before update, L linear is the loss function of the linear model calculated based on the new information, η is the learning rate, β updated，0 is the updated bias term, β previous,0 is the bias term before update.

[0061] Specifically, using the input feature data and target values in the new information, first calculate the loss function value of the linear model. The loss function adopts the mean square error form, which sums the squares of the differences between the actual target values of the new data and the predicted values of the linear model and then takes the average to reflect the current prediction error of the model. For the adjustment of the regression coefficients and bias terms, the partial derivatives are calculated for each parameter based on the loss function to determine the direction and magnitude of the parameter adjustment. The calculation of the partial derivatives is completed through the chain rule. The specific steps are to expand and decompose the error contribution of each term in the loss function and calculate the partial derivative values in combination with the influence weights of each feature value. Subsequently, the gradient descent method is used to update the regression coefficients and bias terms. Each iteration determines the adjustment step size according to the learning rate η. The setting of the learning rate needs to balance the convergence speed and model stability. Usually, a small positive value is taken to avoid over-adjustment. The updated regression coefficient β updated,i and bias term β are fed back into the linear model to recalculate the predicted values to evaluate the adjustment effect. The adjustment process continues until the decrease in the loss function value is less than the set threshold or the maximum number of iterations is reached. Finally, the optimized regression coefficients and bias terms are generated for subsequent prediction of the linear model.

[0062] S507: Adjust the network parameters of the non - linear model according to the following formula based on the newly added information at the current moment: where, φ updated is the updated network parameter, φ previous is the network parameter before update, L nonlinear is the loss function of the non - linear model calculated based on the newly added information, and η is the learning rate.

[0063] Specifically, the input feature matrix and the target value in the newly added information are input into the network structure of the non - linear model. The network parameters include the weight matrix and the bias term, which define the non - linear mapping ability of the model. First, calculate the output value of the network through forward propagation. The input feature matrix is passed through each layer of the network in turn. The calculation of each layer of the network includes the weighted accumulation of the input value and the weight matrix, the linear superposition of the bias term, and the non - linear transformation of the activation function. The activation function usually selects ReLU or Sigmoid to enhance the network's ability to express non - linear features. Subsequently, calculate the loss function value by comparing the network output obtained through forward propagation with the actual target value in the newly added information. The loss function uses the mean square error or the cross - entropy function, specifically depending on the distribution form of the target value. Use the backpropagation algorithm to calculate the partial derivative of the loss function with respect to each network parameter. During the calculation of the partial derivative, perform chain - rule differentiation on the parameters from the output layer to the input layer in turn, layer - by - layer expanding the dependence relationship of the loss function on the network parameters. Combine the current gradient information and use the gradient descent method to update each parameter. When updating the parameters, each weight matrix or bias term is adjusted according to the following rules: first calculate the update amount according to the current gradient information. The size of the update amount is determined by the learning rate η. The learning rate is generally set to a small positive value to ensure the smoothness of the update. Subtract the calculated update amount from the current parameter value to generate the new network parameter. The updated parameters are reapplied to the network structure for the next round of forward propagation and backpropagation. The parameter adjustment process continues until the decrease amplitude of the loss function value is less than the set threshold or the maximum number of iterations is reached, and finally the updated network parameters are obtained.

[0064] S508: Update the weight coefficient of the optimized soft - sensor model according to the following formula based on the newly added information at the current moment: where, λ updated,k is the updated weight coefficient, λ previous,k is the k - th weight coefficient before update, L comb is the combined error loss function of the newly added information in the optimized soft - sensor model, and η is the learning rate.

[0065] Specifically, first, the input feature matrix and the target value within the current time window are extracted from the newly added information. The input feature matrix is passed to the sub-models of the optimized soft sensor model, and the output values of each sub-model are calculated. The sub-model output values are weighted and combined according to the current weight coefficient λ previous,k to generate the predicted value of the current soft sensor model. Subsequently, the predicted value is compared with the target value, and the combined error loss function L comb is calculated using the difference. The loss function usually adopts the mean square error form to measure the overall prediction deviation of the soft sensor model. By taking the derivative of the loss function L comb , the contribution of each weight coefficient to the error is calculated. Specifically, the partial derivative is calculated for each λk, and this partial derivative reflects the direction and degree of the impact of adjusting the weight on reducing the loss. Combining the calculated partial derivatives, the gradient descent method is used to update each weight coefficient. During the update, the weight adjustment amount is controlled by the learning rate η, and the learning rate is set to a small positive value to ensure the smoothness and convergence of the update process. After each update, the combined error value is recalculated to evaluate the adjustment effect. During the adjustment process, it is necessary to ensure that the weight coefficients satisfy the constraint condition, that is, the sum of all weight coefficients is equal to 1. Through the normalization operation, the weight coefficients are normalized after each update to ensure that the constraint condition always holds. The update process is iterated until the change amplitude of the loss function value is lower than the set threshold or the maximum number of iterations is reached, and finally the updated weight coefficients are generated.

