Test-time adaptive fine coal ash content prediction method based on latent space concept drift evaluation

Abstract

The application discloses a test-time adaptive clean coal ash content prediction method based on hidden space concept drift evaluation, and belongs to the technical field of coal product quality evaluation. In view of the problem of insufficient generalization ability of a traditional prediction model caused by non-stationary working conditions of a dense medium coal separation process, a double-input mechanism of a sample-offset sample pair is constructed, hidden space features of process parameters are extracted by using an LSTNet network, and a dynamic offset of clean coal ash content is directly estimated by calculating hidden space differences between a current sample and offset samples. A composite loss function is constructed by fusing a target loss and a spectral regularized reconstruction loss, and the structural stability of the hidden space representation is enhanced. In the test stage, only the measured ash content at the last moment and the offset value predicted by the model are combined, so that the adaptive correction of the ash content value under the new working condition can be realized. The application effectively improves the accuracy, interpretability and robustness of the clean coal ash content prediction under the conditions of working condition change and cross-scene, and is suitable for intelligent production control scenes of coal preparation plants.

Description

Technical Field

[0001] This invention belongs to the field of coal product quality evaluation technology, specifically involving a test-time adaptive method for predicting the ash content of clean coal based on implicit space concept drift assessment. Background Technology

[0002] Ash content in clean coal is a core indicator for evaluating the quality of coal products. Effective control and accurate prediction of ash content in clean coal are of paramount importance for improving product quality, optimizing resource utilization, reducing production costs, and enhancing production stability.

[0003] Traditionally, the determination of clean coal ash content has mainly relied on laboratory chemical analysis methods, which calculate the ash content by burning coal samples and measuring the mass of residual minerals. While this method offers high accuracy, it is complex, time-consuming, and exhibits significant time lag, making it difficult to meet the demands of modern coal preparation production systems for real-time online monitoring and rapid control.

[0004] Current mainstream data-driven prediction methods are mostly based on direct modeling of collected samples, but this paradigm faces three challenges: First, the internal mechanism of the model is difficult to explain, resulting in a serious "black box" problem; second, the non-stationary nature of the heavy media coal preparation process makes it difficult for the model to guarantee prediction accuracy in the early stages of changes in operating conditions and when the amount of data is insufficient; third, there are significant differences in coal quality characteristics and process parameters among different coal preparation plants, making it difficult to directly transfer and apply high-precision models trained in a single scenario.

[0005] Concept drift, or the change in the mapping relationship between input features and target variables, is a significant cause of the aforementioned problems—the same process parameters can lead to drastically different sorting results in different coal preparation plants, directly limiting the model's cross-scenario generalization ability. To overcome this challenge, this invention learns the dynamic evolution law of clean coal ash content from the latent space features of process parameters. By combining the ash content value at the previous moment with the change during the current period, it achieves accurate prediction of the ash content value at the current moment. This technical approach not only enhances the interpretability of the prediction process but also effectively adapts to the non-stationary characteristics of the sorting process, significantly improving the model's response accuracy in the early stages of changes in operating conditions. More importantly, although the absolute ash content values ​​of different coal types or coal preparation plants vary significantly, their drift patterns often have cross-scenario universality. Therefore, knowledge transfer and rapid modeling with few samples can be achieved, providing new technical support for improving the economic efficiency and intelligence level of coal preparation plants. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention proposes a test-time adaptive method for predicting the ash content of clean coal based on the concept of latent space drift assessment. It constructs sample-offset sample pairs for comparative learning, and achieves accurate prediction of clean coal ash content by learning the drift patterns of the samples. The method is rationally designed, overcomes the shortcomings of existing technologies, and demonstrates excellent performance.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The test-time adaptive method for predicting the ash content of clean coal based on the implicit space concept drift assessment includes the following steps: Step 1: Collect process parameters and corresponding clean coal ash content in the heavy medium coal preparation process to construct a coal preparation dataset; Step 2: Preprocess the coal preparation dataset and divide it into training and testing sets; Step 3: Based on the preprocessed data, construct sample-offset sample pairs, wherein the sample-offset sample pairs include the sample at the current time and its corresponding offset sample; Step 4: Input the sample and the offset sample into the pre-built LSTNet model respectively, and extract the latent space features of the sample and the latent space features of the offset sample. Step 5: Based on the latent space features of the sample and the latent space features of the offset sample, calculate the target offset loss, target loss and reconstruction loss, and use the weighted combination of the three as the final loss function to train the model. Step 6: When deploying and applying the model, obtain the input sample at the current time and the input sample at the previous time. Use the trained LSTNet network to extract the latent space features of the two and calculate the difference to obtain the predicted ash offset. Then add the ash offset to the measured ash content at the previous time to obtain the predicted value of clean coal ash content at the current time.

