Intelligent evaluation method for grinding system efficiency based on large model

Abstract

The present application relates to the technical field of ore dressing automation, and discloses a kind of intelligent evaluation method of grinding system efficiency based on large model.The method includes: the double input channel of efficiency model is constructed based on natural language prompt and time sequence model respectively, the first output channel of efficiency model is constructed in the input channel of time sequence model side, the second output channel of efficiency model is constructed based on the double input channel combining large language model;Real-time data of the grinding system to be scored is collected, and the real-time data is input into the efficiency model, and the score of the grinding system to be scored is obtained based on the evaluation of the first output channel and the second output channel combined with consistency loss constraint, to solve the problem that the existing grinding efficiency algorithm for grinding system is not stable and accurate enough.

Description

Technical Field

[0001] This invention relates to the field of mineral processing automation technology, and in particular to an intelligent evaluation method for the efficiency of grinding systems based on a large model. Background Technology

[0002] With the continuous improvement of automation and intelligent manufacturing in mines, the intelligent monitoring and optimization of the operating efficiency and product fineness of grinding systems, as a key link in the mineral processing flow, has become a core issue that urgently needs to be addressed in the industrial field. Traditional methods for evaluating the efficiency of grinding systems mostly rely on manual experience or single physical models. These methods generally suffer from bottlenecks such as low data utilization, delayed intelligent response, and insufficient generalization ability, making it difficult to meet the needs of modern mines for efficient, intelligent, and precise process control.

[0003] In recent years, with the continuous advancement of industrial data acquisition and information fusion technologies, soft sensing and intelligent diagnostic methods based on multi-source time-series data and statistical features have gradually emerged, bringing new possibilities for the intelligent evaluation of grinding systems. However, existing methods still have many shortcomings in terms of feature fusion depth, model generalization ability, and output interpretability. Specifically, current soft sensing and intelligent diagnostic methods mainly include the following categories:

[0004] First, traditional methods based on statistical modeling and signal processing, such as multiple linear regression and principal component analysis (PCA), while possessing a certain theoretical foundation, have limited ability to model the highly nonlinear and dynamic changes in operating conditions within grinding systems, making them ill-suited for complex conditions and sudden anomalies. Second, shallow machine learning and classic deep learning methods, such as support vector machines (SVM) and traditional neural networks (LSTM, GRU), while improving the ability to model temporal dynamic features, generally rely on information extraction from a single path, making it difficult to fully integrate multi-source heterogeneous features. Furthermore, these methods suffer from insufficient interpretability of output results, failing to provide clear and reliable intelligent decision-making suggestions for industrial operators; simultaneously, their generalization ability and robustness remain limited when encountering new operating conditions or extreme anomalies, hindering their widespread application in practical industrial scenarios. Summary of the Invention

[0005] This invention provides a solution to address the problem that existing algorithms for improving grinding efficiency in grinding systems are not stable or accurate enough.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] This invention provides an intelligent evaluation method for the efficiency of a grinding system based on a large model, comprising the following steps:

[0008] Step 1: Construct an efficiency model for the grinding system. The efficiency model includes a dual-input channel, a first output channel, and a second output channel. The dual-input channel is constructed based on natural language prompts and a temporal model. The first output channel is constructed as an input channel on the temporal model side. The second output channel is constructed based on the dual-input channel combined with a large language model.

[0009] Step 2: Collect real-time data of the grinding system to be evaluated, input the real-time data into the efficiency model, the efficiency model obtains the first evaluation on the dynamic time series based on the first output channel, and obtains the second evaluation on the fusion features of time series and natural language based on the second output channel, and obtains the score of the grinding system to be evaluated based on the first evaluation and the second evaluation combined with the consistency loss constraint.

[0010] Furthermore, the dual input channel includes the following construction based on natural language prompts and a temporal model: constructing a first input channel based on natural language prompts and a second input channel based on a temporal model;

[0011] The first input channel obtains statistical features based on the input data, then obtains natural language prompt words based on the statistical features, and obtains the first output features about the natural language after segmenting, encoding and linear transformation of the natural language prompt words;

[0012] The second input channel normalizes and expands the input data, and then obtains the second output feature about the dynamic time series through the time series model.

