Machine learning-based gangue filling slurry fluidity prediction method

Abstract

This invention relates to the fields of mining engineering and intelligent manufacturing technology, and discloses a machine learning-based method for predicting the flowability of gangue backfill slurry. The method includes the following steps: collecting physicochemical characteristic data of different batches of gangue samples, determining the minimum flow concentration, maximum flow concentration, and optimal flow concentration of the backfill slurry for each sample, and constructing an initial training dataset; constructing a multi-layer neural network model and training the model offline using the initial training dataset; obtaining physicochemical characteristic data of the new batch of gangue to be prepared, and inputting it into the trained model to obtain the corresponding flow concentration prediction result. By collecting multi-dimensional physicochemical characteristics of gangue to construct a training dataset, and relying on a multi-layer neural network model to capture the complex nonlinear relationship between characteristics and slurry flowability, this method overcomes the limitations of traditional experience-dependent prediction methods and effectively avoids prediction bias caused by fluctuations in gangue characteristics.

Description

Technical Field

[0001] This invention relates to the fields of mining engineering and intelligent manufacturing technology, specifically to a machine learning-based method for predicting the fluidity of gangue backfill slurry. Background Technology

[0002] Coal mine gangue, as a major solid waste generated during coal mining, is prone to environmental risks such as land encroachment, ecological damage, spontaneous combustion, and leaching pollution when stored for a long time. Its resource utilization for underground backfilling has become a key technical path to practice the concept of green mining and control surface subsidence. The fluidity of gangue backfill slurry is the core indicator that determines the stability of the backfilling process. It is directly related to the pumpability of the slurry, the safety of pipeline transportation, and the final strength of the backfill. Once the fluidity is out of control, it will lead to slurry segregation at best, and pipeline blockage or insufficient strength of the backfill at worst, which will seriously affect the mine's production efficiency and safety.

[0003] Currently, the industry mainly relies on three types of technical solutions for controlling and predicting slurry flowability. The first type is the traditional experience-based method, which presets slurry parameters based on historical data under fixed gangue sources or specific working conditions. However, in actual production, changes in gangue sources, extension of mining layers, or adjustments to crushing processes can cause significant fluctuations in chemical composition and particle size characteristics. This type of method cannot dynamically adapt to these changes and can only adjust parameters through repeated trial and error, which is not only inefficient but also prone to material waste and engineering risks. The second type is a technical solution that focuses on local optimization. Some technologies improve the flowability of specific batches of slurry by adding admixtures, but they have not established a universal predictive model between gangue characteristics and flowability, and therefore cannot provide a predictive model for slurry with variable characteristics. One type provides general guidance for gangue, while another type designs a dedicated pulping system. However, its flowability control still relies on preset solid-liquid ratios and manual fine-tuning based on experience, lacking intelligent response capabilities to dynamic changes in gangue characteristics. The third type is optimization technology driven by experiments or simple models. Some technologies optimize the slurry ratio through laboratory experiments to ensure flowability, but they suffer from long experimental cycles and cumbersome operations, making it difficult to meet the real-time requirements of industrial production. Other technologies introduce multi-source information sensing for flowability control, but their core models have limited ability to characterize the complex nonlinear relationship between the multi-dimensional characteristics of gangue and flowability, and lack a self-learning mechanism. When new fluctuation patterns appear in gangue characteristics, the model cannot adjust itself to maintain prediction accuracy.

