A Machine Learning-Based Method and System for Evaluating the Treatment Effect of Mine Backfilling Materials

Abstract

This invention discloses a machine learning-based method and system for evaluating the effectiveness of mine backfill material remediation, relating to the field of mining engineering. The method includes: collecting historical multi-source data on the mine backfill material remediation process and preprocessing it to obtain a training sample set; classifying the variables in the training sample set into causal variables, intermediate variables, and outcome variables; employing a causal discovery algorithm to learn the causal structure of the causal variables, intermediate variables, and outcome variables, generating a causal graph structure in the form of a directed acyclic graph (DAG); and constructing a hierarchical network connection architecture based on the variable connections determined in the DAG causal graph structure. This invention utilizes the trained model to evaluate and predict real-time data, and generates a quantified causal path and an interpretable evaluation report based on the causal graph and model parameters.

Description

Technical Field

[0001] This invention relates to the field of mine backfilling remediation engineering, and in particular to a method and system for evaluating the remediation effect of mine backfilling materials based on machine learning. Background Technology

[0002] In the field of mining engineering, underground mining activities will create a large number of goaf areas, which can easily cause ground pressure disasters such as surrounding rock instability and surface collapse. Goaf filling treatment is currently the most mainstream treatment method: solid waste, cement and water are mixed to form a filling slurry, which is transported to the underground goaf area. After solidification and curing, it forms a filling body with a certain strength, which plays a supporting and controlling role for the surrounding rock strata.

[0003] The mechanical properties of the backfill directly determine the effectiveness of goaf backfilling treatment. Traditional quality inspection methods rely on destructive testing such as core drilling and point load tests after the backfill has cured, which has inherent limitations such as information lag, high cost, and inability to reflect the entire process of hydration reaction and strength evolution. With the advancement of sensing technology, distributed temperature measurement, acoustic monitoring, and other methods can acquire real-time multi-physics time-series data inside the backfill. To handle high-dimensional heterogeneous data, data-driven methods based on machine learning have become the mainstream technology. The existing solution uses a deep learning model to establish a nonlinear mapping from process monitoring data to final performance indicators, enabling rapid and non-destructive prediction of the goaf backfilling treatment effect and improving evaluation efficiency.

[0004] Current machine learning-based evaluation methods essentially establish complex correlation models, but their internal decision-making processes lack interpretability. While they can achieve high predictive accuracy, they fail to reveal the intrinsic causal mechanisms, from causal variables such as material proportions and process parameters, to intermediate process variables such as hydration reactions and microstructure evolution, and finally to mechanical properties. When evaluation results are abnormal, existing technologies struggle to provide a clear path analysis to trace the root cause of performance defects, failing to effectively guide engineers in targeted and precise optimization of filling proportions or construction processes. This limits the decision support value of evaluation results in closed-loop quality control. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a machine learning-based method for evaluating the treatment effect of mine backfill materials, which solves the technical problem that existing evaluation methods lack interpretability and cannot reveal the inherent causal mechanism of treatment effect, making it difficult for evaluation results to guide precise process optimization.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for evaluating the treatment effect of mine backfill materials based on machine learning, which includes collecting historical multi-source data of the mine backfill material treatment process and preprocessing it to obtain a training sample set of the mine backfill material treatment process; The variables in the training sample set are divided into causal variables, intermediate variables and outcome variables. A causal discovery algorithm is used to learn the causal structure of the causal variables, intermediate variables and outcome variables, and generate a causal graph structure in the form of a directed acyclic graph. Based on the variable connection relationships determined in the causal graph structure in the form of a directed acyclic graph, a hierarchical network connection architecture is constructed. A physical information causal reasoning network model is constructed on a hierarchical network connection architecture, and the physical information causal reasoning network model is trained using the training sample set; Real-time multi-source data of the mine backfill material treatment process to be evaluated is input into a trained physical information causal reasoning network model to obtain evaluation prediction values. Based on the causal graph structure in the form of a directed acyclic graph and the parameters of the physical information causal reasoning network model, a causal path and evaluation explanation report are generated.

[0008] As a preferred embodiment of the machine learning-based method for evaluating the effectiveness of mine backfill material remediation as described in this invention, the training sample set for the mine backfill material remediation process includes: Historical multi-source data on the treatment process of mine backfill materials were collected and standardized and normalized to obtain a training sample set of the treatment process of mine backfill materials.

[0009] As a preferred embodiment of the machine learning-based method for evaluating the treatment effect of mine backfill materials according to the present invention, the causal graph structure in the form of a directed acyclic graph includes, Based on the physical meaning of each variable in the training sample set and the temporal stages in the governance process, the variables in the training sample set are clearly divided into causal variables, intermediate variables and outcome variables. The algorithm for causal discovery combines causal variables, intermediate variables, and outcome variables with the constraints of physical rules governing time irreversibility and mass-energy conservation. It then tests the statistical independence among the causal variables, intermediate variables, and outcome variables, applies physical rules, and infers the causal direction between the variables. Based on the causal direction between variables, causal structure learning is performed on causal variables, intermediate variables and outcome variables to determine the causal connection topology between variables; Based on the causal connection topology between variables, a causal graph structure in the form of a directed acyclic graph is generated, which includes causal directions and causal strength weights.

