A soft measurement method for ball mill particle size based on a time-series large model

Abstract

This invention relates to the field of engineering process control technology and discloses a soft measurement method for ball mill particle size based on a time-series large model. A multi-feature fusion module for extracting multi-scale features is constructed based on Convld K3, Convld K5, and Convld K1. A soft measurement model is built by combining multi-head attention and a large language module with fixed weights. Field data of ball mill grinding is acquired and input into the soft measurement model. The soft measurement model generates a first feature and a query matrix. A key matrix and a value matrix are generated according to the fixed weights of the large language module. A second feature is generated based on the query matrix, key matrix, and value matrix, and fused with the first feature to obtain the fused feature. The particle size prediction result corresponding to the field data is then obtained through the large language module. This method solves the problems of existing soft measurement models, such as excessive reliance on large-scale sample data, insufficient modeling ability for complex nonlinear process parameters, and low prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a soft measurement method for ball mill particle size based on a time-series large model. Background Technology

[0002] As a key piece of equipment in industrial grinding processes, the particle size of the ball mill directly affects the grinding effect and process stability. In existing technologies, real-time online particle size detection suffers from high equipment costs, large response delays, and difficulties in data acquisition. Therefore, constructing soft measurement models based on other readily available process parameters has become a research hotspot in industry.

[0003] However, traditional soft measurement methods often require a large amount of sample data to train the model in order to effectively capture complex process characteristics and nonlinear relationships. In real-world scenarios where data is scarce, the prediction accuracy is unsatisfactory and the ball mill discharge particle size cannot be accurately measured. Summary of the Invention

[0004] This invention provides a soft measurement method for ball mill particle size based on a time-series large model, in order to solve the problems of existing soft measurement models being overly dependent on large-scale sample data, having insufficient modeling ability for complex nonlinear process parameters, and having low prediction accuracy.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] This invention provides a soft measurement method for ball mill particle size based on a time-series large model, comprising the following steps:

[0007] Step 1: Construct a multi-feature fusion module for extracting multi-scale features based on Convld K3, Convld K5 and Convld K1. Construct a soft measurement model based on the query matrix, key matrix and value matrix of multi-head attention combined with the large language module with fixed weights and the multi-feature fusion module.

[0008] Step 2: Obtain the field data of ball mill grinding and input it into the soft measurement model. The soft measurement model generates the first feature and query matrix based on the field data. The soft measurement model generates the key matrix and value matrix based on the fixed weight of the big language module. The second feature is generated based on the query matrix, key matrix and value matrix, and fused with the first feature to obtain the fused feature. Then, the granularity prediction result corresponding to the field data is obtained through the big language module.

[0009] By employing a large language module with fixed weights, the rich semantic knowledge accumulated by the large model during the pre-training stage can be fully utilized, achieving high prediction accuracy and stability even under limited sample data conditions, thereby effectively reducing the requirement for large data volume.

[0010] By utilizing the Transformer architecture with multi-head attention mechanism and large language modules, the local and global features in the input data can be fully modeled from multiple perspectives, effectively capturing the complex nonlinear and dynamic dependencies between ball mill process parameters, and improving the system's response capability to granularity changes under complex operating conditions.

[0011] Furthermore, the multi-feature fusion module includes an input layer, a Convld K3 layer, a Convld K5 layer, a Convld K1 layer, a split layer, a concat layer, and a first output layer;

[0012] The input layer is sequentially connected to the Convld K3 layer, the split layer, the cancat layer, the Convld K1 layer, and the first output layer. The split layer and the cancat layer are also connected through the Convld K5 layer.

[0013] Furthermore, the soft measurement model is constructed as follows: a first multi-feature fusion module is set in the query matrix generation branch of multi-head attention, the fixed weights of the large language model are used as the input of the key matrix and value matrix generation branch, a second multi-feature fusion module is set in the splicing fusion of multi-head attention, and the Transformer framework of the large language model is set in the stage of fusing feature extraction granularity prediction results.