[0004] These existing technologies share three major drawbacks: First, they cannot effectively capture the complex coupling relationship between the physicochemical properties of gangue and the fluidity of slurry, resulting in insufficient prediction accuracy. Second, they lack adaptive capabilities, making it difficult to cope with dynamic fluctuations in gangue properties, especially unsuitable for large mines with multiple working faces and multiple sources of gangue. Third, they lack real-time performance and practicality, either relying on cumbersome experiments or failing to achieve continuous model optimization. Ultimately, this leads to frequent risks of fluidity loss during gangue backfilling, restricting the efficiency of gangue resource utilization and the intelligent upgrading of mine backfilling processes. Therefore, there is an urgent need for an intelligent slurry fluidity prediction method that can deeply integrate the multi-dimensional characteristics of gangue, possess high-precision prediction and continuous self-optimization capabilities, in order to overcome the bottlenecks of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a machine learning-based method for predicting the fluidity of gangue backfill slurry, thus solving the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based method for predicting the fluidity of gangue backfill slurry, comprising the following steps: Physicochemical characteristic data of different batches of gangue samples were collected to determine the minimum flow concentration, maximum flow concentration and optimal flow concentration of the filling slurry corresponding to each sample, and to construct an initial training dataset. Construct a multi-layer neural network model and train it offline using the initial training dataset; Obtain the physicochemical characteristic data of the new batch of gangue to be prepared, input it into the trained model to obtain the corresponding flow concentration prediction results; Based on the prediction results, the slurry is prepared and its actual flow properties are monitored. The relevant data from the new batch is used as a sample for incremental training of the model. Based on the updated model, the flow concentration range of subsequent batches of slurry is recommended to guide the adjustment of slurry preparation parameters.

[0007] Preferably, the physicochemical characteristic data includes chemical composition data and particle size characteristic data. The chemical composition data includes the mass fractions of silicon dioxide, aluminum oxide, calcium oxide, and carbon, and the particle size characteristic data includes the average particle size. The data are obtained by a component analyzer and a particle size analyzer.

[0008] Preferably, the initial training dataset uses physicochemical characteristic data as input features and minimum flow concentration, maximum flow concentration, and optimal flow concentration as output targets. Abnormal fluctuation data are removed during the data acquisition process to ensure the integrity and effectiveness of the dataset.

[0009] Preferably, the neural network model includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is consistent with the number of feature types in the physicochemical feature data. The number of nodes in the output layer is three, corresponding to the minimum flow concentration, the maximum flow concentration, and the optimal flow concentration, respectively. The number of nodes in the hidden layer is determined by cross-validation. The model uses a Sigmoid or ReLU activation function.

[0010] Preferably, during model training, the initial training dataset is divided into a training set, a validation set, and a test set, and the Levenberg-Marquardt, Adam, or RMSprop optimization algorithm is used. The training termination condition is set to reach a preset number of iterations or the model prediction error is lower than a set threshold.

[0011] Preferably, the incremental training adopts a transfer learning or online learning strategy. The new sample data includes physicochemical feature data, actual preparation concentration, and actual flow performance parameters. During training, a small learning rate is used to fine-tune the model weights, and the model performance is monitored through a validation set to avoid overfitting.

[0012] Preferably, the optimal flow concentration is determined by comprehensively considering slurry flowability, pumpability, filling strength, material cost, and energy consumption, and the recommended concentration range is dynamically adjusted based on model prediction results and actual working conditions.

[0013] Preferably, the physicochemical characteristic data of the new batch of gangue are collected before pulping, and the collected indicators are consistent with the characteristic indicators of the initial training dataset. The prediction results are output in the form of a clear concentration range.

[0014] A machine learning-based system for predicting the flowability of gangue-filled slurry includes a data acquisition module, a model training module, a prediction module, a feedback optimization module, and a parameter guidance module. The data acquisition module is used to detect and store physicochemical characteristic data of gangue and slurry flow performance data. The model training module is used to construct a neural network model and perform offline training and incremental training. The prediction module is used to load the trained model and output the predicted flow concentration. The feedback optimization module is used to collect actual slurry preparation data and trigger incremental training of the model. The parameter guidance module is used to generate slurry parameter adjustment suggestions based on the prediction results.

[0015] This invention provides a machine learning-based method for predicting the flowability of gangue backfill slurry. It has the following beneficial effects: 1. This invention constructs a training dataset by collecting multi-dimensional physicochemical characteristics of gangue, and relies on a multi-layer neural network model to capture the complex nonlinear relationship between the characteristics and the fluidity of the slurry. It breaks away from the limitations of traditional experience-dependent prediction methods, effectively avoids prediction deviations caused by fluctuations in gangue characteristics, ensures that the predicted flow concentration range can accurately match the requirements for stable slurry delivery and filling, and reduces engineering problems such as pipe blockage and insufficient filling strength caused by uncontrolled fluidity. It provides key support for the stable operation of the filling process.