[0010] As a preferred embodiment of the machine learning-based method for evaluating the treatment effect of mine backfill materials according to the present invention, the hierarchical network connection architecture includes, Based on the topological sorting depth of nodes in a causal graph structure in the form of a directed acyclic graph, the network hierarchy is identified. Based on the network hierarchy, the input layer, hidden layer and output layer of the hierarchical network connection architecture are defined. According to the specific causal connections in the causal graph structure in the form of a directed acyclic graph, neuron connections are established between the corresponding layers of the hierarchical network connection architecture. Based on the causal graph structure in the form of a directed acyclic graph, initial weight constraints are assigned to the neuron connections already established in the hierarchical network connection architecture. Based on the established neuron connections with initial weight constraints, a hierarchical network connection architecture is obtained.

[0011] As a preferred embodiment of the machine learning-based method for evaluating the treatment effect of mine backfill materials according to the present invention, the training of the physical information causal reasoning network model includes, In the hierarchical network connection architecture, the network layers and neuron activation functions of the physical information causal reasoning network model are defined; The physical consistency constraint loss term is obtained by the physical rule calculation module based on the time-varying sequence of wave velocity, time-varying sequence of attenuation coefficient, time-varying sequence of hydration heat release rate and time-varying sequence of equivalent porosity index in the training sample set. Based on the rational information causal reasoning network model, a total loss function is constructed, consisting of a data fitting loss term, a physical consistency constraint loss term, and a causal structure regularization loss term. The total loss function and the training sample set of the mine backfill material treatment process are input into the optimization algorithm. The parameters of the physical information causal inference network model are updated iteratively by minimizing the total loss function, and the total loss function value and prediction performance on the validation set are monitored. When the total loss function value drops below a preset threshold and the fluctuation range of the validation set prediction performance stabilizes within a preset range, the trained physical information causal reasoning network model is obtained.

[0012] As a preferred embodiment of the machine learning-based method for evaluating the treatment effect of mine backfill materials according to the present invention, the evaluated predicted values ​​include: Real-time multi-source data of the mine backfill material treatment process to be evaluated is collected and preprocessed in the same way as in the training phase to obtain preprocessed real-time multi-source data. The preprocessed real-time multi-source data is input into the trained physical information causal reasoning network model for forward propagation calculation to obtain the evaluation prediction value.

[0013] As a preferred embodiment of the machine learning-based method for evaluating the treatment effect of mine backfill materials according to the present invention, the generation of the causal path and evaluation explanation report includes, Based on the connection weight parameters of the causal graph structure in the form of a directed acyclic graph and the trained physical information causal reasoning network model, the causal path from the causal variable to the evaluation prediction value is traced back. Based on the causal path, calculate the quantitative impact weight of each causal variable in the causal path on the evaluation prediction value; Integrate causal paths with corresponding quantitative impact weights to generate causal path and assessment explanation reports.

[0014] Secondly, the present invention provides a machine learning-based system for evaluating the treatment effect of mine backfill materials, including a processing module that collects historical multi-source data on the treatment process of mine backfill materials and performs preprocessing to obtain a training sample set of the treatment process of mine backfill materials. The learning module classifies the variables in the training sample set into causal variables, intermediate variables, and outcome variables. It then uses a causal discovery algorithm to learn the causal structure of the causal variables, intermediate variables, and outcome variables, generating a causal graph structure in the form of a directed acyclic graph. The building module constructs a hierarchical network connection architecture based on the variable connection relationships determined in the causal graph structure in the form of a directed acyclic graph. The training module constructs a physical information causal reasoning network model on a hierarchical network connection architecture and trains the physical information causal reasoning network model using the training sample set. The prediction module inputs real-time multi-source data of the mine backfill material treatment process to be evaluated into a trained physical information causal reasoning network model to obtain the evaluation prediction value. Based on the causal graph structure in the form of a directed acyclic graph and the parameters of the physical information causal reasoning network model, it generates the causal path and evaluation explanation report.

[0015] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the machine learning-based method for evaluating the treatment effect of mine backfill materials as described in the first aspect of the present invention.

[0016] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning-based method for evaluating the treatment effect of mine backfill materials as described in the first aspect of the present invention.

[0017] The beneficial effects of this invention are as follows: by collecting and preprocessing historical multi-source data to construct a training sample set; then, by integrating a causal discovery algorithm with physical rule constraints, the causal structure between causal variables, intermediate variables and outcome variables in the sample is learned, generating a weighted directed acyclic graph structure; based on this causal graph structure, a hierarchical network connection architecture is constructed; and on this basis, a causal inference network model integrating physical consistency constraints is constructed for training; the trained model is used to evaluate and predict real-time data; and based on the causal graph and model parameters, a quantitative causal path and an interpretable evaluation report are generated by backtracking. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a machine learning-based method for evaluating the effectiveness of mine backfill material remediation.