[0014] By leveraging a multi-head attention mechanism to perform parallel computation on the query matrix, key matrix, and value matrix, the system captures local and global features from multiple perspectives, effectively handling the interactions between complex and nonlinear process parameters. Combined with the Transformer architecture's ability to extract long-range dependencies, the system achieves a balance between accuracy and computational efficiency in real-time online prediction, a requirement that is often difficult to meet simultaneously in existing technologies.

[0015] Furthermore, in step 2, after acquiring the field data, feature extraction is performed to obtain statistical features, and then the statistical features are input into the soft measurement model;

[0016] The statistical characteristics include central tendency characteristics, dispersion characteristics, distribution characteristics, and data distribution characteristics;

[0017] The central trend features include the mean and the median;

[0018] The dispersion characteristics include variance, standard deviation, root mean square error, and bias;

[0019] The distribution characteristics include the maximum value, minimum value, first quartile, second quartile, third quartile, kurtosis, skewness, and percentile.

[0020] The data distribution characteristics include the autocorrelation coefficient.

[0021] Furthermore, the soft measurement model specifically includes data branches, multi-head branches, and output branches based on large language modules;

[0022] The data branch generates a first feature based on the statistical characteristics of the field data;

[0023] The multi-head branch generates a query matrix based on the statistical characteristics of the field data, generates a key matrix and a value matrix according to the fixed weight of the large language module, and then generates a second feature based on the query matrix, key matrix and value matrix;

[0024] The output branch obtains the first feature and the second feature and fuses them to obtain the fused feature. The long-distance dependency of the fused feature is extracted by the large language module to generate the granular prediction result corresponding to the field data.

[0025] Furthermore, the data branch sequentially includes a statistical feature layer, a prompt layer, a tokenizer layer, an embedding layer, and a linear layer;

[0026] The statistical feature layer transmits statistical features based on the central trend features, dispersion features, distribution features, and data distribution features of the field data to the prompt layer. The prompt layer converts the statistical features into prompt words and transmits them to the tokenizer layer for word segmentation. Then, the first feature is generated through vector transformation in the embedding layer and linear processing in the linear layer.

[0027] By comprehensively extracting various statistical features, this data is transformed into prompt words, which are then processed by a tokenizer to form the model input. This approach, which combines statistical features with prompt word construction, allows the model to incorporate prior knowledge during the feature representation stage, significantly outperforming traditional methods that rely solely on numerical data input.

[0028] Furthermore, the multi-head branch includes a normalization layer, an unfold layer, a first multi-feature fusion module, a first linear layer, a second linear layer, an MHA layer, and a second feature fusion module;

[0029] The normalization layer normalizes the data and transmits it to the unfold layer. The unfold layer extracts sliding window features from the normalized data. The first multi-feature fusion module performs feature transformation based on the sliding window features to generate a query matrix.

[0030] The first linear layer generates a key matrix based on the fixed weights of the large language module, and the second linear layer generates a value matrix based on the fixed weights of the large language module.

[0031] The MHA layer performs multi-head attention calculations based on the query matrix, key matrix, and value matrix, and then combines these with the second feature fusion module to generate the second feature.

[0032] Furthermore, the output branch is based on a large language module, including a concat layer, a large language module, a third linear layer, and a second output layer;

[0033] The concat layer, the large language module, the third linear layer, and the second output layer are connected in sequence.

[0034] The concat layer obtains the first feature and the second feature, fuses them to obtain a fused feature, and transmits it to the large language module. The large language module extracts the encoding vector of the long-distance dependency of the fused feature and transmits it to the third linear layer. The third linear layer generates the granular prediction result corresponding to the field data based on the encoding vector and outputs it through the second output layer.

[0035] By freezing pre-trained weights, the model can significantly reduce computational costs during training and inference while maintaining high prediction accuracy, thereby meeting the stringent requirements of industrial control systems for real-time performance and stability, and enabling online prediction and rapid response.

[0036] Furthermore, the soft measurement model constructs an objective function based on the actual particle size value of the ball mill grinding output and the predicted particle size value predicted by the soft measurement model;

[0037] The objective function MSE is expressed by the following formula;

[0038] ;

[0039] in, It is the total number of samples; It is the first The true granularity value of each sample; It is the first The predicted granularity value for each sample.

[0040] Furthermore, the training process of the soft measurement model employs adaptive learning and regularization mechanisms to optimize the model.