[0013] Furthermore, the first input channel sequentially includes a statistical feature layer, a prompt word layer, a word segmentation layer, an encoding layer, and a first linear layer. The first output feature output by the first linear layer and the second output feature output by the second input channel are input into the second output channel for feature fusion, and then combined with the large language model to output the second evaluation.

[0014] Furthermore, the real-time data includes ore hardness, particle size distribution, feed rate, mill speed, media filling rate, and energy consumption, and the statistical characteristics include the maximum value, minimum value, mean, standard deviation, peak value, peak interval, and linear trend of the real-time data.

[0015] Furthermore, the second input channel sequentially includes a first normalization layer, an expansion layer, and a temporal model layer. The first output channel is connected to the temporal model layer. The second output feature output by the temporal model layer is obtained and linearly transformed by the second linear layer set in the first output channel to output the first evaluation. The second output feature output by the temporal model layer is also input into the second output channel with the first output feature output by the first output channel for feature fusion and then combined with the large language model to output the second evaluation.

[0016] Furthermore, the time-series model layer is constructed based on linear transformation and gated neural networks;

[0017] The time-series model layer performs linear transformation and gated neural network processing on the acquired features respectively. Based on the linear transformation, a first linear feature is obtained, and then a first enhanced feature is obtained through the gated neural network. The first linear feature and the first enhanced feature are added together, and after linear transformation and normalization, a second output feature is output.

[0018] Through the above design, by deeply integrating multi-dimensional time-series data and statistical features, the global state and dynamic changes of the grinding system can be fully captured, enabling accurate identification and perception of complex process conditions, which is significantly better than the traditional single data source modeling method.

[0019] Furthermore, the time-series model layer includes an input layer, a linear branch, a gated neural network branch, and an output branch. The input ends of the linear branch and the gated neural network branch are both connected to the input layer, and the output ends of the linear branch and the gated neural network branch are both connected to the output branch.

[0020] The linear branches include a third linear layer and a fourth linear layer in sequence;

[0021] The gated neural network branch includes a recurrent neural network, a first gated sub-branch, and a first multiplication layer. The first gated sub-branch is located between the recurrent neural network and the first multiplication layer. The first gated sub-branch includes a fifth linear layer and a first linear activation layer.

[0022] The output branch includes an addition layer, a sixth linear layer, a second normalization layer, and an output layer, wherein the fourth linear layer and the first multiplication layer are both connected to the addition layer.

[0023] Through the above design, combining linear transformation and gated recurrent neural network (LSTM), and through multi-layer fusion and gating mechanism, the feature expression ability and sequence modeling ability are improved, enabling deep fusion of linear features and temporal features. The LSTM embeds an adaptive gating adjustment mechanism to improve the expression of key features, and the normalization processing after fusion enhances the stability and generalization ability of the layer.

[0024] Furthermore, the second output channel is constructed based on a large language model with weighted freezing;

[0025] The second output channel receives the first output feature and the second output feature, performs feature fusion on the first output feature and the second output feature to obtain the fused feature, obtains the hidden vector based on the fused feature and the large language model with weight freezing, obtains the second linear feature after linear transformation on the hidden vector, performs linear transformation and linear activation based on the second linear feature and the second linear feature element-wise multiplied to obtain the reinforcement feature, and obtains the second evaluation after linear transformation on the reinforcement feature.

[0026] Furthermore, the backbone of the second output channel is connected in the order of fusion layer, large language model layer, seventh linear layer, second multiplication layer and eighth linear layer. A second gated sub-branch is set between the seventh linear layer and the second multiplication layer of the second output channel. The second gated sub-branch includes the ninth linear layer and the second linear activation layer.

[0027] Through the above design, structured natural language prompts are automatically generated using statistical features and combined with a large language model (LLM). This not only improves the model's reasoning ability and generalization performance, but also outputs expert-level process diagnostic suggestions, greatly enhancing the interpretability and practicality of the system output.

[0028] Relying on the pre-trained knowledge and prompting word mechanism of the large language model, it can maintain high accuracy in condition identification and efficiency evaluation even with limited sample data, meeting the application needs of real industrial sites where data is scarce and conditions are complex.