[0008] Furthermore, in step 1, the process parameters include sampling date, sampling time, sorting density, coal slime content, magnetic material content, hydrocyclone inlet pressure, and liquid level in the mixing tank.

[0009] Furthermore, in step 2, the preprocessing includes outlier detection and missing value processing; the ratio of the training set to the test set is 9:1.

[0010] Further, the method for constructing the offset sample is as follows: calculate a random offset, and offset the current sample backward by the random offset to obtain the offset sample; the sample-offset sample pair is denoted as... ,in, For the current time sample, Let the sample be an offset sample, and satisfy: ; in, This is a random offset.

[0011] Further, in step 4, the LSTNet model includes convolutional neural network layers, recurrent neural network layers, and recurrent skip neural network layers; for input samples The process of extracting its latent space features is as follows: Step 4.1: Extract local features of samples using a convolutional neural network. ; Step 4.2: Use a recurrent neural network to analyze the local features. Perform short-term dependency modeling to obtain features ; Step 4.3: Use a recurrent skip neural network to analyze the local features. Perform long-term dependency modeling to obtain features ; Step 4.4, and By concatenating the data, the latent space features of the samples are obtained. ; The overall calculation process is expressed as follows: ; ; ; ; in, This represents a two-dimensional convolution operation. Indicates the GRU layer. This indicates skipping GRU layers. This indicates a splicing operation.

[0012] Furthermore, in step 5, the formula for calculating the target offset loss is: ; ; in, Indicates target offset loss. Describes the L1 loss function. Represents the total number of samples. and These represent the two input values ​​in the L1 loss. and These are the latent space features of the current sample and the offset sample, respectively. and These represent the ash content of clean coal at the current moment and the ash content of clean coal at the next moment, respectively.

[0013] The target loss is calculated using the mean square error, and the formula is as follows: ; in, Indicates target loss. This represents the predicted representation obtained by sequentially passing the latent space features of the current sample through a projection network and a prediction network. This represents the target representation obtained by applying a projection network to the latent space features of the offset samples. The formula for calculating the reconstruction loss is: ; in, Indicates the reconstruction loss. This represents the spectral regularization loss function; The formula for calculating the final loss function is as follows: ; ; in, This represents the final loss function. Indicates sample Intermediate losses, The weights are used to balance the target loss and the reconstruction loss. The global weights are the intermediate losses.

[0014] Furthermore, in step 6, after the model training is completed, the final trained model is evaluated using a test set, and the best-performing model is applied to the actual heavy media coal preparation process to achieve accurate prediction of the ash content of clean coal.

[0015] The beneficial technical effects of this invention are as follows: 1. Enhanced model interpretability. Traditional clean coal ash content prediction models struggle to trace the specific causes of fluctuations in predicted values, limiting their practical guidance in actual production. This invention, by predicting the dynamic change patterns of clean coal ash content, visually presents the key factors influencing ash content fluctuations, significantly improving the model's interpretability and thus enhancing its guidance for on-site operations and process adjustments.

[0016] 2. Enhanced proactive adaptability to non-stationary processes. When concept drift occurs, traditional models often require accumulating a large number of new samples to complete parameter adjustments, making it difficult to guarantee prediction accuracy in the early stages of drift. This invention extends the modeling object from static mapping to dynamic learning of drift patterns. Through forward-looking prediction of drift trends, it achieves rapid response to new operating conditions, significantly improving the accuracy and stability of ash content prediction during periods of changing operating conditions, bringing better economic benefits to coal preparation plants.

