Intelligent control method, device and apparatus for backfilling body replacing coal pillar

Abstract

The application provides a filling body replacement coal pillar intelligent control method, equipment and device, and belongs to the field of coal mine filling automation control; the method first constructs a unified multi-dimensional data set; through a multi-framework set Frameworks, high-dimensional feature mapping of environment state and execution parameters is realized; further, a double mutual information constraint collaborative representation training model is utilized to obtain a filling state-execution collaborative representation with low redundancy and high consistency; on this basis, a regression prediction model is adopted to realize accurate mapping from the collaborative representation to the hydraulic control instruction, forming an intelligent control kernel; finally, the kernel is embedded into the control equipment, and through real-time communication with the on-site gateway, the whole process closed loop of data access, instruction generation and equipment control is completed; the application overcomes the defects of traditional coal pillar replacement, such as reliance on manual experience parameter adjustment, low control precision and serious waste of filling materials, and realizes autonomous decision and fine control of the filling body replacement coal pillar process.

Description

Technical Field

[0001] This application relates to the field of automated control technology for coal mine backfilling, and in particular to an intelligent control method, equipment and device for replacing coal pillars with backfilling bodies. Background Technology

[0002] Replacing coal pillars with backfill is a key technology for controlling ground pressure and maintaining surrounding rock stability. The core challenge lies in the accurate perception, intelligent decision-making, and real-time closed-loop control of the backfilling process. Currently, research and practice in this field mainly focus on backfilling process control based on human experience or single sensor data, manually intervening in the backfilling state by monitoring parameters such as pressure, flow rate, or displacement. However, this type of control has low automation and limited data utilization, making it difficult to achieve real-time and accurate control of the backfill evolution process under complex geological conditions.

[0003] Existing filling control technologies have significant shortcomings. On the one hand, traditional filling control systems rely heavily on single-point monitoring and threshold judgment, failing to achieve dynamic correlation and fusion of multi-dimensional sensor information. This results in delayed perception of filling status, ground pressure changes, and equipment response, allowing only passive adjustments. On the other hand, the lack of comprehensive analysis of multi-source information such as filling equipment operating data, geological parameters, and acoustic emission signals prevents the system from accurately identifying filling anomalies and predicting potential instability risks, leading to insufficient intelligent decision-making capabilities.

[0004] In recent years, although some studies have attempted to introduce automated monitoring and data-driven models, they often only achieve parameter optimization in localized stages, lacking systematic integration and real-time adaptive control mechanisms, making it difficult to meet the safety and efficiency requirements of complex mining environments. Therefore, there is an urgent need for an intelligent control method for coal pillar replacement in filling bodies that can achieve multi-dimensional sensing data fusion, intelligent decision-making, and real-time closed-loop control, in order to realize the full automation, intelligence, and real-time operation of the filling process and improve coal mine safety production and resource recovery efficiency. Summary of the Invention

[0005] To address the aforementioned technical issues, this application proposes an intelligent control method, equipment, and device for replacing coal pillars with filling materials.

[0006] The technical solution adopted in this application is: an intelligent control method for coal pillar replacement by filling body, based on smart mines, including the following steps:

[0007] S1: Multi-dimensional sensing data integration during the filling process: Unifying and integrating data from different types of sensors into a multi-dimensional data set that can be used for intelligent control and analysis. This multidimensional dataset Includes multidimensional sensing set of filling area Hydraulic actuator parameter set Simultaneously, a set of instruction tags for filling equipment was prepared. ;

[0008] S2: For multidimensional data sets Feature extraction: Utilizing a multi-framework set of multidimensional data sets Multi-source heterogeneous data are projected in a unified dimension to form a feature set. Feature set Including multi-dimensional sensing sets for filling areas Feature extraction results And parameter sets for hydraulic actuators Feature extraction results ,in It is a set of environmental state characteristics. For the set of execution parameter features;

[0009] S3: Filling State - Execution Co-representation Training: Constructing the Filling State - Execution Co-representation Trainer By filling the state-execution co-representation trainer Set of environmental state characteristics With the set of execution parameters Perform joint training to obtain a filled state with cooperative properties and minimized redundant information - execute cooperative representation. And the optimal filling state - execute the co-representation trainer ;

[0010] S4: Device Command Prediction Training Based on Cooperative Representation: Constructing a Device Command Predictor To fill the state - perform collaborative representation As input, fill the device instruction label set P with labels, and train the device instruction predictor. Ultimately, the optimal device instruction predictor is obtained. ;

[0011] S5: Construct the intelligent control kernel for coal pillar replacement with filling body: This involves using the model framework (Frameworks) obtained in steps S1-S4. and The sequentially stacked organization forms the intelligent filling body replacement coal pillar control core. The core receives data from various sensors at the coal pillar replacement site, and the sensor data is sequentially processed through Frameworks-> -> Subsequently, instructions regarding the hydraulic actuators were received. ;

[0012] S6: Develop embedded intelligent control software for filling body replacement coal pillars: The software encapsulates the kernel built in step S5 and includes control interfaces for hydraulic actuators, directly automating the control of various devices involved in the filling body replacement coal pillar process.