[0020] Figure 2 This is a schematic diagram of a machine learning-based system for evaluating the effectiveness of mine backfill material remediation. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "an embodiment" or "an embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The phrase "an embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Reference Figures 1-2 This is one embodiment of the present invention, which provides a machine learning-based method for evaluating the treatment effect of mine backfill materials, including the following steps: S1. Collect historical multi-source data on the treatment process of mine backfill materials, and preprocess the data to obtain a training sample set of the treatment process of mine backfill materials.

[0025] S1.1 Collect historical multi-source data on the treatment process of mine backfill materials, and perform standardization and normalization processing to obtain a training sample set of the treatment process of mine backfill materials.

[0026] Furthermore, historical multi-source data on the mine backfill material treatment process are read, and missing value imputation and outlier removal operations are performed on the historical multi-source data on the mine backfill material treatment process. Standardization processing is performed on each feature dimension of the historical multi-source data on the mine backfill material treatment process to unify the units of measurement, and normalization processing is performed on each feature dimension of the historical multi-source data on the mine backfill material treatment process to map their values ​​to a unified interval. The standardized and normalized historical multi-source data on the mine backfill material treatment process are then sorted and merged to obtain the training sample set of the mine backfill material treatment process.

[0027] S2. The variables in the training sample set are divided into causal variables, intermediate variables and result variables. A causal discovery algorithm is used to learn the causal structure of the causal variables, intermediate variables and result variables, and generate a causal graph structure in the form of a directed acyclic graph.

[0028] S2.1 Based on the physical meaning of each variable in the training sample set and the temporal stages that occur during the governance process, the variables in the training sample set are clearly divided into causal variables, intermediate variables and outcome variables.

[0029] Furthermore, based on the engineering meaning of each variable in the training sample set for the mine backfill material remediation process and the temporal order in which the variable observations were recorded during the remediation process, the variables in the training sample set for the mine backfill material remediation process are divided into three categories. Causal variables are those set before or at the start of the remediation operation and are considered to be actively adjustable or used as input conditions; their observations do not depend on other variables in the remediation process. Intermediate variables are those obtained through continuous or intermittent monitoring during the remediation process, reflecting the dynamic evolution of the internal state of the backfill body; their observations are influenced by causal variables and may affect outcome variables. Outcome variables are those obtained through testing after the remediation process is completed, used to characterize the final performance of the backfill body; their observations are the output of the entire remediation process. By performing the above-mentioned determination of physical meaning and temporal order on each variable in the training sample set for the mine backfill material remediation process, the classification of causal variables, intermediate variables, and outcome variables is completed.

[0030] Specifically, the prior physical meaning and time stage division introduce strong guidance for causal direction search and semantic constraints on variable roles into the subsequent causal discovery algorithm. For example, the division ensures that the peak time of hydration heat release is not mistakenly identified as the cause of the initial parameter of cement usage, because the former is later in time than the latter. The division not only reduces the search complexity of causal learning, but also shifts the learning objective from discovering statistical associations between arbitrary variables to revealing specific transmission mechanisms from cause-effect processes to effects. S2.2. Input the causal discovery algorithm that integrates the causal variables, intermediate variables and result variables with the constraints of the physical rules of time irreversibility and mass-energy conservation, test the statistical independence among the causal variables, intermediate variables and result variables and apply physical rule constraints, and infer the causal direction between the variables.

[0031] Furthermore, all data corresponding to the causal, intermediate, and outcome variables that have been segmented are input into a causal discovery algorithm that incorporates physical rule constraints. The algorithm's execution process comprises two core stages: statistical independence testing and application of physical rule constraints. In the statistical independence testing stage, the algorithm performs a conditional independence hypothesis test on any two variables among the causal, intermediate, and outcome variables, given one or more other subsets of variables, obtaining probability values ​​indicating whether the two variables are independent under given conditions. In the physical rule constraint application stage, the algorithm reviews and corrects the statistical test results based on embedded domain physical rules. Specific applied physical rules include the time irreversibility rule and the mass-energy conservation rule.

[0032] Specifically, the time irreversibility rule states that if the observation time of variable A is always no later than the observation time of variable B, then it is forbidden to infer that variable B is the cause of variable A. The mass-energy conservation rule states that in relationships involving mass or energy flow, the causal variable must be upstream in the direction of energy or mass transfer. When the causal direction inferred from the statistical independence test conflicts with any of the above physical rules, that causal direction will be rejected, and the direction that conforms to the physical rule constraints will be adopted, or it will be determined that there is no direct causal direction. Through iterative execution of the testing and constraint process, the inferred causal direction for each pair of variables is output.

[0033] S2.3. Based on the causal direction between variables, perform causal structure learning on causal variables, intermediate variables and outcome variables to determine the causal connection topology between variables.

[0034] Furthermore, based on the inference results from the causal directions between variables that have passed the physical rule constraint verification, a global causal structure learning is performed on the causal variables, intermediate variables, and outcome variables. The goal of causal structure learning is to determine a unique directed acyclic graph structure in which the set of directed edges must be consistent with the causal directions between all variable pairs, and the graph structure can explain or imply the observed conditional independence relationships between all variable pairs.