[0017] 3. Achieving cross-condition knowledge transfer and few-sample learning. Traditional methods are limited by the significant differences in coal quality characteristics, equipment parameters, and process conditions among different coal preparation plants, making effective cross-scenario transfer difficult. This invention utilizes the similarity of the clean coal ash drift mechanism across scenarios, and by modeling the inherent laws of ash change, it achieves model adaptation with only a small number of samples under new operating conditions. This feature effectively improves the model's generalization ability and transfer efficiency, providing strong support for the intelligent promotion and deployment of coal preparation plants. Attached Figure Description

[0018] Figure 1 This is a flowchart of the test-time adaptive method for predicting the ash content of clean coal based on the concept of latent space drift assessment of the present invention.

[0019] Figure 2 This is a structural diagram of the LSTNet model used in this invention.

[0020] Figure 3 This is a diagram of the GRU layer structure in the LSTNet model used in this invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be further described below with reference to specific examples: Test-time adaptive method for predicting clean coal ash content based on implicit space concept drift assessment, such as Figure 1 As shown, it includes the following steps: Step 1: Collect process parameters and corresponding clean coal ash content in the heavy medium coal preparation process to construct a coal preparation dataset; Based on various sensors deployed in the heavy media coal preparation process at the coal preparation plant, as well as manual records, the system collected key process parameters, including separation density, coal slime content, and magnetic material content, along with corresponding clean coal ash content indicators. Process parameters included multiple dimensions such as sampling date, sampling time, separation density, coal slime content, magnetic material content, hydrocyclone inlet pressure, and combined medium tank level. Building upon this, through in-depth data mining and systematic organization, a heavy media coal preparation process dataset was constructed for subsequent modeling and analysis.

[0022] The data used in this invention comes from the actual production process of a coal preparation plant under Shandong Energy Zaozhuang Mining Group, covering the ash content of clean coal and key process parameters in the heavy media separation process. Through in-depth mining and systematic analysis of this data, solid data support is provided for subsequent prediction of clean coal ash content. Tables 1 and 2 show examples of datasets for process parameters and clean coal ash content, respectively. Key process parameters include sampling date, sampling time, magnetic content, coal slime content, separation density, hydrocyclone inlet pressure, and combined media tank level, with a sampling frequency of once per minute. In actual production, the coal preparation plant organizes production in three 8-hour shifts, and each shift conducts 3 to 5 raw coal float-sink experiments to determine the raw coal ash content according to the work schedule. To fully utilize the temporal information contained in historical data, this invention adopts the following strategy to construct samples: using the clean coal ash sampling time as the prediction target, process parameter data from the 8 hours prior to that time are selected as input features to predict the offset between the clean coal ash content at that time and the clean coal ash content at the previous time. Using the methods described above, a total of 18 months of heavy media coal preparation process data from the plant were collected and organized, and a dataset containing 3,290 samples was finally constructed.

[0023] Table 1 Example of process parameter data Table 2 Example of clean coal ash content data Step 2: Preprocess the coal preparation dataset and divide it into training and test sets; To enhance the model's ability to extract information from the heavy media coal preparation process, a systematic data preprocessing was performed on the collected raw dataset. Preprocessing included outlier detection and missing value handling to improve data quality and consistency. Specific operations included sensor error correction and erroneous record correction, aiming to eliminate noise and distortion introduced during data acquisition and ensure the accuracy and consistency of the input data. After cleaning and reconstruction, the data was transformed into a structured format suitable for subsequent modeling input requirements, laying the foundation for the model to fully learn the latent features in the data. The data preprocessing stage effectively reduced prediction bias caused by raw data quality issues and played a crucial role in improving the accuracy of clean coal ash content prediction.

[0024] After preprocessing, the obtained data is randomly divided into training and testing sets in a 9:1 ratio for subsequent model training and evaluation.

[0025] Step 3: Based on the preprocessed data, construct sample-offset sample pairs, which include the sample at the current time and its corresponding offset sample; The method for constructing the offset sample is as follows: calculate a random offset, and offset the current sample backward by the random offset to obtain the offset sample; the sample-offset sample pair is denoted as... ,in, For the current time sample, Let the sample be an offset sample, and satisfy: ; in, This is a random offset.

[0026] Based on the sample and the offset, an offset sample is extracted, and a sample-offset sample pair is constructed.