[0066] S509: Generate an updated soft sensor model based on the updated regression coefficients, updated bias terms, updated network parameters, and updated weight coefficients through the following formula: The formula for the updated linear predicted value is The formula for the updated non-linear predicted value is, Y nonlinear,updated = f(X new , φ updated ), where Xi is the input feature matrix collected in real time, Xnew is the original value of the i-th feature in the current feature set, and the combined formula for the updated soft sensor model is, M updated = λ updated,1 ·Y linear,updated + λ updated,2 ·Y nonlinear,updated , M up d ated is the updated soft sensor model, and λupdated,1 and λupdated,2 are the weights of the updated linear model and non-linear model, respectively.

[0067] Specifically, according to the input feature matrix collected in real time, the linear prediction value is calculated by using the updated regression coefficients and bias terms. Multiply each input feature value by the corresponding updated regression coefficient and sum them up, and then add the updated bias term to generate the linear prediction value. Then, input the newly added feature set at the current moment into the updated non-linear model, adjust the weights and biases of the network layer through the updated network parameters, and calculate the non-linear mapping result through the recursive activation function to obtain the non-linear prediction value. Then, combine the calculated linear prediction value and non-linear prediction value with the updated linear model weights and non-linear model weights, and perform weighted combination on the two prediction values according to the weight distribution ratio to generate the final updated soft sensor model.

[0068] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not imply the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0069] In one embodiment, a dynamic grinding particle size soft sensor device is provided, which corresponds to the grinding particle size soft sensor method in the above embodiment one by one. As Figure 8 shown, the dynamic grinding particle size soft sensor device includes an input feature preprocessing module, a feature optimization and linear modeling module, a non-linear modeling module, a preliminary soft sensor modeling module, a dynamic correction and incremental learning module, and a particle size prediction and process adjustment suggestion module. The detailed description of each functional module is as follows: The input feature preprocessing module is used to obtain the input feature information in the grinding production process, perform denoising processing on the input feature information, and obtain the preprocessed input feature information; The feature optimization and linear modeling module is used to expand the preprocessed input feature information according to a time window to generate a time series feature matrix including a historical length; The non-linear modeling module is used to analyze the time series feature matrix based on the stepwise regularization feature ranking method to form an optimized input feature matrix. Based on the optimized input feature matrix, a linear model is established by using the weighted multi-dimensional regression method to obtain the linear prediction value; The preliminary soft sensor modeling module is used to obtain the actual measurement value of the particle size detection device in the on-site working condition, calculate the error between the linear prediction value and the actual measurement value to obtain the error value. Based on the error value, a non-linear model is established by using a dynamic weighted recursive network to obtain the non-linear prediction value, and the linear prediction value and the non-linear prediction value are combined according to the weight to generate a preliminary soft sensor model; The dynamic correction and incremental learning module is used to globally optimize the combined weights of the preliminary soft sensor model through a multi-objective optimization algorithm, generate an optimized soft sensor model, and monitor the statistical characteristics of the input feature information in the grinding production process in real time. When the statistical characteristics exceed the set threshold, it triggers the dynamic correction mechanism. According to the dynamic correction mechanism, the parameters of the optimized soft sensor model are updated through an incremental learning method to obtain an updated soft sensor model; The particle size prediction and process adjustment suggestion module is used to predict the grinding particle size in real time based on the updated soft sensor model, output the particle size prediction value, and generate grinding process parameter adjustment suggestions according to the particle size prediction value.

[0070] Optionally, the non-linear modeling module includes: The linear prediction value sub-module is used to establish a linear model based on the weighted multi-dimensional regression method according to the following formula to obtain a linear prediction value: where, Y linear is the linear prediction value, β 0 is the bias term of the linear model, n is the number of input features; w i is the weight of the i-th feature, β i is the regression coefficient of the i-th feature, x i is the value of the i-th feature.