[0035] Specifically, the learning process typically takes place on a complete graph containing all causal, intermediate, and outcome variables. Edges are added or removed progressively based on causal direction constraints, and the graph structure is tested to ensure compatibility with the conditional independence relationships observed in the data. The final determined causal connection topology is a directed acyclic graph (DAG), where the nodes are causal, intermediate, and outcome variables, and the edges represent direct causal relationships learned from the data and validated by physical rules. The entire topology fully depicts the causal transmission network between variables.

[0036] S2.4. Based on the causal connection topology between variables, generate a causal graph structure in the form of a directed acyclic graph with causal directions and causal strength weights.

[0037] Furthermore, based on the causal connection topology among the causal variables, intermediate variables, and outcome variables, a final directed acyclic graph (DAG) causal graph structure is generated. The generation process includes instantiating the nodes and directed edges in the causal connection topology as graphical elements, and calculating and assigning a weight value representing the strength of the causal effect to each directed edge in the DAG causal graph structure. The weight values ​​are calculated based on data from the training sample set of the mine backfill material treatment process, using regression analysis, structural equation modeling, or other causal effect quantification methods to assess the expected change in the outcome variable when the causal variable undergoes a unit change, while controlling for the influence of other related variables. After completing the weight calculation and assignment, the resulting DAG causal graph structure not only demonstrates the causal direction between variables but also quantifies the strength of each causal relationship through edge weights, forming a network model with complete causal semantics and quantitative information.

[0038] S3. Based on the variable connection relationships determined in the causal graph structure in the form of a directed acyclic graph, construct a hierarchical network connection architecture.

[0039] S3.1. Based on the topological sorting depth of nodes in a causal graph structure in the form of a directed acyclic graph, the network hierarchy is identified.

[0040] Furthermore, a topology sorting algorithm is performed on the causal graph structure in the form of a directed acyclic graph (DAG). The longest path length from each node in the DAG to all root nodes with an in-degree of zero is calculated, and this longest path length is defined as the topology sorting depth of that node. The topology sorting depth values ​​of all nodes in the DAG are compared, and nodes with the same topology sorting depth are grouped into the same level. Based on the increasing order of topology sorting depth from the root node to the terminal node, all nodes are assigned to different levels, thus identifying the hierarchical structure of the network inherent in the DAG causal graph structure.

[0041] It should be noted that directed acyclic graphs ,in It is a set of nodes. It is a set of directed edges, nodes ∈ topology depth The definition is as follows: ; in, For the longest causal path constraint, For nodes The topological depth function, For nodes The topological sort depth, For nodes The in-degree function.

[0042] Specifically, the topological depth of a node essentially reflects the shortest number of steps required for causal information transmission, i.e., the generation of a variable in the causal chain. For example, the root node (such as initial process parameters) has a topological depth of 0, the intermediate variables it directly affects have a depth of 1, the next-level variables affected by these intermediate variables have a depth of 2, and the final performance metric is located at the maximum depth layer. This hierarchical division based on causal path length ensures that the subsequently constructed layered network connection architecture is completely consistent with the actual causal sequence and logical dependencies in terms of information transmission order.

[0043] S3.2. Based on the network hierarchy, define the input layer, hidden layer, and output layer of the hierarchical network connection architecture. According to the specific causal connections in the causal graph structure in the form of a directed acyclic graph, establish neuron connections between the corresponding layers of the hierarchical network connection architecture.

[0044] Furthermore, based on the network hierarchy, layers of the hierarchical network connection architecture are defined. The set of nodes contained in the layer with a topological sorting depth of zero is defined as the input layer of the hierarchical network connection architecture. The set of nodes contained in the layer with the largest topological sorting depth is defined as the output layer of the hierarchical network connection architecture. The set of nodes contained in all layers with topological sorting depths between zero and the maximum value is defined as the hidden layers of the hierarchical network connection architecture, arranged in ascending order of topological sorting depth. After completing the layer definition, each directed edge in the causal graph structure in the form of a directed acyclic graph is traversed. For each directed edge from node A to node B, the layers to which node A and node B belong are found in the hierarchical network connection architecture. A connection is established between these two layers, from the neuron corresponding to node A to the neuron corresponding to node B, thereby transforming the causal connection relationship in the causal graph structure in the form of a directed acyclic graph into cross-layer neuron connections in the hierarchical network connection architecture.

[0045] S3.3. Based on the causal graph structure in the form of a directed acyclic graph, assign initial weight constraints to the neuron connections already established in the hierarchical network connection architecture.

[0046] Furthermore, based on the causal graph structure in the form of a directed acyclic graph (DAG), an initial weight constraint is assigned to each established neuron connection in the hierarchical network connection architecture. For an edge connecting two neurons, the causal strength weight on the corresponding directed edge in the causal graph structure is consulted. After processing this causal strength weight through a scaling function, it is used as the initial weight value for the corresponding neuron connection in the hierarchical network connection architecture. The initial weight value is not used as a fixed parameter, but rather as the parameter initialization point at the start of subsequent network training, and a soft constraint or prior is applied to the optimization process of this parameter. For example, by adding a regularization term related to the deviation of this initial weight to the loss function, the trained connection weights are encouraged to take values ​​in its vicinity, thereby injecting the prior strength information provided by the causal graph into the learning process of the neural network.