[0041] Beneficial effects:

[0042] This invention provides a soft measurement method for ball mill particle size based on a time-series large model. Combining the powerful expressive capabilities of a large language module and the advantages of a multi-head attention mechanism, it can effectively handle complex, nonlinear data and provide accurate ball mill particle size prediction. Due to the fixed weights of the large language module, the soft measurement model can not only operate efficiently with limited training resources but also automatically adapt to changing data patterns under different operating conditions. The soft measurement model has good scalability, can handle large-scale datasets, and can be easily extended to adapt to other complex industrial tasks. This solves the problems of existing soft measurement models, such as excessive reliance on large-scale sample data, insufficient modeling ability for complex nonlinear process parameters, and low prediction accuracy. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the network structure of the multi-feature fusion module in an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the network structure of the soft measurement model in an embodiment of the present invention. Detailed Implementation

[0045] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0047] This application provides a soft measurement method for ball mill particle size based on a time-series large model, including the following steps:

[0048] Step 1: Construct a multi-feature fusion module for extracting multi-scale features based on Convld K3, Convld K5 and Convld K1. Construct a soft measurement model based on the query matrix, key matrix and value matrix of multi-head attention combined with the large language module with fixed weights and the multi-feature fusion module.

[0049] Please see details. Figure 1 The multi-feature fusion module includes an input layer, Convld K3 layer, Convld K5 layer, Convld K1 layer, split layer, concat layer, and first output layer.

[0050] The input layer is connected in sequence to the Convld K3 layer, the split layer, the cancat layer, the Convld K1 layer, and the first output layer. The split layer and the cancat layer are also connected through the Convld K5 layer.

[0051] Here, the multi-feature fusion module first performs preliminary feature extraction on the input signal using a one-dimensional convolutional layer with a kernel size of 3, generating a feature map with a higher dimensionality. Then, the feature map is divided into two equal parts along the channel dimension to extract and preserve information at different scales. One branch continues to use the original features without any transformation, aiming to maintain the original attributes of the input data; the other branch uses a convolutional layer with a kernel size of 5 to obtain local features within a larger receptive field, thereby capturing richer spatial semantic information. Finally, the outputs of the two branches are concatenated along the channel dimension and fused through a 1×1 convolution to form the final output. This design not only enhances the model's sensitivity to multi-scale features but also improves the efficiency of feature fusion, helping to better integrate local region information while preserving the original information, thus improving the overall model performance.

[0052] The construction of the soft measurement model includes: setting up a first multi-feature fusion module in the query matrix generation branch of multi-head attention; using the fixed weights of the large language model as input to the key matrix and value matrix generation branch; setting up a second multi-feature fusion module in the concatenation and fusion of multi-head attention; and setting up the Transformer framework of the large language model in the stage of fusing feature extraction granularity prediction results. Both the first and second multi-feature fusion modules are multi-feature fusion modules; the terms "first" and "second" are used here only for distinction.

[0053] Step 2: Obtain the field data of ball mill grinding and input it into the soft measurement model. The soft measurement model generates the first feature and query matrix based on the field data. The soft measurement model generates the key matrix and value matrix based on the fixed weight of the big language module. The second feature is generated based on the query matrix, key matrix and value matrix, and fused with the first feature to obtain the fused feature. Then, the granularity prediction result corresponding to the field data is obtained through the big language module.

[0054] In this embodiment, the collected field data includes feed rate, grinding sound, water feed rate, hydrocyclone feed concentration, and hydrocyclone overflow particle size. After collection, the aforementioned data are subjected to conventional processing such as feature extraction, data cleaning, and normalization.

[0055] Here, feature extraction involves extracting relevant features based on the characteristics of the original data using methods such as signal processing and statistical analysis. For example, trend and periodic features can be extracted from time series data, and frequency components of signals can be extracted from frequency domain data. These extracted features help the model better capture the dynamic patterns of ball mill operation, thereby improving prediction accuracy. Data cleaning includes handling missing and outlier values. Common methods include interpolation imputation (e.g., linear interpolation), mean imputation, or outlier removal to ensure data integrity and accuracy. This avoids data issues affecting the model's learning process. Normalization involves converting features to a distribution with a mean of 0 and a standard deviation of 1. Normalization scaling features to a specific range (e.g., 0 to 1) helps improve the model's convergence speed and stability, ensuring consistent numerical ranges for each feature and preventing certain features from having an excessive impact on the model training process.