[0029] By introducing a multi-branch feature fusion structure and gating mechanism, the contribution of each feature branch in the final judgment can be dynamically adjusted, which significantly improves the stability and robustness of the model in a variable and complex environment and reduces the sensitivity to abnormal or noisy data.

[0030] Furthermore, the efficiency model employs a main loss and a consistency constraint term to construct a loss function during training. Both the main loss and the consistency constraint term are constructed using mean squared error. The main loss performs regression fitting on the first evaluation, the second evaluation, and the true label. The consistency constraint term constrains the consistency between the outputs of the first evaluation and the second evaluation.

[0031] The loss function is expressed by the following formula:

[0032] ;

[0033] in, Represents the loss function; Indicates the main loss; This represents the weight hyperparameter; Indicates consistency constraints;

[0034] The main loss is expressed by the following formula:

[0035] ;

[0036] in, This indicates the sample size for the first evaluation and the second evaluation. and They represent the first First and second evaluations of each sample; Indicates the true label;

[0037] The constraint is expressed by the following formula:

[0038] .

[0039] By employing multi-branch collaborative output and introducing consistency loss constraints, the consistency and reliability of the system's output results are effectively improved, ensuring more stable and credible intelligent evaluation results. This approach not only focuses on the accurate fitting of the model's two branches to the actual process state but also effectively enhances the output stability and generalization ability of the branch fusion model through consistency constraints. This design balances the utilization of multi-source information with the reliability of the final judgment, providing theoretical support and an engineering foundation for intelligent diagnostic systems for complex industrial processes.

[0040] Beneficial effects:

[0041] This invention provides an intelligent evaluation method for grinding system efficiency based on a large model. By using multi-source statistics and time-series feature extraction, the method enhances the model's ability to perceive the global and dynamic working conditions. By using natural prompts, the method enhances the model's expert knowledge expression and interpretability, thus solving the problems of insufficient stability and accuracy of existing algorithms for grinding system efficiency.

[0042] The model inference and gating mechanism take into account both generalization ability and feature adaptive optimization, while consistency constraints ensure the reliability of multi-output fused features.

[0043] Multi-stage feature processing and fusion optimization design ensure that the model has high computational efficiency while maintaining prediction accuracy, which can meet the actual requirements of online real-time intelligent monitoring and optimization in industrial sites, and facilitate the engineering deployment and promotion of the system. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the network structure of the efficiency model in an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the network structure of the timing model layer in an embodiment of the present invention. Detailed Implementation

[0046] 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.

[0047] 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.

[0048] Please see Figure 1 This application provides an intelligent evaluation method for the efficiency of a grinding system based on a large model, comprising the following steps:

[0049] Step 1: Construct an efficiency model for the grinding system. The efficiency model includes a dual-input channel, a first output channel, and a second output channel. The dual-input channel is constructed based on natural language prompts and a temporal model. The first output channel is constructed as an input channel on the temporal model side. The second output channel is constructed based on the dual-input channel combined with a large language model.

[0050] The real-time data includes ore hardness, particle size distribution, feed rate, mill speed, media filling rate, and energy consumption. In this embodiment, the above data needs to be collected continuously and in real time during the process operation, and the original data is recorded in a time series manner.

[0051] Specifically, the first input channel of the efficiency model is constructed based on natural language prompts, and the second input channel of the efficiency model is constructed based on the time series model;

[0052] The first input channel obtains statistical features based on the input data, then obtains natural language prompt words based on the statistical features, and obtains the first output features after segmenting, encoding, and linear transformation of the natural language prompt words.

[0053] Statistical characteristics include maximum, minimum, mean, standard deviation, peak value, peak interval, and linear trend. If the... The time-series characteristics of a physical quantity (real-time data) are as follows: ,in .

[0054] The maximum value is then expressed by the following formula:

[0055] ;

[0056] in, Indicates the first The maximum value of a physical quantity ; The total number of sampling times. express Time of the first The value of a physical quantity;

[0057] The minimum value is expressed by the following formula:

[0058] ;

[0059] in, Indicates the first The minimum value of a physical quantity;

[0060] The mean is expressed by the following formula:

[0061] ;

[0062] in, Indicates the first The average of the physical quantities;

[0063] Standard deviation is expressed by the following formula:

[0064] ;

[0065] in, Indicates the first The standard deviation of a physical quantity characterizes volatility.