[0047] S3.4. Based on the established neuron connections with initial weight constraints, a hierarchical network connection architecture is obtained.

[0048] Furthermore, by integrating all neuron connections and the assigned initial weight constraints, a complete, structurally defined, and knowledge-initialized hierarchical network connection architecture is formed. The number of layers, the number of neurons in each layer, and the connection patterns between layers are uniquely determined by a causal graph structure in the form of a directed acyclic graph, thus obtaining the underlying computational graph skeleton prepared for constructing the physical information causal reasoning network model, i.e., the hierarchical network connection architecture.

[0049] S4. Construct a physical information causal reasoning network model on a hierarchical network connection architecture, and train the physical information causal reasoning network model using the training sample set.

[0050] S4.1. On the hierarchical network connection architecture, define the network layers and neuron activation functions of the physical information causal reasoning network model.

[0051] Furthermore, in the hierarchical network connection architecture, network layers of the physical information causal reasoning network model are defined. The input layer, hidden layer, and output layer of the hierarchical network connection architecture directly serve as the corresponding network layers of the physical information causal reasoning network model. Activation functions are specified for the neurons of each layer of the physical information causal reasoning network model. The input layer typically uses a linear activation function or directly passes the input, while the hidden layer can choose to use ReLU, Tanh, or other nonlinear activation functions to introduce nonlinear mapping capabilities. The output layer selects an activation function based on the nature of the predicted value being evaluated; for example, a linear activation function is used for regression prediction, while a Sigmoid activation function can be used for probability output. This is achieved by defining network layers and activation functions.

[0052] S4.2 The physical rule calculation module obtains the physical consistency constraint loss term based on the time-varying sequence of wave velocity, time-varying sequence of attenuation coefficient, time-varying sequence of hydration heat release rate and time-varying sequence of equivalent porosity index in the training sample set.

[0053] Furthermore, during model training, the physical rule calculation module calculates the physical consistency constraint loss term based on the true values ​​of the time-varying sequences of wave velocity, attenuation coefficient, hydration heat release rate, and equivalent porosity index contained in the training sample set of the mine backfill material treatment process. This is combined with the time-varying sequence of the equivalent porosity index predicted by the physical information causal inference network model in the current iteration step. At each time step, the squared error between the predicted and true time-varying sequences of the equivalent porosity index is calculated, and the residuals of the three physical relationships are calculated, corresponding to the wave velocity-porosity relationship, attenuation coefficient-porosity relationship, and hydration heat release rate-porosity relationship, respectively. The Huber loss function is applied to each residual to obtain the value of the physical consistency constraint loss term.

[0054] The expression for the physical consistency constraint loss term is: ; in, For the physical consistency constraint loss term, The total number of time steps for the time series in the training sample set. For the index of time steps, In time step The predicted time-varying sequence of the equivalent porosity index, In time step The true value of the time-varying sequence of the equivalent porosity exponent. For the first Adaptive weight coefficients corresponding to each physical relation constraint. An index for physical relation constraints. for, In time step No. The residual of a physical relationship.

[0055] S4.3 Based on the rational information causal reasoning network model, construct the total loss function of the data fitting loss term, the physical consistency constraint loss term, and the causal structure regularization loss term.

[0056] Furthermore, based on the defined network structure and activation function of the physical information causal inference network model, a total loss function is constructed for model training. The total loss function consists of three parts: a data fitting loss term, a physical consistency constraint loss term, and a causal structure regularization loss term. The data fitting loss term measures the difference between the final output prediction of the physical information causal inference network model and the true result variable values ​​in the training sample set of the mine backfill material treatment process; it typically uses mean squared error or cross-entropy loss. The causal structure regularization loss term aims to encourage the magnitude of neuron connection weights in the trained physical information causal inference network model to maintain consistency with the prior causal strength weights of corresponding edges in the causal graph structure in the form of a directed acyclic graph. For example, L2 norm can be used to penalize the deviation of connection weights from the initial weight constraints. These three losses are weighted by preset coefficients to form the total loss function.

[0057] S4.4 Input the total loss function and the training sample set of the mine backfill material treatment process into the optimization algorithm, and iteratively update the parameters of the physical information causal inference network model by minimizing the total loss function, and monitor the total loss function value and the prediction performance on the validation set.