[0013] Furthermore, step S1 specifically includes:

[0014] S1.1: Data sensing and acquisition related to coal pillar replacement by filling body: Through industrial bus or industrial IoT platform, synchronously collect filling environment data and hydraulic actuator data at different replacement points, and at the same time, collect or record corresponding filling equipment instruction information;

[0015] The filling environment data includes multi-dimensional sensing data such as filling medium and pipeline parameters, surrounding rock and ground pressure information, and environmental and safety conditions. The hydraulic actuator data includes actuator elements, power source supply and return oil parameters, control valves and pipeline control parameters.

[0016] S1.2: Loading of data related to coal pillar replacement by filling body: Load the collected multi-dimensional sensing data and filling equipment instruction information into the data lake warehouse system of the data platform for persistent storage and management;

[0017] S1.3: Preprocess the multidimensional sensing data collected in step S1.1;

[0018] S1.4: Construction of Multidimensional Perception Dataset for Filling Process: The preprocessed data is manually labeled and verified based on expert knowledge to form an approved multidimensional perception set for the filling area. .

[0019] Furthermore, step S2 specifically includes:

[0020] S2.1: Multi-framework-based multi-dimensional sensing set for filling regions Unified Feature Mapping: A multi-framework neural network architecture is used to perform feature mapping on various data sources to form a unified set of environmental state features. And the optimal set of multiple frameworks after training;

[0021] S2.2: Hydraulic Actuator Parameter Set Feature mapping: for Correlated time-series data and discrete parameter data are mapped using a single feature extraction network to obtain the feature set of execution parameters of the hydraulic actuator. ;in, Feature dimensions and environmental state feature set in step S2.1 The feature dimensions are consistent.

[0022] Furthermore, step S2.1 specifically includes:

[0023] S2.1.1: Data Type Classification: Classify data types according to their data type. If divided into four subsets, then ,in Represents time-series data. Represents spatial displacement type data, Represents image / point cloud type data, Represents discrete data types;

[0024] S2.1.2: Multi-framework construction for feature mapping: Corresponding neural network frameworks are constructed to achieve efficient feature mapping based on the feature distribution and structural characteristics of different types of data. The data uses the LSTM framework. The class data uses the GCN framework. The class data uses the ResNet framework. The data adopts the MLP framework, and each neural network framework contains an encoder module and a decoder module, forming a symmetrical input-reconstruction structure;

[0025] S2.1.3: Feature Training Based on Reconstruction Loss: Define the reconstruction loss function as:

[0026] ;

[0027] make ;

[0028] in, express , Decoders representing neural network frameworks corresponding to different types of data. Encoders that represent the neural network frameworks corresponding to different types of data;

[0029] By minimizing the aforementioned reconstruction loss, gradient descent iterative optimization is performed on the four neural network frameworks. This process yields feature representations with high information preservation and discriminative power. and optimal framework ;

[0030] S2.1.4: Constructing a set of environmental state features Frameworks: Following the method in step S2.1.3, the four types of data and their corresponding network frameworks are trained respectively, and finally summarized into the following two sets: , .

[0031] Furthermore, the filling state-execution co-representation trainer It adopts a hierarchical, multi-branch fusion architecture, including the following modules:

[0032] The perception initialization module consists of an initial connection layer, a batch normalization (BN) layer, and an activation function, and is used to receive a set of environmental state features. With the set of execution parameters Data in the middle;

[0033] Geo-force feature extraction module: contains multiple residual units and downsampling convolutions to capture the spatial gradient features of geological and stress fields, and realize the structured encoding of geo-force signals;

[0034] The filling medium evolution coding module consists of a continuous stack of residual blocks, combined with batch normalization and ReLU activation, to adapt to multi-scale feature expression. It is used to characterize the changes in physical properties of high-water filling bodies under different ratios and hydration stages, forming a spatiotemporal coding of material states.

[0035] Hydraulic control response encoding module: Combining multi-channel residual blocks and local attention mechanisms to model dynamic sequence features;

[0036] Collaborative fusion inference module: Includes cross-channel fusion unit and nonlinear projection layer to achieve feature alignment in multi-view latent space.

[0037] Furthermore, step S3 specifically includes:

[0038] S3.1: Constructing the Filling State - Executing the Cooperative Representation Trainer ;

[0039] S3.2: Filling State - Execution Intrinsic Information Redundancy Reduction: By minimizing the redundant mutual information between the filling region environmental state characteristics and the execution parameter characteristics, invalid correlations are reduced, including:

[0040] S3.2.1: Calculate the collaborative fusion reasoning representation: using the filling state-execution collaborative representation trainer built in step S3.1. Calculate about and Fusion reasoning representation and The calculation formula is: ;

[0041] S3.2.2: Calculate the co-occurrence matrix among fused inference representations The calculation formula is: Then normalize the matrix. Its formula is ,in Representation matrix The sum of all elements in the matrix It is a square matrix with dimension D;

[0042] S3.2.3: Calculate the filling status - execute the intrinsic information redundancy. The calculation formula is:

[0043] :