[0016] 2. This invention, through an incremental training mechanism combined with transfer learning or online learning strategies, can continuously integrate new batch data generated during the actual pulping process into model updates. It eliminates the need to conduct complex experiments or reconstruct the model for fluctuating gangue raw materials, and can quickly adapt to the dynamic changes in gangue characteristics under different origins, mining layers, or crushing processes. It is especially suitable for mining scenarios with multiple working faces and multiple sources of gangue, greatly improving the adaptability of the backfilling process to raw material fluctuations and ensuring the stability of the prediction effect in long-term production.

[0017] 3. This method combines high practical value with intelligent empowerment. On the one hand, it determines the optimal flow concentration by comprehensively considering slurry fluidity, pumpability, filling strength, material cost, and energy consumption, and outputs clear recommended concentration ranges and slurry parameter adjustment suggestions, directly reducing on-site trial and error costs and improving slurry efficiency and filling operation safety. On the other hand, it relies on a systematic modular design to achieve intelligent connection of the entire process from data input to parameter guidance, promoting the intelligent upgrading of coal mine gangue resource utilization and mine filling technology, taking into account engineering performance, economic and environmental benefits, and meeting the needs of green mine construction. Attached Figure Description

[0018] Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a schematic diagram of the five-factor neural network prediction model for gangue slurry concentration of the present invention; Figure 3 This is a schematic diagram illustrating the correlation distribution of various variables in an embodiment of the present invention; Figure 4 This is a schematic diagram of neural network prediction training and testing according to the present invention; Figure 5 This is a comparison chart of the actual and predicted values ​​for each sample in this invention; Figure 6 This is a comparison chart of the predicted values ​​and actual responses of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see the appendix Figure 1 - Appendix Figure 6 This invention provides a machine learning-based method for predicting the flowability of gangue backfill slurry, comprising the following steps: Physicochemical characteristic data of different batches of gangue samples were collected to determine the minimum flow concentration, maximum flow concentration and optimal flow concentration of the filling slurry corresponding to each sample, and to construct an initial training dataset. The physicochemical feature data includes chemical composition data and average particle size. The chemical composition data includes the mass fractions of silicon dioxide, aluminum oxide, calcium oxide, and carbon. The particle size data includes the average particle size. The data was acquired using a component analyzer and a particle size analyzer. The initial training dataset uses the physicochemical feature data as input features and sets the minimum flow concentration, maximum flow concentration, and optimal flow concentration as output targets. Abnormal fluctuations were removed during data acquisition to ensure the integrity and validity of the dataset.

[0021] Specifically, the initial dataset used in this study comes from field practice data of gangue backfill slurry from typical mines in five regions: Shaanxi, Guangxi, Henan, and Shandong. Therefore, Table 1 contains five sets of sample data, corresponding to five different types of gangue. This study uses the mass fraction of the main chemical components of gangue (SiO2, Al2O3, CaO) and the average particle size as the main parameters for detection and analysis. These indicators were chosen as the model input layer based on the following two considerations: First, SiO2, Al2O3, and CaO are the main chemical components of gangue, which have a decisive influence on its physicochemical properties and backfilling performance; second, gangue is usually classified according to the SiO2 and Al2O3 content and the Al2O3 / SiO2 ratio, which can classify coal gangue into four types: aluminous, claystone, sandstone, and calcareous gangue. Furthermore, gangue is a carbon-based rock, containing a certain proportion of carbon, which is an important characteristic that distinguishes it from other solid waste ores. Therefore, in addition to the analysis of the main chemical components, the influence of carbon content also needs to be considered. This study conducted a loss on ignition test on the samples according to the national standard GB / T35986-2018 "Determination of Loss on Ignition of Coal Gangue", and calculated the carbon content of the gangue accordingly.

[0022] The concentrations of the slurry formed from gangue were collected through literature and reports and summarized in Table 1, forming the initial dataset. The collected physicochemical characteristics of the gangue were used as input features, and the corresponding minimum and maximum slurry concentrations were used as output targets to construct the initial training dataset for the BP neural network. The final initial training dataset of the collected gangue slurry formation data is shown in Table 1.