[0066] The peak value is represented by the following formula:

[0067] ;

[0068] in, Represents the set of all local maxima;

[0069] The peak spacing is expressed by the following formula:

[0070] ;

[0071] in, Indicates the first The average interval between adjacent peaks in a group reflects the periodicity of the sequence's fluctuations; For the first The time difference between adjacent peaks. Specifically, Indicates the first The moment when the peak occurs, The time of its next adjacent peak;

[0072] A linear trend is represented by the following formula:

[0073] ;

[0074] in, This refers to the trend slope obtained by performing a linear regression fit on the data segment, which is used to quantify the trend of data change over time. Indicates the first Time point or time series index of a sample For the corresponding data observations, and They are respectively and The mean, Representing time variables (such as sampling time, sequence number, etc.) is the foundation of linear trend analysis. This refers to observed values ​​(such as process parameters, characteristic indicators, etc.).

[0075] Regarding the network structure of the first input channel, the first input channel sequentially includes a statistical feature layer, a prompt word layer, a word segmentation layer, an encoding layer, and a first linear layer. The first output feature output by the first linear layer and the second output feature output by the second input channel are input into the second output channel for feature fusion, and then combined with the large language model to output the second evaluation.

[0076] The second input channel normalizes and expands the input data, and then obtains the second output feature through a time series model.

[0077] Regarding the network structure of the second input channel, the second input channel sequentially includes a first normalization layer, an expansion layer, and a temporal model layer. The first output channel is connected to the temporal model layer. The second output feature of the temporal model layer is obtained and linearly transformed by the second linear layer set in the first output channel to output the first evaluation. The second output feature of the temporal model layer is also input into the second output channel with the first output feature of the first output channel for feature fusion and then combined with the large language model to output the second evaluation.

[0078] Specifically, the normalization process used here is Norm processing, which eliminates the influence of different physical quantities' dimensions. The normalized data is then divided into segments of length [length missing] using a sliding window (unfold). The sequence, with a common parameter range of . The segmented sequence is input into a time-series model layer (GFLSTM) to extract temporal dynamic features. Subsequently, a linear layer performs feature transformation on the output to obtain the temporal feature vector. It can be expressed as the following formula:

[0079] ;

[0080] in, This represents the data after normalization. Indicates based on window length Expand the time series data, Usually taken ; This represents the output features of the temporal model layer, which is used to extract temporal dependencies. The number of hidden units is also mentioned. Usually set to ; This represents a linear transformation, which further transforms the output features of the time series model layer.

[0081] Please see Figure 2 In the second input channel, the temporal model layer is constructed based on linear transformation and gated neural network. The temporal model layer processes the acquired features by linear transformation and gated neural network respectively. The first linear feature is obtained based on linear transformation, and the first enhanced feature is obtained through gated neural network. The first linear feature and the first enhanced feature are added together, and the second output feature is output after linear transformation and normalization.

[0082] The network structure of the temporal model layer includes an input layer, a linear branch, a gated neural network branch, and an output branch. The input ends of the linear branch and the gated neural network branch are connected to the input layer, and the output ends of the linear branch and the gated neural network branch are connected to the output branch.

[0083] Specifically, the linear branches include the third linear layer and the fourth linear layer in sequence;

[0084] The gated neural network branch includes a recurrent neural network, a first gated sub-branch, and a first multiplication layer. The first gated sub-branch is located between the recurrent neural network and the first multiplication layer. The first gated sub-branch includes a fifth linear layer and a first linear activation layer.

[0085] The output branch includes an additive layer, a sixth linear layer, a second normalization layer, and an output layer. The fourth linear layer and the first multiplication layer are both connected to the additive layer.

[0086] The second normalization layer here also uses Norm normalization to improve the model's convergence speed and generalization performance.

[0087] For the second output channel of the efficiency model, a large language model is constructed based on weight freezing.