[0058] Furthermore, the total loss function and a complete training sample set of the mine backfill material treatment process are input into an optimization algorithm, such as stochastic gradient descent or its variants. The workflow of the optimization algorithm is as follows: a batch of data is sampled from the training sample set of the mine backfill material treatment process and input into the physical information causal inference network model under the current parameter state. Forward propagation is performed to obtain the predicted value. Based on the predicted value and the true value, the total loss function value corresponding to the current batch of data is obtained. Backpropagation is performed to obtain the gradient of the total loss function with respect to each trainable parameter of the physical information causal inference network model. Using the update rule in the optimization algorithm, the parameters of the physical information causal inference network model are iteratively updated according to the gradient to minimize the total loss function. During the iterative update process, two indicators are continuously monitored: first, the value of the total loss function on the training set, observing its downward trend; second, the prediction performance of the physical information causal inference network model on an independently partitioned validation set, such as calculating the root mean square error or coefficient of determination of the validation set.

[0059] S4.5 When the total loss function value drops below the preset threshold and the fluctuation range of the validation set prediction performance stabilizes within the preset range, the trained physical information causal reasoning network model is obtained.

[0060] Furthermore, a preset threshold for the total loss function value is set, such as a positive decimal close to zero. A preset interval for the fluctuation range of the predicted performance on the validation set is also set, such as the difference between the upper and lower limits of the performance indicator's change over multiple consecutive evaluations. During the iterative update process, stopping conditions are continuously evaluated. The iterative update process stops when both of the following conditions are met: Condition 1, the total loss function value drops below the preset threshold. Condition 2, the fluctuation range of the predicted performance indicator on the validation set stabilizes within the preset interval over multiple consecutive evaluation periods, i.e., the performance no longer shows significant improvement or decline. When the iterative update process stops due to the satisfaction of the above conditions, the parameters of the physical information causal inference network model at this point are the parameters obtained from the final training, thus obtaining the trained physical information causal inference network model.

[0061] S5. Input the real-time multi-source data of the mine backfill material treatment process to be evaluated into the trained physical information causal reasoning network model to obtain the evaluation prediction value. Based on the causal graph structure in the form of a directed acyclic graph and the parameters of the physical information causal reasoning network model, generate the causal path and evaluation explanation report.

[0062] S5.1 Collect real-time multi-source data of the mine backfill material treatment process to be evaluated and perform preprocessing consistent with the training phase to obtain preprocessed real-time multi-source data.

[0063] Furthermore, real-time multi-source data on the mine backfill material remediation process to be evaluated is collected. The data type, source, and format of the collected data should be consistent with the historical multi-source data collected when constructing the training sample set for the mine backfill material remediation process. The collected real-time multi-source data is processed using the same preprocessing sequence as in the training phase, including but not limited to data cleaning, outlier handling, missing value imputation, temporal alignment, and standardization / normalization. The data processed in this way is called preprocessed real-time multi-source data, and its data structure and scale are consistent with the training sample set for the mine backfill material remediation process, serving as input to the trained physical information causal inference network model.

[0064] S5.2 Input the preprocessed real-time multi-source data into the trained physical information causal reasoning network model for forward propagation calculation to obtain the evaluation prediction value.

[0065] Furthermore, the preprocessed real-time multi-source data is input into the trained physical information causal reasoning network model. The trained physical information causal reasoning network model performs a complete forward propagation computation based on its fixed parameter set and the defined network layers and activation functions constituting the forward computation function. The forward propagation computation calculates the activation values ​​of neurons in each layer sequentially according to the hierarchy and connection order defined in the hierarchical network connection architecture, generating one or more output values ​​at the output layer. This value generated from the output layer represents the evaluation result of the current infill treatment effect.

[0066] The expression for the predicted value is: in, To evaluate the predicted values, For preprocessed real-time multi-source data, The parameter set of the trained physical information causal reasoning network model. This is a forward computation function.

[0067] S5.3. Based on the connection weight parameters of the causal graph structure in the form of a directed acyclic graph and the trained physical information causal reasoning network model, the causal path from the causal variable to the evaluation prediction value is traced back.

[0068] Furthermore, in the causal graph structure in the form of a directed acyclic graph, the result variable node corresponding to the evaluation prediction value is located. Starting from this node, all direct parent nodes pointing to this node are searched along the opposite direction of the directed edge. The parent node is the direct cause of the result. For each direct parent node found, its direct parent node is searched along the opposite direction of the directed edge. The search continues until the root node with an in-degree of zero, i.e. the cause variable, is backtracked. One or more directed paths from the cause variable, through a series of intermediate variables, to the result variable are recorded.

[0069] S5.4 Based on the causal path, calculate the quantitative influence weight of each causal variable in the causal path on the evaluation prediction value.

[0070] Furthermore, a causal path refers to a continuous sequence of directed edges that starts from a causal variable node with zero in-degree in a causal graph structure in the form of a directed acyclic graph, follows directed edges, passes through one or more intermediate variable nodes, and finally reaches the result variable node representing the evaluation prediction value. The nodes and edges on this path completely characterize the causal transmission chain from the initial controllable process or material parameters to the final performance index. For a determined causal path, a trained physical information causal reasoning network model is used to perform forward propagation calculations on the preprocessed real-time multi-source data of the current input to obtain the evaluation prediction value. Simultaneously, backpropagation automatic differentiation technology is used to calculate the effect of the evaluation prediction value on the causal path. The partial derivative of the input value of each causal variable on the causal path at each observation time step is used to sum the absolute values ​​of the partial derivatives of each causal variable at all relevant time steps. This sum represents the total influence strength of the causal variable on the evaluation prediction value under the current input state. The total influence strength of all causal variables on the causal path is summed to obtain the cumulative influence strength of all causal variables. The ratio of the total influence strength of each causal variable to the cumulative influence strength of all causal variables is the quantitative influence weight of the causal variable. This weight value is between zero and one, reflecting the relative contribution importance of the causal variable to the evaluation prediction result in this specific evaluation scenario.