[0027] Step 4: Input the sample and the offset sample into the pre-built LSTNet model respectively, and extract the latent space features of the sample and the latent space features of the offset sample. The LSTNet model structure is as follows: Figure 2 As shown, it includes convolutional neural network layers, recurrent neural network layers, and recurrent skip neural network layers; for input samples The process of extracting its latent space features is as follows: Step 4.1: Input the obtained samples and offset samples into the model respectively, and use a Convolutional Neural Network (CNN) to extract the local features of the samples. ; Step 4.2: Use a Recurrent Neural Network (RNN) to process the local features. Perform short-term dependency modeling to obtain features ; Step 4.3: Use a recurrent skip neural network to analyze the local features. Perform long-term dependency modeling to obtain features ; Step 4.4: Concatenate the short-term dependency modeling implemented by RNN with the long-term dependency modeling implemented by Skip-RNN. and By concatenating the data, the latent space features of the samples are obtained. ; The overall calculation process is expressed as follows: ; ; ; ; in, This represents a two-dimensional convolution operation. Indicates the GRU layer. This indicates skipping GRU layers. This indicates a splicing operation.

[0028] GRU layer structure as follows Figure 3 As shown, the specific calculation formula is as follows: ; ; ; ; in, For the present The input value of the time unit, and These are the hidden states at the current time and the previous time, respectively. This indicates the candidate state of the current work unit. and These represent the reset door and the update door, respectively. , and Represents the weight matrix. For the sigmoid function, The tanh activation function is used. This indicates the dot product operation.

[0029] Step 5: Based on the latent space features of the sample and the latent space features of the offset sample, calculate the target offset loss, target loss and reconstruction loss, and use the weighted combination of the three as the final loss function to train the model. The formula for calculating the target offset loss is: ; ; in, Indicates target offset loss. Describes the L1 loss function. Represents the total number of samples. and These represent the two input values ​​in the L1 loss. and These are the latent space features of the current sample and the offset sample, respectively. and These represent the ash content of clean coal at the current moment and the ash content of clean coal at the next moment, respectively.

[0030] The target loss is calculated using the mean square error, and the formula is as follows: ; in, Indicates target loss. This represents the predicted representation obtained by sequentially passing the latent space features of the current sample through a projection network and a prediction network. This represents the target representation obtained by passing the latent space features of the offset samples through a projection network (Projector), with the mean squared error constraining the two representations. The formula for calculating the reconstruction loss is: ; in, Indicates the reconstruction loss. This represents the spectral regularization loss function; The formula for calculating the final loss function is as follows: ; ; in, This represents the final loss function. Indicates sample Intermediate losses, The weights are used to balance the target loss and the reconstruction loss. λ represents the global weights for the intermediate loss. Experimental data shows that in this embodiment, λ and α are set to 0.6 and 0.1, respectively.

[0031] Furthermore, to achieve optimal model performance, this invention adjusts the designed hyperparameters. Specifically, the learning rate is 0.001; the Adam optimizer is used; the batch size is 1024; and the number of training epochs is 1600. Simultaneously, to prevent overfitting, this invention monitors the changes in various metric values ​​during training, and saves the current optimal model parameters when the loss decreases.

[0032] Step 6: When deploying and applying the model, obtain the input sample at the current time and the input sample at the previous time. Use the trained LSTNet network to extract the latent space features of the two and calculate the difference to obtain the predicted ash offset. Then add the ash offset to the measured ash content at the previous time to obtain the predicted value of clean coal ash content at the current time.

[0033] After training the model, the final model is evaluated using a test set, and the best-performing model is applied to the actual heavy media coal preparation process to achieve accurate prediction of clean coal ash content.

[0034] The model's final performance is measured by comparing the predicted ash content with the actual difference in clean coal ash content, and calculating indicators such as relative absolute error, average reconstruction loss, and average difference loss. The evaluation results on the test set are: relative absolute error of 0.0945, average reconstruction loss of 0.1239, and average difference loss of 0.0244. Applying the evaluated model to the actual coal preparation process ultimately achieves accurate prediction of clean coal ash content, further improving the efficiency of the heavy media coal preparation process.