[0088] Specifically, the second output channel receives the first output feature and the second output feature, performs feature fusion on the first output feature and the second output feature to obtain the fused feature, obtains the hidden vector based on the fused feature and the large language model with weight freezing, obtains the second linear feature after linear transformation on the hidden vector, performs linear transformation and linear activation based on the second linear feature and the second linear feature element-wise multiplied with the second linear feature to obtain the reinforcement feature, and obtains the second evaluation after linear transformation on the reinforcement feature.

[0089] The main trunk of the second output channel is connected in the order of fusion layer, large language model layer, seventh linear layer, second multiplication layer and eighth linear layer. A second gated sub-branch is set between the seventh linear layer and the second multiplication layer of the second output channel. The second gated sub-branch includes the ninth linear layer and the second linear activation layer.

[0090] The first and second evaluations output by the efficiency model are used to characterize the achievement of fineness standards and efficiency changes, respectively. The first evaluation reflects whether the material fineness meets the process requirements under the current operating conditions, for example, using a -200µm particle size fraction greater than 60% as the criterion for achieving fineness standards. The second evaluation reflects the changes in efficiency, for example, using the percentage increase in efficiency as the judgment criterion. Through comprehensive analysis of these two evaluation values, the operating condition discrimination results can be mapped to the set {+100, +50, 0, -50, -100}, with the following specific meanings:

[0091] +100: Fineness meets standards and efficiency is improved;

[0092] +50: Fineness meets the standard, but efficiency is reduced;

[0093] 0: Fineness meets the standard and efficiency remains unchanged;

[0094] -50: Fineness not met but efficiency improved;

[0095] -100: Fineness not met and efficiency reduced;

[0096] The specific discrimination method is as follows:

[0097] The first evaluation is used to determine whether the fineness meets the standard. If the fineness (e.g., the -200µm particle size fraction) is greater than 60%, it is judged as meeting the fineness standard; otherwise, it is considered as not meeting the standard.

[0098] The second evaluation is used to judge the change in efficiency. If the efficiency increase is greater than or equal to 3%, it is judged as an increase in efficiency; if the efficiency change is between -3% and 3% and the fineness meets the standard, it is judged as normal fluctuation and output as 0; if the efficiency increase is less than -3%, it is judged as a decrease in efficiency.

[0099] The comprehensive discriminant mapping is as follows:

[0100] When the fineness meets the standard and the efficiency increases, the output is +100;

[0101] When the fineness meets the standard but the efficiency decreases, the output is +50;

[0102] When the fineness meets the standard and the efficiency change is between -3% and 3%, the output is 0, which is considered normal fluctuation.

[0103] When the fineness does not meet the standard but the efficiency is increased, the output is -50;

[0104] When the fineness does not meet the standard and the efficiency decreases, output -100.

[0105] The criterion for achieving the fineness standard can be set to a particle size fraction of -200µm greater than 60%, and the criterion for increased efficiency can be set to an efficiency improvement of greater than or equal to 3%. The specific thresholds for the first and second evaluations can be further adjusted and set according to actual process requirements or model output characteristics.

[0106] During training, the efficiency model uses a main loss and a consistency constraint term to construct the loss function. Both the main loss and the consistency constraint term are constructed using mean squared error. The main loss performs regression fitting on the first evaluation, the second evaluation, and the true label, while the consistency constraint term constrains the consistency of the outputs of the first evaluation and the second evaluation.

[0107] The loss function is expressed by the following formula:

[0108] ;

[0109] in, Represents the loss function; Indicates the main loss; This represents the weight hyperparameter; Indicates consistency constraints;

[0110] The principal loss is expressed by the following formula:

[0111] ;

[0112] in, This indicates the sample size for the first evaluation and the second evaluation. and They represent the first First and second evaluations of each sample; Indicates the true label;

[0113] Constraints are expressed by the following formula:

[0114] .