[0071] Optionally, the preliminary soft sensor modeling module includes: The error-driven input sub-module is used to use the error value as the modeling target of the dynamic weighted recursive network and use the optimized input feature matrix as the input of the dynamic weighted recursive network; The dynamic weight and bias adjustment sub-module is used to adjust the weights and biases of the hidden nodes according to the dynamic changes of the error value in the hidden layer of the dynamic weighted recursive network. The dynamic adjustment of the weights and biases is based on the following formula: wh,i=wh, i-1 -α·Δw h,i-1 , where, w h,i and b h,i are the weights and biases of the i-th hidden node respectively, α is the adaptive step coefficient, Δw h,i-1 and Δb h,i-1 are the weight and bias update amplitudes caused by the previous error respectively; The recursive activation calculation sub-module is used to recursively calculate the output of each node through a recursive activation function in the hidden layer of the dynamic weighted recursive network. The recursive activation function is represented by the following formula: h i =g(w h,i ·x i +b h,i +γ·h i-1), where \(h_i\) is the output of the \(i\)-th hidden node, and \(x\) i is the input feature value, and \(h\) i-1 is the recursive output of the previous node, and \(\gamma\) is the recursive gain coefficient; The non - linear prediction generation sub - module is used to obtain the non - linear prediction value at the output layer of the dynamic weighted recursive network through the following calculation formula: where \(Y\) nonlinear is the non - linear prediction value, \(\varphi_i\) is the output weight of the \(i\)-th hidden node, and \(h\) i is the output value of the \(i\)-th hidden node.

[0072] Optionally, the preliminary soft - sensor modeling module further includes: The preliminary soft - sensor model generation sub - module is used to obtain the preliminary soft - sensor model through the following formula: \(M\) initial \(= f\) linear (X)\(\cdot\lambda\) 1 \(+ f\) nonlinear (X, E)\(\cdot\lambda\) 2 , where \(M\) initial is the preliminary soft - sensor model, \(f\) linear (X) represents the mapping of the linear model to the input feature matrix, \(f\) nonlinear (X, E) represents the mapping of the non - linear model based on the input feature matrix and the error value, and \(\lambda\) 1 and \(\lambda\) 2 are the weight coefficients of the linear model and the non - linear model respectively, and satisfy \(\lambda_1+\lambda_2 = 1\).

[0073] Optionally, the dynamic correction and incremental learning module includes: The soft - sensor model weight optimization sub - module is used to obtain the optimized soft - sensor model through the following soft - sensor model calculation formula: where \(M\) optimized is the optimized soft - sensor model, \(M\) k represents the output of the \(k\)-th sub - model, and \(\lambda\) k opt is the optimized weight of the \(k\)-th sub - model; The optimized soft - sensor model generation sub - module is used to solve \(\lambda\) k opt based on the following objective function: where \(J(\lambda)\) is the objective function value, \(Z\) l is the actual particle size measurement value of the sample, \(q\) is the number of samples used in the optimization process, \(\beta\) is the adjustment coefficient, and \(\lambda\) k is the \(k\)-th initial weight. By iteratively optimizing \(J(\lambda)\) and adjusting \(\lambda\) k , \(\lambda\) kopt , substitute λ k opt into the calculation formula of the soft sensor model to obtain the optimized soft sensor model.

[0074] Optionally, the dynamic correction and incremental learning module further includes: A statistical characteristic real-time monitoring sub-module for real-time monitoring of the statistical characteristics of the input feature information, where the statistical characteristics include mean, variance, and distribution form; A distribution change alarm trigger sub-module for comparing the mean, variance, and distribution form with the historical reference benchmark respectively. When the change amplitude of any one of the mean, variance, or distribution form exceeds the set threshold, a distribution change alarm message is generated. After generating the distribution change alarm message, the dynamic correction mechanism is started.