[0056] Statistical characteristics include central tendency characteristics, dispersion characteristics, distribution characteristics, and data distribution characteristics;

[0057] The central tendency features include the mean and the median; the mean is the average of the data, and the median is the middle value after the data is sorted. If there are an even number of values, the middle two are taken and the mean is calculated.

[0058] Dispersion characteristics include variance, standard deviation, root mean square error, and bias;

[0059] Distribution characteristics include maximum value, minimum value, first quartile, second quartile, third quartile, kurtosis, skewness, and percentiles; the first quartile is the 25% of data that are less than this number, the second quartile is 50%, the third quartile is 75%, and the percentile divides the data into 100 equal parts to represent the relative position of the data.

[0060] Data distribution characteristics include the autocorrelation coefficient.

[0061] For information on the network structure of soft sensor models, please refer to [link / reference]. Figure 2 The soft measurement model specifically includes data branches, multi-head branches, and output branches based on large language modules;

[0062] The data branch generates the first feature based on the statistical characteristics of the field data;

[0063] The multi-branch approach generates a query matrix based on the statistical characteristics of the field data, generates a key matrix and a value matrix based on the fixed weights of the large language module, and then generates a second feature based on the query matrix, key matrix, and value matrix.

[0064] The output branch obtains the first and second features and fuses them to obtain the fused features. The long-distance dependency relationship of the fused features is extracted through the large language module to generate the granular prediction results corresponding to the field data.

[0065] The data branches include, in sequence, a statistical feature layer, a hint layer, a tokenizer layer, an embedding layer, and a linear layer;

[0066] The statistical feature layer transmits statistical features based on the central trend features, dispersion features, distribution features, and data distribution features of the field data to the prompt layer. The prompt layer converts the statistical features into prompt words and transmits them to the tokenizer layer for word segmentation. Then, the first feature is generated through vector transformation in the embedding layer and linear processing in the linear layer.

[0067] By employing statistical feature extraction, prompt word construction, data normalization, and a multi-stage feature fusion process, this invention achieves comprehensive analysis of multi-dimensional data. It not only preserves the key information of the original data but also enhances the semantic expression and modeling capabilities of the data through prompt words and multi-scale feature fusion, ensuring that the prediction model fully captures the inherent patterns of the data.

[0068] The multi-branch structure includes a normalization layer, an unfold layer, a first multi-feature fusion module, a first linear layer, a second linear layer, an MHA layer, and a second feature fusion module.

[0069] The normalization layer normalizes the field data and transmits it to the unfold layer. The unfold layer extracts sliding window features from the normalized data. The first multi-feature fusion module performs feature transformation based on the sliding window features to generate a query matrix.

[0070] Specifically, the first linear layer generates a key matrix based on the fixed weights of the large language module, and the second linear layer generates a value matrix based on the fixed weights of the large language module.

[0071] The MHA layer performs multi-head attention calculations based on the query matrix, key matrix, and value matrix, and then combines these with the second feature fusion module to generate the second feature.

[0072] The output branches are based on large language modules, including the concat layer, large language modules, the third linear layer, and the second output layer;

[0073] The concat layer, the large language module, the third linear layer, and the second output layer are connected in sequence;

[0074] The concat layer obtains the first and second features, fuses them to obtain fused features, and transmits them to the large language module. The large language module extracts the encoding vector of the long-distance dependency of the fused features and transmits it to the third linear layer. The third linear layer generates granular prediction results corresponding to the field data based on the encoding vector and outputs them through the second output layer.

[0075] Regarding the training of the soft measurement model, the soft measurement model constructs an objective function based on the actual particle size value of the ball mill grinding output and the predicted particle size value predicted by the soft measurement model;

[0076] The objective function MSE is expressed by the following formula;

[0077] ;

[0078] in, It is the total number of samples; It is the first The true granularity value of each sample; It is the first The predicted granularity value for each sample.