[0023] Table 1 Dataset for Gangue Slurry Formation Model

[0024]

[0025] A multi-layer neural network model was constructed and trained offline using the initial training dataset. The neural network model consists of an input layer, hidden layers, and an output layer. The number of nodes in the input layer corresponds to the number of feature types in the physicochemical feature data. The output layer has three nodes, corresponding to the minimum flow concentration, maximum flow concentration, and optimal flow concentration, respectively. The number of nodes in the hidden layer was determined using cross-validation. The model employs either a Sigmoid or ReLU activation function. During model training, the initial training dataset was divided into training, validation, and test sets. Levenberg-Marquardt, Adam, or RMSprop optimization algorithms were used. The training termination condition was set to reach a preset number of iterations or for the model's prediction error to fall below a set threshold.

[0026] Specifically, a multi-layer BP neural network model is established, with the number of input layer nodes equal to the number of physicochemical characteristic factors of gangue collected in step 1 (e.g., SiO2, Al2O3, CaO, C content, average particle size, etc.), and the number of output layer nodes is 3 (corresponding to the predicted minimum, maximum, and optimal slurry concentrations). Before model training, data standardization and correlation analysis are performed on the initial dataset; after data analysis, the correlation statistics between the data variables are as follows: Figure 3 As shown in the figure, the scatter plot and correlation coefficient analysis reveal that the correlation between the predictor variables is low. The highest correlation is between SiO2 and CaO, at -0.6817. The low correlation between the variable data makes it suitable for model training.

[0027] The learning and training process of a backpropagation (BP) neural network consists of two phases: forward propagation and backward propagation. During forward propagation, sample data is passed from the input layer to the output layer through the transfer function of the hidden layers. If the output layer does not produce the desired output, the backpropagation process begins. The error signal is returned along the original forward propagation path, and the network connection weights are modified using mean squared error and gradient descent to adjust the mean squared error between the actual network output and the guiding learning signal. This process is repeated until the specified error requirement is met or the maximum number of training iterations is reached. Figure 4 The basic principle of the neural network model is shown. The initial training dataset is divided into training and test sets, and the Levenberg-Marquardt algorithm is used to train the BP neural network model offline. During training, the learning rate, maximum number of iterations, and training objective are set, and the model performance is monitored through the validation set to prevent overfitting. The initial data used for the model comes from the experimental data in Table 1, and a portion of the samples are randomly selected proportionally as the training set, while the remaining samples are used for model testing and validation.

[0028] Twenty samples were selected for training the neural network; the remaining five samples were used as a test set to test the model's generalization ability. The processed sample data were input into the established neural network model. For the 20 samples in the training set, a 5-fold cross-validation was used to create a validation set to test the model and prevent overfitting. Specific steps are as follows: Figure 4 As shown.

[0029] In this invention, the number of training iterations for the network is fixed at 1000, the number of neurons in the hidden layer of the neural network is set to 10, the learning rate is 0.1, and the training target is specified as 0.01. Using the content of the main components (SiO2, Al2O3, CaO, C) and the average particle size as the five input factors of the neural network, a five-factor BP neural network model is established to predict the slurry concentration of gangue. The prediction results are as follows: Figure 5 As shown.

[0030] Figure 5 This is a comparison chart of the predicted and actual values ​​of gangue slurry concentration. The chart shows that the model's predictions generally reflect the fluctuations in the actual values ​​well, except for sample 11 where the predicted and observed values ​​have relatively large errors. To further verify the effectiveness of the neural network, experimental and predicted values ​​are compared, see... Figure 6 Most of the points fall on or are distributed on both sides of the 45° line, indicating that the predicted values ​​and experimental values ​​are in good agreement.

[0031] The physicochemical characteristics data of the new batch of gangue to be prepared are obtained and input into the trained model to obtain the corresponding flow concentration prediction results. The optimal flow concentration is determined by comprehensively considering the slurry flowability, pumpability, filling strength, material cost and energy consumption. The recommended concentration range is dynamically adjusted according to the model prediction results and actual working conditions.

[0032] Specifically, a coal mine in northern Shaanxi is located in the Yushen mining area of ​​the Jurassic coalfield in northern Shaanxi. It is a mine under construction with a relatively simple geological structure. During coal seam mining, the generated gangue mainly consists of mudstone, silty mudstone, and argillaceous siltstone. Therefore, the overall lithological characteristics of the mine's gangue are dominated by mudstone, carbonaceous mudstone, silty mudstone, and argillaceous siltstone. In the actual backfilling process, the mine did not distinguish between gangue of different properties, but instead mixed all the gangue and crushed them uniformly before using them for backfilling. It is worth noting that before the backfilling work began, a systematic chemical composition test was conducted on various types of gangue. The specific test results are shown in Table 2.