[0075] Optionally, the dynamic correction and incremental learning module further includes: A new information acquisition sub-module for acquiring the new information at the current moment according to the dynamic correction mechanism; A linear model parameter adjustment sub-module for adjusting the regression coefficient and bias term of the linear model based on the new information at the current moment through the following formula: where, β updated，i is the updated regression coefficient, β previous，i is the regression coefficient before update, L linear is the loss function of the linear model calculated based on the new information, η is the learning rate, β updated，0 is the updated bias term, β previous,0 is the bias term before update; A non-linear model parameter adjustment sub-module for adjusting the network parameters of the non-linear model based on the new information at the current moment through the following formula: where, φ updated is the updated network parameter, φ previous is the network parameter before update, L nonlinear is the loss function of the non-linear model calculated based on the new information, η is the learning rate; A weight optimization and update sub-module for updating the weight coefficients of the optimized soft sensor model based on the new information at the current moment through the following formula: where, λ updated,k is the updated weight coefficient, λ previous,k is the k-th weight coefficient before update, L combThe combined error loss function for the new information in the optimized soft sensor model, where η is the learning rate; the soft sensor model generation sub-module is used to generate an updated soft sensor model based on the updated regression coefficients, updated bias terms, updated network parameters, and updated weight coefficients through the following formula: The formula for the updated linear prediction value is The formula for the updated non-linear prediction value is, Y nonlinear,updated = f(X new , φ updated ), where Xi is the input feature matrix collected in real time, Xnew is the original value of the i-th feature in the current feature set, and the combined formula for the updated soft sensor model is, M updated = λ updated,1 ·Y linear,updated + λ updated,2 ·Y nonlinear,updated , M updated is the updated soft sensor model, and λupdated,1 and λupdated,2 are the updated linear model weight and non-linear model weight respectively.

[0076] For the specific limitations of the dynamic grinding particle size soft sensor device, reference can be made to the limitations of the grinding particle size soft measurement method in the above text, which will not be elaborated here. Each module in the above dynamic grinding particle size soft sensor device can be implemented in whole or in part by software, hardware, and their combination. Each of the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so as to facilitate the processor to call and execute the operations corresponding to each of the above modules.

[0077] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 9 shown. The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program, when executed by the processor, implements a grinding particle size soft measurement method.

[0078] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the following steps are implemented: Obtain the input feature information in the grinding production process, perform denoising processing on the input feature information, and obtain the preprocessed input feature information; Expand the preprocessed input feature information according to a time window to generate a time series feature matrix including the historical length; Analyze the time series feature matrix based on the stepwise regularization feature ranking method to form an optimized input feature matrix. Based on the optimized input feature matrix, establish a linear model using the weighted multi-dimensional regression method to obtain a linear prediction value; Obtain the actual measurement value of the particle size detection device in the on-site working condition, calculate the error between the linear prediction value and the actual measurement value to obtain an error value. Based on the error value, perform non-linear modeling using a dynamic weighted recursive network to obtain a non-linear prediction value, and combine the linear prediction value and the non-linear prediction value according to weights to generate a preliminary soft sensor model; Globally optimize the combined weights of the preliminary soft sensor model through a multi-objective optimization algorithm to generate an optimized soft sensor model. Real-time monitor the statistical characteristics of the input feature information in the grinding production process. When the statistical characteristics exceed the set threshold, trigger a dynamic correction mechanism. According to the dynamic correction mechanism, update the parameters of the optimized soft sensor model through an incremental learning method to obtain an updated soft sensor model; Based on the updated soft sensor model, perform real-time prediction on the grinding particle size, output the particle size prediction value, and generate suggestions for adjusting the grinding process parameters according to the particle size prediction value.

[0079] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented: Obtain the input feature information in the grinding production process, perform denoising processing on the input feature information, and obtain the preprocessed input feature information; Expand the preprocessed input feature information according to a time window to generate a time series feature matrix including the historical length; Analyze the time series feature matrix based on the stepwise regularization feature ranking method to form an optimized input feature matrix. Based on the optimized input feature matrix, establish a linear model using the weighted multi-dimensional regression method to obtain a linear prediction value; Obtain the actual measurement value of the particle size detection device in the on-site working condition, calculate the error between the linear prediction value and the actual measurement value to obtain an error value. Based on the error value, perform non-linear modeling using a dynamic weighted recursive network to obtain a non-linear prediction value, and combine the linear prediction value and the non-linear prediction value according to weights to generate a preliminary soft sensor model; The combined weights of the preliminary soft sensor model are globally optimized through a multi-objective optimization algorithm to generate an optimized soft sensor model, and the statistical characteristics of the input feature information during the grinding production process are monitored in real time. When the statistical characteristics exceed the set threshold, a dynamic correction mechanism is triggered. According to the dynamic correction mechanism, the parameters of the optimized soft sensor model are updated through an incremental learning method to obtain an updated soft sensor model; Based on the updated soft sensor model, the grinding particle size is predicted in real time, and the predicted particle size value is output. Adjustment suggestions for the grinding process parameters are generated according to the predicted particle size value.