[0115] For training the efficiency model, the batch size was set to 64, and the Adam optimizer (learning rate set to 1e-4, weight decay 1e-5) was used to improve convergence efficiency and suppress overfitting. Mean squared error (MSELoss) was chosen as the loss function to suit the regression task. The total number of training epochs was set to 100, and an early stopping mechanism was introduced: if the validation set metrics showed no improvement for 10 consecutive epochs, training was terminated early to prevent overfitting. Furthermore, a fixed random seed of 42 was used during training to ensure experimental reproducibility. Gradient clipping (maximum norm 5.0) and learning rate scheduling (ReduceLROnPlateau) were enabled when necessary to further improve the model's stability and generalization ability. With the above hardware and parameter configuration, the efficiency and reliability of model training can be fully guaranteed, providing a solid foundation for subsequent performance evaluation and industrial applications.

[0116] The efficiency model constructed above possesses a multi-branch collaborative mechanism, which can fully explore and utilize the dynamic and statistical features in the original data to achieve global and detailed perception of complex working conditions. With the help of natural language prompts and expert knowledge embedding, the interpretability of model reasoning and decision reference value are significantly enhanced. The innovative introduction of gating mechanism and consistency constraint improves the model's generalization ability and output stability. It has good industrial applicability and system scalability, providing a solid technical foundation and support for the intelligent upgrading of mines and the optimization of grinding processes.

[0117] Step 2: Collect real-time data of the grinding system to be evaluated, input the real-time data into the efficiency model, and obtain the score of the grinding system to be evaluated based on the evaluation of the first output channel and the second output channel combined with the consistency loss constraint.

[0118] This embodiment compares and tests various mainstream regression models with the method of this invention. The experimental data covers multi-source time-series data and statistical features of key process parameters during grinding. All models were trained and tested using the same dataset. The comparison models include multiple linear regression, classic LSTM, LSTM fusion with statistical features, Gated LSTM, LLM + statistical features, and the "multi-source fusion large model" (efficiency model) proposed in this invention. The main evaluation metric is mean squared error (MSE). ) and coefficient of determination ( To comprehensively reflect the model's performance in terms of regression accuracy and fitting ability, please refer to Table 1 for specific results;

[0119] Table 1: Comparison of various mainstream regression models with the model of this invention;

[0120]

[0121] As shown in Table 1, the comparative experimental results reveal significant differences in the performance of each model in the grinding system efficiency regression task. Multiple linear regression was used as the baseline model. It is 16.2. The value of 0.81 indicates that its ability to fit the data is limited. The classic LSTM model introduces time series modeling capabilities. It dropped to 12.8. The score was improved to 0.87, indicating a performance increase. Combining statistical features with LSTM allows the model to grasp global and dynamic information more comprehensively. Further reduced to 10.6, Improved to 0.89. Gated LSTM optimizes feature flow through a gating mechanism. It is 9.8. The value is 0.90, indicating a stronger feature representation capability;

[0122] Introducing a large-scale pre-trained language model (LLM) and combining it with statistical features significantly improves model performance. It dropped to 7.4. A score of 0.92 indicates that the model possesses better generalization and nonlinear modeling capabilities. The "multi-source fusion large model" (efficiency model) proposed in this invention, through innovative designs such as multi-branch feature encoding, natural language prompts, gating mechanisms, and consistency constraints, will... It further decreased to 4.1. The accuracy was improved to 0.95, significantly outperforming other comparative models. This result fully verifies the outstanding advantages of our method in integrating multi-source features, improving model accuracy, and enhancing industrial applicability, providing more reliable technical support for the intelligent diagnosis and optimization of grinding systems.

[0123] 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 method for intelligent evaluation of grinding system efficiency based on a large model, characterized in that, Includes the following steps: Step 1: Construct an efficiency model for the grinding system. The efficiency model includes a dual-input channel, a first output channel, and a second output channel. The dual-input channel is constructed based on natural language prompts and a temporal model. The first output channel is constructed as an input channel on the temporal model side. The second output channel is constructed based on the dual-input channel combined with a large language model. The dual input channel includes a construction based on natural language prompts and a temporal model, comprising: constructing a first input channel based on natural language prompts and a second input channel based on a temporal model; The first input channel obtains statistical features based on the input data, then obtains natural language prompt words based on the statistical features, and obtains the first output features about the natural language after segmenting, encoding and linear transformation of the natural language prompt words; The second input channel normalizes and expands the input data, and then obtains the second output feature about the dynamic time series through the time series model; The second input channel includes a first normalization layer, an expansion layer, and a temporal model layer in sequence. The first output channel is connected to the temporal model layer. The second output feature output by the temporal model layer is obtained and linearly transformed by the second linear layer set in the first output channel to output the first evaluation. The second output feature output by the temporal model layer is also input into the second output channel with the first output feature output by the first input channel for feature fusion and then combined with the large language model to output the second evaluation. Step 2: Collect real-time data of the grinding system to be evaluated, input the real-time data into the efficiency model, the efficiency model obtains the first evaluation on the dynamic time series based on the first output channel, and obtains the second evaluation on the fusion features of time series and natural language based on the second output channel, and obtains the score of the grinding system to be evaluated based on the first evaluation and the second evaluation combined with the consistency loss constraint.