[0033]

[0034] Table 2 (Main components and proportions of gangue) The chemical composition and different particle size composition in the batch feature data are transmitted to the BP neural network model that has completed initial training. The predicted minimum and maximum pulping concentration ranges of the batch of gangue are shown in Table 3.

[0035]

[0036] Table 3 (Predicted Initial Concentration Range) Actual preparation and performance monitoring: Based on the predicted concentration range, this batch of gangue slurry was prepared, and the slurry concentration was verified using methods such as rheometer and slump test. The results show that gangue slurry within this range can meet the filling requirements. Slightly lower concentrations will result in particle sedimentation, making slurry formation difficult and causing segregation. Slightly higher concentrations will result in a slump below the maximum flow slump value of 160mm, leading to insufficient flowability and failure to meet pumping requirements. This result not only verifies the accuracy of the predicted concentration but also highlights the reliability of this invention in ensuring the working performance of the slurry, providing strong support for achieving precise control of gangue slurry.

[0037] Based on the predicted results, the slurry is prepared and its actual flow properties are monitored. New batch data is used as samples for incremental training of the model. The updated model is then used to recommend flow concentration ranges for subsequent batches of slurry, guiding adjustments to the slurry preparation parameters. Incremental training employs transfer learning or online learning strategies. The new sample data includes physicochemical characteristics, actual prepared concentrations, and actual flow property parameters. A small learning rate is used to fine-tune the model weights during training, and the model performance is monitored using a validation set to avoid overfitting.

[0038] Recommended dynamic concentration range: Subsequently, during the filling and pulping process at the mine, the pulping records over a period of time, guided by this model, are shown in Table 5 below:

[0039]

[0040] Table 5 (Prediction of Gangue Slurry Concentration in a Mine in Northern Shaanxi) It can be seen that during the long-term backfilling process of gangue slurry, both the gangue composition and the particle size of the backfill material fluctuate significantly. This means that without real-time adjustments, it is difficult to maintain a stable backfilling effect, which may lead to problems such as insufficient backfill strength, segregation and pipe blockage, or resource waste. By adopting our proposed concentration prediction model, the backfill concentration of the mine can be adjusted in real time to ensure the uniformity and stability of the backfill material, effectively improve the backfill strength, avoid pipe blockage, maximize the utilization efficiency of gangue, and ultimately ensure the safe and efficient production of the mine.

[0041] Physicochemical characteristic data of the new batch of gangue were collected before pulping. The collected indicators were consistent with the characteristic indicators of the initial training dataset, and the prediction results were output in the form of a clear concentration range.

[0042] A machine learning-based system for predicting the flowability of gangue-filled slurry includes a data acquisition module, a model training module, a prediction module, a feedback optimization module, and a parameter guidance module. The data acquisition module is used to detect and store physicochemical characteristic data of gangue and slurry flow performance data. The model training module is used to construct a neural network model and perform offline training and incremental training. The prediction module is used to load the trained model and output the predicted flow concentration. The feedback optimization module is used to collect actual slurry preparation data and trigger incremental training of the model. The parameter guidance module is used to generate slurry parameter adjustment suggestions based on the prediction results.