[0080] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0081] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above-mentioned division of each functional unit and module is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0082] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A soft measurement method for grinding particle size, characterized in that: The grinding particle size soft measurement method comprises: Acquiring input feature information in a grinding production process, and performing denoising processing on the input feature information to obtain preprocessed input feature information; Expanding the preprocessed input feature information according to the time window to generate a time series feature matrix including the history length; The time series feature matrix is ​​analyzed based on a stepwise regularization feature sorting method to form an optimized input feature matrix, and a linear model is established based on the optimized input feature matrix using a weighted multidimensional regression method to obtain a linear prediction value; Acquire the actual measurement value of the particle size detection device in the field working condition, perform error calculation between the linear prediction value and the actual measurement value to obtain an error value, perform nonlinear modeling based on the error value using a dynamic weighted recursive network to obtain a nonlinear prediction value, combine the linear prediction value and the nonlinear prediction value according to the weight, and generate a preliminary soft measurement model; The combined weights of the preliminary soft sensor model are globally optimized by a multi-objective optimization algorithm to generate an optimized soft sensor model, and the statistical characteristics of the input feature information in the grinding production process are monitored in real time. When the statistical characteristics exceed a set threshold, a dynamic correction mechanism is triggered. According to the dynamic correction mechanism, the parameters of the optimized soft sensor model are updated by an incremental learning method to obtain an updated soft sensor model; Based on the updated soft sensor model, the grinding particle size is predicted in real time, the particle size prediction value is output, and the grinding process parameter adjustment suggestion is generated according to the particle size prediction value.

2. The grinding particle size soft measurement method according to claim 1 is characterized in that: The method of establishing a linear model based on the optimized input feature matrix and obtaining a linear prediction value by using a weighted multidimensional regression method includes: The weighted multidimensional regression method establishes the linear model based on the following formula to obtain the linear prediction value: Among them, the Y linear is the linear prediction value, β0 is the bias term of the linear model, n is the number of input features; w i is the weight of the i-th feature, the β i is the regression coefficient of the i-th feature, the x i is the value of the ith feature.

3. The grinding particle size soft measurement method according to claim 1 is characterized in that: The method of performing nonlinear modeling based on the error value using a dynamic weighted recursive network to obtain a nonlinear prediction value includes: Using the error value as a modeling target of the dynamic weighted recurrent network, and using the optimized input feature matrix as an input of the dynamic weighted recurrent network; In the hidden layer of the dynamic weighted recursive network, the weights and biases of the hidden nodes are adjusted according to the dynamic changes of the error value, and the dynamic adjustment of the weights and biases is based on the following formula: wh,i=wh, i-1 -α·Δw h,i-1 , Among them, the w h,i and the b h,i are the weight and bias of the ith hidden node, α is the adaptive step coefficient, and Δw h,i-1 and the Δb h,i-1 They are the weight and bias update amplitudes caused by the error in the previous step respectively; In the hidden layer of the dynamic weighted recursive network, the output of each node is recursively calculated by a recursive activation function, and the recursive activation function is expressed by the following formula: h i =g(w h,i ·x i +b h,i +γ·h i-1 ), where hi is the output of the i-th hidden node, xi is the input eigenvalue, hi-1 is the recursive output of the previous node, and γ is the recursive gain coefficient; At the output layer of the dynamic weighted recursive network, the nonlinear prediction value is obtained by the following calculation formula: Wherein Y nonlinear is the nonlinear prediction value, φi is the output weight of the i-th hidden node, h i is the output value of the i-th hidden node.

4. The grinding particle size soft measurement method according to claim 1 is characterized in that: The step of combining the linear prediction value and the nonlinear prediction value according to weights to generate a preliminary soft sensor model comprises: The preliminary soft-sensing model is obtained by the following formula: initial =f linear (X)·λ1+f nonlinear (X, E)·λ2, where the M initial is the preliminary soft sensor model, f linear (X) represents the mapping of the linear model to the input feature matrix, f nonlinear (X, E) represents the mapping of the nonlinear model based on the input feature matrix and the error value, and the λ1 and the λ2 are weight coefficients of the linear model and the nonlinear model respectively, and satisfy λ1+λ2=1.

5. The grinding particle size soft measurement method according to claim 1 is characterized in that: The method of globally optimizing the combined weights of the preliminary soft sensor model by a multi-objective optimization algorithm to generate an optimized soft sensor model comprises: The optimized soft sensor model is obtained by the following soft sensor model calculation formula: Among them, the M optimized is the optimized soft sensor model, M k represents the output of the kth sub-model, λ k opt is the optimization weight of the kth sub-model; The lambda k opt The solution is based on the following objective function: Wherein, J(λ) is the objective function value, Z l is the actual particle size measurement value of the sample, q is the number of samples used in the optimization process, β is the adjustment coefficient, and the λ k is the kth initial weight, and J(λ) is adjusted by iterative optimization. k , we get the λ k opt , the λ k opt Substitute the soft sensor model calculation formula to obtain the optimized soft sensor model.