[0071] The expression for quantifying the influence weight is: ; in, For the first in the causal path The weights of the quantitative impact of each causal variable on the predicted value. To evaluate the predicted values For causal variables At time step The partial derivatives, The total number of causal variables. For the index of the cause variable, For time steps Evaluation of predicted values On the causal path, the first One causal variable The partial derivatives, For the first in the causal path The causal variables at time step The value of , For the first in the causal path The causal variables at time step Specific observations or input values, for, In time step The predicted value of the assessment.

[0072] S5.5 Integrate causal paths with corresponding quantitative impact weights to generate causal path and assessment explanation reports.

[0073] Furthermore, the causal path and the quantified impact weight of each causal variable along the causal path are integrated to generate the final assessment and explanation report. The report content is structured to include the assessed predicted value and the main causal transmission chain leading to that predicted value. For each causal path, the report lists the sequence of intermediate variables from the causal variable to the assessed predicted value. For each causal variable in the path, the report explicitly provides its quantified impact weight. Based on these weights, the report ranks all causal variables by importance and may incorporate domain knowledge to provide qualitative analysis or optimization suggestions for potential process problems indicated by higher-weighted causal variables. The resulting report is the causal path and assessment and explanation report.

[0074] This embodiment also provides a machine learning-based system for evaluating the effectiveness of mine backfill material remediation, including: The processing module collects historical multi-source data on the mine backfill material treatment process and performs preprocessing to obtain a training sample set of the mine backfill material treatment process. The learning module classifies the variables in the training sample set into causal variables, intermediate variables, and outcome variables. It then uses a causal discovery algorithm to learn the causal structure of the causal variables, intermediate variables, and outcome variables, generating a causal graph structure in the form of a directed acyclic graph. The building module constructs a hierarchical network connection architecture based on the variable connection relationships determined in the causal graph structure in the form of a directed acyclic graph. The training module constructs a physical information causal reasoning network model on a hierarchical network connection architecture and trains the physical information causal reasoning network model using the training sample set. The prediction module inputs real-time multi-source data of the mine backfill material treatment process to be evaluated into a trained physical information causal reasoning network model to obtain the evaluation prediction value. Based on the causal graph structure in the form of a directed acyclic graph and the parameters of the physical information causal reasoning network model, it generates the causal path and evaluation explanation report.

[0075] This embodiment also provides a computer device applicable to the evaluation method of mine backfill material treatment effect based on machine learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the evaluation method of mine backfill material treatment effect based on machine learning as proposed in the above embodiment.

[0076] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0077] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for evaluating the treatment effect of mine backfill materials based on machine learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0078] In summary, this invention constructs a training sample set by collecting and preprocessing historical multi-source data; then, by integrating a causal discovery algorithm with physical rule constraints, it learns the causal structure between causal variables, intermediate variables, and outcome variables in the samples, generating a weighted directed acyclic graph (DAG) structure. Based on this causal graph structure, a hierarchical network connection architecture is constructed, and a causal inference network model integrating physical consistency constraints is built and trained. The trained model is used to evaluate and predict real-time data, and a quantitative causal path and an interpretable evaluation report are generated by backtracking based on the causal graph and model parameters.

[0079] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for evaluating the effect of mine filling material management based on machine learning, characterized in that: include, Historical multi-source data on the treatment process of mine backfill materials were collected and preprocessed to obtain a training sample set of the treatment process of mine backfill materials. The variables in the training sample set are divided into causal variables, intermediate variables and outcome variables. A causal discovery algorithm is used to learn the causal structure of the causal variables, intermediate variables and outcome variables, and generate a causal graph structure in the form of a directed acyclic graph. Based on the variable connection relationships determined in the causal graph structure in the form of a directed acyclic graph, a hierarchical network connection architecture is constructed. A physical information causal reasoning network model is constructed on a hierarchical network connection architecture, and the physical information causal reasoning network model is trained using the training sample set. Real-time multi-source data of the mine backfill material treatment process to be evaluated is input into a trained physical information causal reasoning network model to obtain evaluation prediction values. Based on the causal graph structure in the form of a directed acyclic graph and the parameters of the physical information causal reasoning network model, a causal path and evaluation explanation report are generated.

2. The method for evaluating the effect of mine filling material management based on machine learning according to claim 1, characterized in that: The training sample set for the mine backfill material treatment process includes... Historical multi-source data on the treatment process of mine backfill materials were collected and standardized and normalized to obtain a training sample set of the treatment process of mine backfill materials.