2. The intelligent evaluation method for grinding system efficiency based on a large model according to claim 1, characterized in that, The first input channel sequentially includes a statistical feature layer, a prompt word layer, a word segmentation layer, an encoding layer, and a first linear layer. The first output feature output by the first linear layer and the second output feature output by the second input channel are input into the second output channel for feature fusion, and then combined with the large language model to output the second evaluation.

3. The intelligent evaluation method for grinding system efficiency based on a large model according to claim 1, characterized in that, The real-time data includes ore hardness, particle size distribution, feed rate, mill speed, media filling rate, and energy consumption. The statistical characteristics include the maximum value, minimum value, mean, standard deviation, peak value, peak interval, and linear trend of the real-time data.

4. The intelligent evaluation method for grinding system efficiency based on a large model according to claim 1, characterized in that, The time-series model layer is constructed based on linear transformation and gated neural networks; The time-series model layer performs linear transformation and gated neural network processing on the acquired features respectively. Based on the linear transformation, a first linear feature is obtained, and then a first enhanced feature is obtained through the gated neural network. The first linear feature and the first enhanced feature are added together, and after linear transformation and normalization, a second output feature is output.

5. The intelligent evaluation method for grinding system efficiency based on a large model according to claim 4, characterized in that, The time series model layer includes an input layer, a linear branch, a gated neural network branch, and an output branch. The input ends of the linear branch and the gated neural network branch are connected to the input layer, and the output ends of the linear branch and the gated neural network branch are connected to the output branch. The linear branches include a third linear layer and a fourth linear layer in sequence; The gated neural network branch includes a recurrent neural network, a first gated sub-branch, and a first multiplication layer. The first gated sub-branch is located between the recurrent neural network and the first multiplication layer. The first gated sub-branch includes a fifth linear layer and a first linear activation layer. The output branch includes an addition layer, a sixth linear layer, a second normalization layer, and an output layer, wherein the fourth linear layer and the first multiplication layer are both connected to the addition layer.

6. The intelligent evaluation method for grinding system efficiency based on a large model according to any one of claims 1-5, characterized in that, The second output channel is constructed based on a large language model with weighted freezing; The second output channel receives the first output feature and the second output feature, performs feature fusion on the first output feature and the second output feature to obtain the fused feature, obtains the hidden vector based on the fused feature and the large language model with weight freezing, obtains the second linear feature after linear transformation on the hidden vector, performs linear transformation and linear activation based on the second linear feature and the second linear feature element-wise multiplied to obtain the reinforcement feature, and obtains the second evaluation after linear transformation on the reinforcement feature.

7. The intelligent evaluation method for grinding system efficiency based on a large model according to claim 6, characterized in that, The backbone of the second output channel is connected in the order of fusion layer, large language model layer, seventh linear layer, second multiplication layer and eighth linear layer. A second gated sub-branch is set between the seventh linear layer and the second multiplication layer of the second output channel. The second gated sub-branch includes the ninth linear layer and the second linear activation layer.

8. The intelligent evaluation method for grinding system efficiency based on a large model according to claim 6, characterized in that, The efficiency model employs a main loss and a consistency constraint term to construct a loss function during training. Both the main loss and the consistency constraint term are constructed using mean squared error. The main loss performs regression fitting on the first evaluation, the second evaluation, and the true label. The consistency constraint term constrains the consistency between the outputs of the first evaluation and the second evaluation.

Images