[0035] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A test-time adaptive method for predicting the ash content of clean coal based on the implicit space concept drift assessment, characterized in that, Includes the following steps: Step 1: Collect process parameters and corresponding clean coal ash content in the heavy medium coal preparation process to construct a coal preparation dataset; Step 2: Preprocess the coal preparation dataset and divide it into training and testing sets; Step 3: Based on the preprocessed data, construct sample-offset sample pairs, wherein the sample-offset sample pairs include the sample at the current time and its corresponding offset sample; Step 4: Input the sample and the offset sample into the pre-built LSTNet model respectively, and extract the latent space features of the sample and the latent space features of the offset sample. Step 5: Based on the latent space features of the sample and the latent space features of the offset sample, calculate the target offset loss, target loss and reconstruction loss, and use the weighted combination of the three as the final loss function to train the model. Step 6: When deploying and applying the model, obtain the input sample at the current time and the input sample at the previous time. Use the trained LSTNet network to extract the latent space features of the two and calculate the difference to obtain the predicted ash offset. Then add the ash offset to the measured ash content at the previous time to obtain the predicted value of clean coal ash content at the current time.

2. The test-time adaptive method for predicting clean coal ash content based on implicit space concept drift assessment according to claim 1, characterized in that, In step 1, the process parameters include sampling date, sampling time, sorting density, coal slime content, magnetic material content, hydrocyclone inlet pressure, and liquid level in the mixing tank.

3. The test-time adaptive method for predicting clean coal ash content based on implicit space concept drift assessment according to claim 1, characterized in that, In step 2, the preprocessing includes outlier detection and missing value handling; the ratio of the training set to the test set is 9:

1.

4. The test-time adaptive method for predicting clean coal ash content based on implicit space concept drift assessment according to claim 1, characterized in that, The method for constructing the offset sample is as follows: calculate a random offset, and offset the current sample backward by the random offset to obtain the offset sample; the sample-offset sample pair is denoted as... ,in, For the current time sample, Let the sample be an offset sample, and satisfy: ； in, This is a random offset.

5. The test-time adaptive method for predicting clean coal ash content based on implicit space concept drift assessment according to claim 1, characterized in that, In step 4, the LSTNet model includes convolutional neural network layers, recurrent neural network layers, and recurrent skip neural network layers; for input samples The process of extracting its latent space features is as follows: Step 4.1: Extract local features of samples using a convolutional neural network. ; Step 4.2: Use a recurrent neural network to analyze the local features. Perform short-term dependency modeling to obtain features ; Step 4.3: Use a recurrent skip neural network to analyze the local features. Perform long-term dependency modeling to obtain features ; Step 4.4, and By concatenating the data, the latent space features of the samples are obtained. ; The overall calculation process is expressed as follows: ； ； ； ； in, This represents a two-dimensional convolution operation. Indicates the GRU layer. This indicates skipping GRU layers. This indicates a splicing operation.

6. The test-time adaptive method for predicting clean coal ash content based on implicit space concept drift assessment according to claim 1, characterized in that, In step 5, the formula for calculating the target offset loss is: ； ； in, Indicates target offset loss. Describes the L1 loss function. Represents the total number of samples. and These represent the two input values ​​in the L1 loss. and These are the latent space features of the current sample and the offset sample, respectively. and These represent the current and next time values ​​of clean coal ash content, respectively. The target loss is calculated using the mean square error, and the formula is as follows: ； in, Indicates target loss. This represents the predicted representation obtained by sequentially passing the latent space features of the current sample through a projection network and a prediction network. This represents the target representation obtained by applying a projection network to the latent space features of the offset samples. The formula for calculating the reconstruction loss is: ； in, Indicates the reconstruction loss. This represents the spectral regularization loss function; The formula for calculating the final loss function is as follows: ； ； in, This represents the final loss function. Indicates sample Intermediate losses, The weights are used to balance the target loss and the reconstruction loss. The global weights are the intermediate losses.

7. The test-time adaptive method for predicting the ash content of clean coal based on the implicit space concept drift assessment according to claim 1, characterized in that, In step 6, after the model training is completed, the final model obtained by training is evaluated using a test set, and the model with the best performance is applied to the actual heavy media coal preparation process to achieve accurate prediction of the ash content of clean coal.

Images