3. The method for evaluating the effect of mine filling material management based on machine learning according to claim 2, characterized in that: The causal graph structure in the form of a directed acyclic graph includes, Based on the physical meaning of each variable in the training sample set and the temporal stages in the governance process, the variables in the training sample set are clearly divided into causal variables, intermediate variables and outcome variables. The algorithm for causal discovery combines causal variables, intermediate variables, and outcome variables with the constraints of physical rules governing time irreversibility and mass-energy conservation. It then tests the statistical independence among the causal variables, intermediate variables, and outcome variables, applies physical rules, and infers the causal direction between the variables. Based on the causal direction between variables, causal structure learning is performed on causal variables, intermediate variables and outcome variables to determine the causal connection topology between variables; Based on the causal connection topology between variables, a causal graph structure in the form of a directed acyclic graph is generated, which includes causal directions and causal strength weights.

4. The method for evaluating the treatment effect of mine backfill materials based on machine learning as described in claim 3, characterized in that: The layered network connectivity architecture includes, Based on the topological sorting depth of nodes in a causal graph structure in the form of a directed acyclic graph, the network hierarchy is identified. Based on the network hierarchy, the input layer, hidden layer and output layer of the hierarchical network connection architecture are defined. According to the specific causal connections in the causal graph structure in the form of a directed acyclic graph, neuron connections are established between the corresponding layers of the hierarchical network connection architecture. Based on the causal graph structure in the form of a directed acyclic graph, initial weight constraints are assigned to the neuron connections already established in the hierarchical network connection architecture. Based on the established neuron connections with initial weight constraints, a hierarchical network connection architecture is obtained.

5. The method for evaluating the treatment effect of mine backfill materials based on machine learning as described in claim 4, characterized in that: The training of the physical information causal reasoning network model includes... In the hierarchical network connection architecture, the network layers and neuron activation functions of the physical information causal reasoning network model are defined; The physical consistency constraint loss term is obtained by the physical rule calculation module based on the time-varying sequence of wave velocity, time-varying sequence of attenuation coefficient, time-varying sequence of hydration heat release rate and time-varying sequence of equivalent porosity index in the training sample set. Based on the rational information causal reasoning network model, a total loss function is constructed, consisting of a data fitting loss term, a physical consistency constraint loss term, and a causal structure regularization loss term. The total loss function and the training sample set of the mine backfill material treatment process are input into the optimization algorithm. The parameters of the physical information causal inference network model are updated iteratively by minimizing the total loss function, and the total loss function value and prediction performance on the validation set are monitored. When the total loss function value drops below a preset threshold and the fluctuation range of the validation set prediction performance stabilizes within a preset range, the trained physical information causal reasoning network model is obtained.

6. The method for evaluating the treatment effect of mine backfill materials based on machine learning as described in claim 5, characterized in that: The predicted evaluation values ​​include, Real-time multi-source data of the mine backfill material treatment process to be evaluated is collected and preprocessed in the same way as in the training phase to obtain preprocessed real-time multi-source data. The preprocessed real-time multi-source data is input into the trained physical information causal reasoning network model for forward propagation calculation to obtain the evaluation prediction value.

7. The method for evaluating the treatment effect of mine backfill materials based on machine learning as described in claim 6, characterized in that: The generated causal path and evaluation explanation report includes, Based on the connection weight parameters of the causal graph structure in the form of a directed acyclic graph and the trained physical information causal reasoning network model, the causal path from the causal variable to the evaluation prediction value is traced back. Based on the causal path, calculate the quantitative impact weight of each causal variable in the causal path on the evaluation prediction value; Integrate causal paths with corresponding quantitative impact weights to generate causal path and assessment explanation reports.

8. A machine learning-based system for evaluating the effectiveness of mine backfill material remediation, based on the machine learning-based method for evaluating the effectiveness of mine backfill material remediation as described in any one of claims 1 to 7, characterized in that: include, The processing module collects historical multi-source data on the mine backfill material treatment process and performs preprocessing to obtain a training sample set of the mine backfill material treatment process. The learning module classifies the variables in the training sample set into causal variables, intermediate variables, and outcome variables. It then uses a causal discovery algorithm to learn the causal structure of the causal variables, intermediate variables, and outcome variables, generating a causal graph structure in the form of a directed acyclic graph. The building module constructs a hierarchical network connection architecture based on the variable connection relationships determined in the causal graph structure in the form of a directed acyclic graph. The training module constructs a physical information causal reasoning network model on a hierarchical network connection architecture and trains the physical information causal reasoning network model using the training sample set. The prediction module inputs real-time multi-source data of the mine backfill material treatment process to be evaluated into a trained physical information causal reasoning network model to obtain the evaluation prediction value. Based on the causal graph structure in the form of a directed acyclic graph and the parameters of the physical information causal reasoning network model, it generates the causal path and evaluation explanation report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the machine learning-based method for evaluating the treatment effect of mine backfill materials as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning-based method for evaluating the treatment effect of mine backfill materials as described in any one of claims 1 to 7.

Images