[0079] The training process of the soft measurement model employs adaptive learning and regularization mechanisms to optimize the model.

[0080] The Adam optimizer with an adaptive learning rate is used during training, with the initial learning rate set to... Their attenuation coefficients are respectively set as and The batch size for training was set to 64, the number of training epochs was set to 100, and finally, a regularization mechanism was introduced, using L2 regularization, with the regularization coefficient set to [value missing]. This is to constrain the model weights, thereby improving its generalization ability and preventing the model from overfitting the training data.

[0081] The soft measurement model is not only applicable to online soft measurement of ball mill particle size in this embodiment, but also has a wide range of generalization capabilities. It can be extended to soft measurement tasks in other process industries such as mineral processing, metallurgy, and chemical engineering. It provides a new technical solution for complex industrial data analysis and real-time monitoring, and has good industrial application prospects and economic benefits.

[0082] Finally, the soft measurement model constructed using the ball mill particle size soft measurement method based on a time-series large model provided by this invention was compared with conventional linear regression models, support vector regression models, random forest models, and neural network models to predict the same 5000 sets of ball mill data (the training sample set of conventional large models includes hundreds of thousands of data sets), and the results were based on the mean squared error (MSE) and coefficient of determination (COP). For a horizontal comparison, please refer to Table 1;

[0083] Table 1: Horizontal comparison results of mean square error and coefficient of determination.

[0084]

[0085] As shown in the table, the soft sensor model performs best in terms of mean squared error (MSE), with an MSE of 0.035, significantly lower than other traditional methods. Linear regression has the highest MSE at 0.065, demonstrating its limitations in handling complex nonlinear relationships. Support vector regression (SVR) and random forest have relatively smaller MSEs of 0.055 and 0.048 respectively, but still lower than the soft sensor model. The neural network has an MSE of 0.040, which is better than traditional methods, but still inferior to the model of this invention.

[0086] In the coefficient of determination ( In the comparison, the soft measurement model also performed excellently. It reached 0.95, far exceeding other methods. The index is used to measure the goodness of fit of a model; the closer the value is to 1, the stronger the predictive power of the model. Linear regression... With a score of only 0.60, it indicates poor performance in capturing complex relationships within the data. Support vector regression... The value is 0.75, slightly higher than linear regression, but still not comparable to deep learning models. Random forest... The value is 0.80, which is the value of the neural network. The value of 0.88 indicates their strong fitting ability in some nonlinear tasks, but none of them can reach the level of soft measurement models.

[0087] According to MSE and The comparison results clearly show that the soft measurement model provides the most accurate and stable prediction results in the ball mill particle size prediction task. Traditional methods such as linear regression and SVR, while applicable in certain situations, perform significantly worse than methods based on large-scale language models (LLM) when dealing with complex nonlinear data. By combining frozen LLM pre-trained weights and a multi-head attention mechanism, the model of this invention demonstrates extremely strong generalization ability and high prediction accuracy when processing complex industrial data, providing a more powerful soft measurement tool for ball mill particle size prediction.

[0088] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A soft measurement method for ball mill particle size based on a time-series large model, characterized in that, Includes the following steps: Step 1: Construct a multi-feature fusion module for extracting multi-scale features based on Convld K3, Convld K5 and Convld K1. Construct a soft measurement model based on the query matrix, key matrix and value matrix of multi-head attention combined with the large language module with fixed weights and the multi-feature fusion module. The multi-feature fusion module includes an input layer, Convld K3 layer, Convld K5 layer, Convld K1 layer, split layer, concat layer, and a first output layer; The input layer is sequentially connected to the Convld K3 layer, the split layer, the cancat layer, the Convld K1 layer, and the first output layer. The split layer and the cancat layer are also connected through the Convld K5 layer. Step 2: Obtain the field data of ball mill grinding and input it into the soft measurement model. The soft measurement model generates the first feature and query matrix based on the field data. The soft measurement model generates the key matrix and value matrix based on the fixed weight of the big language module. The second feature is generated based on the query matrix, key matrix and value matrix, and fused with the first feature to obtain the fused feature. Then, the granularity prediction result corresponding to the field data is obtained through the big language module.