6. The grinding particle size soft measurement method according to claim 1, characterized in that: The real-time monitoring of the statistical characteristics of the input feature information in the grinding production process, when the statistical characteristics exceed a set threshold, triggering a dynamic correction mechanism, includes: real-time monitoring of the statistical characteristics of the input feature information, the statistical characteristics including mean, variance and distribution form; The mean, the variance and the distribution form are compared with historical reference benchmarks respectively. When the change amplitude of any one of the mean, the variance or the distribution form exceeds the set threshold, a distribution change alarm message is generated. After the distribution change alarm message is generated, the dynamic correction mechanism is activated.

7. The grinding particle size soft measurement method according to claim 1, characterized in that: The method of updating the parameters of the optimized soft sensor model by an incremental learning method according to the dynamic correction mechanism to obtain an updated soft sensor model includes: According to the dynamic correction mechanism, obtaining the newly added information at the current moment; Based on the newly added information at the current moment, the regression coefficient and bias term of the linear model are adjusted by the following formula: Among them, the β updated，i is the updated regression coefficient, the β previous，i is the regression coefficient before updating, the L linear is the linear model loss function calculated based on the newly added information, η is the learning rate, The β updated，0 is the updated bias term, the β previous,0 is the bias term before updating; Based on the newly added information at the current moment, the network parameters of the nonlinear model are adjusted by the following formula: Among them, the φ updated is the updated network parameter, the φ previous is the network parameter before updating, the L nonlinear is the nonlinear model loss function calculated based on the newly added information, and η is the learning rate; Based on the newly added information at the current moment, the weight coefficient of the optimized soft sensor model is updated by the following formula: Among them, λ updated,k is the updated weight coefficient, λ previous,k is the kth weight coefficient before update, L comb is the combined error loss function of the newly added information in the optimized soft sensor model, η is the learning rate; based on the updated regression coefficient, the updated bias term, the updated network parameter and the updated weight coefficient, the updated soft sensor model is generated by the following formula: The updated linear prediction value formula is: The updated nonlinear prediction value formula is: nonlinear,updated =f(X new ,φ updated ), where X i is the input feature matrix collected in real time, X new is the original value of the i-th feature in the current feature set, and the combination formula of the updated soft sensor model is: updated =λ updated,1 ·Y linear,updated +λ updated,2 ·Y nonlinear,updated , the M updated is the updated soft measurement model, and λupdated,1 and λupdated,2 are the updated linear model weight and nonlinear model weight respectively.

8. A dynamic grinding particle size soft measurement device, characterized in that: The dynamic grinding particle size soft measurement device comprises: An input feature preprocessing module is used to obtain input feature information in the grinding production process, perform denoising on the input feature information, and obtain preprocessed input feature information; A feature optimization and linear modeling module, used to expand the preprocessed input feature information according to a time window to generate a time series feature matrix including a history length; A nonlinear modeling module is used to analyze the time series feature matrix based on a stepwise regularization feature sorting method to form an optimized input feature matrix, and to establish a linear model based on the optimized input feature matrix using a weighted multidimensional regression method to obtain a linear prediction value; A preliminary soft measurement modeling module is used to obtain the actual measurement value of the particle size detection device in the field working condition, perform error calculation between the linear prediction value and the actual measurement value to obtain the error value, perform nonlinear modeling based on the error value using a dynamic weighted recursive network to obtain a nonlinear prediction value, and combine the linear prediction value and the nonlinear prediction value according to the weight to generate a preliminary soft measurement model; A dynamic correction and incremental learning module is used to globally optimize the combined weights of the preliminary soft sensor model through a multi-objective optimization algorithm to generate an optimized soft sensor model, monitor the statistical characteristics of the input feature information in the grinding production process in real time, and trigger a dynamic correction mechanism when the statistical characteristics exceed a set threshold. According to the dynamic correction mechanism, the parameters of the optimized soft sensor model are updated through an incremental learning method to obtain an updated soft sensor model; The particle size prediction and process adjustment suggestion module is used to make real-time predictions on the grinding particle size based on the updated soft measurement model, output the particle size prediction value, and generate grinding process parameter adjustment suggestions according to the particle size prediction value.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the grinding particle size soft measurement method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the grinding particle size soft measurement method according to any one of claims 1 to 7 are implemented.

Images