[0043] Specifically, the core principle of this prediction system is to construct a data-driven closed-loop prediction and optimization mechanism through machine learning technology to achieve accurate prediction of the flowability of gangue-filled slurry and intelligent guidance of slurry preparation parameters. First, the data acquisition module captures and stores basic data on the physicochemical characteristics of gangue and the corresponding slurry flowability, providing training materials for the model. The model training module builds a neural network model based on this data, completes the initial model construction through offline training, and supports subsequent incremental training to adapt to data changes. The prediction module calls the well-trained model, inputs new gangue-related data, and directly outputs the flow concentration prediction result. The feedback optimization module continuously collects real data from the actual slurry preparation process, compares it with the prediction results, and triggers incremental model training to continuously improve the model's prediction accuracy. The parameter guidance module transforms the prediction results into specific slurry preparation parameter adjustment suggestions, which in turn guide the actual slurry preparation operation, forming a complete workflow of data acquisition, model training, prediction output, feedback optimization, and parameter guidance, realizing efficient prediction of slurry flowability and dynamic optimization of the slurry preparation process.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based method for predicting the fluidity of gangue backfill slurry, characterized in that, Includes the following steps: Physicochemical characteristic data of different batches of gangue samples were collected to determine the minimum flow concentration, maximum flow concentration and optimal flow concentration of the filling slurry corresponding to each sample, and to construct an initial training dataset. Construct a multi-layer neural network model and train it offline using the initial training dataset; Obtain the physicochemical characteristic data of the new batch of gangue to be prepared, input it into the trained model to obtain the corresponding flow concentration prediction results; Based on the prediction results, the slurry is prepared and its actual flow properties are monitored. The relevant data from the new batch is used as a sample for incremental training of the model. Based on the updated model, the flow concentration range of subsequent batches of slurry is recommended to guide the adjustment of slurry preparation parameters.

2. The method for predicting the fluidity of gangue backfill slurry based on machine learning according to claim 1, characterized in that, The physicochemical characteristic data includes chemical composition data and particle size characteristic data. The chemical composition data includes the mass fractions of silicon dioxide, aluminum oxide, calcium oxide, and carbon, and the particle size characteristic data includes the average particle size. The data are obtained by a component analyzer and a particle size analyzer.

3. The method for predicting the fluidity of gangue backfill slurry based on machine learning according to claim 1, characterized in that, The initial training dataset uses physicochemical characteristic data as input features and minimum flow concentration, maximum flow concentration, and optimal flow concentration as output targets. Abnormal fluctuation data are removed during the data acquisition process to ensure the integrity and effectiveness of the dataset.

4. The method for predicting the fluidity of gangue backfill slurry based on machine learning according to claim 1, characterized in that, The neural network model includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is the same as the number of feature types in the physicochemical feature data. The number of nodes in the output layer is three, corresponding to the minimum flow concentration, the maximum flow concentration, and the optimal flow concentration, respectively. The number of nodes in the hidden layer is determined by cross-validation. The model uses either a Sigmoid or ReLU activation function.

5. The method for predicting the fluidity of gangue backfill slurry based on machine learning according to claim 1, characterized in that, During model training, the initial training dataset is divided into a training set, a validation set, and a test set. The Levenberg-Marquardt, Adam, or RMSprop optimization algorithms are used, and the training termination condition is set to reach a preset number of iterations or the model prediction error is lower than a set threshold.

6. The method for predicting the fluidity of gangue backfill slurry based on machine learning according to claim 1, characterized in that, The incremental training employs transfer learning or online learning strategies. The new sample data includes physicochemical characteristic data, actual prepared concentrations, and actual flow performance parameters. During training, a small learning rate is used to fine-tune the model weights, and the model performance is monitored through a validation set to avoid overfitting.

7. The method for predicting the fluidity of gangue backfill slurry based on machine learning according to claim 1, characterized in that, The optimal flow concentration is determined by comprehensively considering slurry flowability, pumpability, filling strength, material cost, and energy consumption. The recommended concentration range is dynamically adjusted based on model prediction results and actual working conditions.

8. The method for predicting the fluidity of gangue backfill slurry based on machine learning according to claim 1, characterized in that, The physicochemical characteristic data of the new batch of gangue were collected before pulping. The collected indicators were consistent with the characteristic indicators of the initial training dataset, and the prediction results were output in the form of a clear concentration range.

9. A machine learning-based system for predicting the flowability of gangue backfill slurry, used in the machine learning-based method for predicting the flowability of gangue backfill slurry as described in any one of claims 1-8, characterized in that, It also includes a data acquisition module, a model training module, a prediction module, a feedback optimization module, and a parameter guidance module. The data acquisition module is used to detect and store the physicochemical characteristics data of gangue and the flow properties data of slurry. The model training module is used to build a neural network model and perform offline training and incremental training. The prediction module is used to load the trained model and output the flow concentration prediction results. The feedback optimization module is used to collect actual pulping data and trigger incremental training of the model. The parameter guidance module is used to generate pulping parameter adjustment suggestions based on the prediction results.

Images