2. The soft measurement method for ball mill particle size based on a time-series large model according to claim 1, characterized in that, The soft measurement model is constructed as follows: a first multi-feature fusion module is set in the query matrix generation branch of multi-head attention, the fixed weights of the large language model are used as the input of the key matrix and value matrix generation branch, a second multi-feature fusion module is set in the splicing fusion of multi-head attention, and the Transformer framework of the large language model is set in the stage of fusing feature extraction granularity prediction results.

3. The soft measurement method for ball mill particle size based on a time-series large model according to claim 1, characterized in that, In step 2, after acquiring the field data, feature extraction is performed to obtain statistical features, and then the statistical features are input into the soft measurement model; The statistical characteristics include central tendency characteristics, dispersion characteristics, distribution characteristics, and data distribution characteristics; The central trend features include the mean and the median; The dispersion characteristics include variance, standard deviation, root mean square error, and bias; The distribution characteristics include the maximum value, minimum value, first quartile, second quartile, third quartile, kurtosis, skewness, and percentile. The data distribution characteristics include the autocorrelation coefficient.

4. The soft measurement method for ball mill particle size based on a time-series large model according to claim 2, characterized in that, The soft measurement model specifically includes data branches, multi-head branches, and output branches based on large language modules; The data branch generates a first feature based on the statistical characteristics of the field data; The multi-head branch generates a query matrix based on the statistical characteristics of the field data, generates a key matrix and a value matrix according to the fixed weight of the large language module, and then generates a second feature based on the query matrix, key matrix and value matrix; The output branch obtains the first feature and the second feature and fuses them to obtain the fused feature. The long-distance dependency of the fused feature is extracted by the large language module to generate the granular prediction result corresponding to the field data.

5. The soft measurement method for ball mill particle size based on a time-series large model according to claim 4, characterized in that, The data branches sequentially include a statistical feature layer, a prompt layer, a tokenizer layer, an embedding layer, and a linear layer; The statistical feature layer transmits statistical features based on the central trend features, dispersion features, distribution features, and data distribution features of the field data to the prompt layer. The prompt layer converts the statistical features into prompt words and transmits them to the tokenizer layer for word segmentation. Then, the first feature is generated through vector transformation in the embedding layer and linear processing in the linear layer.

6. The soft measurement method for ball mill particle size based on a time-series large model according to claim 4, characterized in that, The multi-head branch includes a normalization layer, an unfold layer, a first multi-feature fusion module, a first linear layer, a second linear layer, an MHA layer, and a second feature fusion module; The normalization layer normalizes the field data and transmits it to the unfold layer. The unfold layer extracts sliding window features from the normalized data. The first multi-feature fusion module performs feature transformation based on the sliding window features to generate a query matrix. The first linear layer generates a key matrix based on the fixed weights of the large language module, and the second linear layer generates a value matrix based on the fixed weights of the large language module. The MHA layer performs multi-head attention calculations based on the query matrix, key matrix, and value matrix, and then combines these with the second feature fusion module to generate the second feature.

7. The soft measurement method for ball mill particle size based on a time-series large model according to claim 4, characterized in that, The output branch is based on a large language module, including a concat layer, a large language module, a third linear layer, and a second output layer; The concat layer, the large language module, the third linear layer, and the second output layer are connected in sequence. The concat layer obtains the first feature and the second feature, fuses them to obtain a fused feature, and transmits it to the large language module. The large language module extracts the encoding vector of the long-distance dependency of the fused feature and transmits it to the third linear layer. The third linear layer generates the granular prediction result corresponding to the field data based on the encoding vector and outputs it through the second output layer.

8. The soft measurement method for ball mill particle size based on a time-series large model according to claim 1, characterized in that, The soft measurement model constructs an objective function based on the actual particle size value produced by the ball mill grinding and the predicted particle size value predicted by the soft measurement model; The objective function MSE is expressed by the following formula; ； in, It is the total number of samples; It is the first The true granularity value of each sample; It is the first The predicted granularity value for each sample.

9. The soft measurement method for ball mill particle size based on a time-series large model according to claim 1, characterized in that, The training process of the soft measurement model employs adaptive learning and regularization mechanisms to optimize the model.

Images