[0044] in The first smoothing parameter is configurable. Representation matrix The Line number Column elements, Indicates the first The sum of actions: , Indicates the first Sum of columns: ;

[0045] S3.2.4: Redundancy Removal Training: To minimize To achieve this, update the filling state using gradient descent and execute the co-representation trainer. Parameters;

[0046] S3.2.5: Repeat steps S3.2.1-S3.2.4 for a total of Second-rate;

[0047] S3.3: Filling State - Execution of Cooperative Action Coupling Matching: Maximizing Coupling Matching under the Same Working Condition and Mutual information, aligning positive example representations, and achieving dynamic coordination between states and actions, including:

[0048] S3.3.1: Computational coupling between fusion inference representations First, obtain the fusion reasoning representation using the exact same method as in step S3.2.1. and ,assumed and The first in The rows are respectively and Calculate the coupling using the following formula : ,in express and The number of rows; then normalize the matrix in the same way as in step S3.2.2. Finally, we calculate the weighted average to get: The matrix It is also a square matrix, with the same dimension D as the matrix in step S3.2.2. The dimensions are consistent;

[0049] S3.3.2: Calculating Fill State - Performing Coupled Matching The calculation formula is:

[0050] ;

[0051] in This is a configurable second smoothing parameter. Representation matrix The Line number Column elements, Indicates the first The sum of actions: , Indicates the first Sum of columns: ;

[0052] S3.3.3: Matching Enhancement Training: To Maximize To achieve this, update the filling state using gradient descent and execute the co-representation trainer. Parameters;

[0053] S3.3.4: Repeat steps S3.3.1-S3.3.3 (total). Second-rate;

[0054] S3.4: Filling State - Execution Co-representation Construction: Based on the training results of steps S3.1-S3.3, obtain the filling state - execution co-representation. The calculation formula is: = .

[0055] Furthermore, the device instruction predictor It is composed of the following levels stacked in the order of introduction:

[0056] Collaborative information normalization layer: Contains initial convolution, batch normalization, and ReLU activation units, used to perform collaborative representation on the filling state. Scale standardization and feature enhancement are performed to eliminate statistical differences between different physical quantities;

[0057] Device state parsing block: Consists of two residual units, including convolution, batch normalization, and shorting structures, used to extract filling state - perform collaborative representation. The underlying information reflects the operating status, stability, and health characteristics of the hydraulic actuator;

[0058] Working condition feature analysis block: stacked residual units with downsampling convolution, used to identify macroscopic working condition patterns under different geological and backfilling environments and capture the changing patterns of backfilling state;

[0059] Control signal sensing block: By combining multi-channel convolution with batch normalization and adding attention modules to some channels, it is used to enhance the potential feature dimensions most relevant to hydraulic control commands and learn key control factors.

[0060] Predictive inference layer: Composed of high-order residual units and nonlinear projection layers, it is used to synthesize multi-scale features from the previous stage to form a nonlinear mapping capability for the target command;

[0061] Hydraulic control prediction head: a multi-layer fully connected network used to output continuous hydraulic command vectors.

[0062] Furthermore, step S4 specifically includes:

[0063] S4.1: Building a Device Command Predictor ;

[0064] S4.2: Based on Filling State - Execution Cooperative Representation The device instruction prediction training includes:

[0065] S4.2.1: Calculation of equipment instruction regression error, the calculation formula is as follows: ,in and These represent the instruction tag sets for filling equipment. and filling state - execution co-representation The Instructions and collaborative representations Indicates the number of data points. Instructions indicating a prediction;

[0066] S4.2.2: Device Command Prediction Training: To Maximize To achieve this, the collaborative representation trainer is updated using gradient descent. Parameters;

[0067] S4.3.3: Repeat steps S4.3.1-S4.3.2 for a total of This process is repeated to obtain the final optimized model. .

[0068] A smart control device for filling material replacement coal pillars, used to execute control commands output by the aforementioned smart control method for filling material replacement coal pillars, comprises two parts: hardware and software.

[0069] The hardware consists of a computing and control module, a data access module, an execution driver module, a communication network module, a power protection module, and a human-computer interaction module.

[0070] Among them, the computing and control module is the core of the equipment, which is used to run the kernel of the intelligent control software for the filling body and to perform collaborative representation reasoning and hydraulic command generation.

[0071] The data access module is responsible for acquiring multi-dimensional sensor and equipment status data through the field gateway;

[0072] The execution drive module converts the calculated control commands into real-time control signals for the hydraulic pump, valves and related actuators, enabling precise operation of the coal pillar replacement process;

[0073] The communication network module provides a high-speed data interaction channel between the device and the field gateway, the upper-level monitoring system and the industrial Internet platform, and supports Ethernet and industrial bus communication;

[0074] The power protection module provides a stable power supply for the equipment;

[0075] The human-computer interaction submodule is used for on-site information display, parameter adjustment, and emergency control operations;

[0076] The software component consists of an operating system module, intelligent control software, and an execution driver module.

[0077] The operating system module provides the basic environment and system services for device operation.

[0078] The intelligent control software is the intelligent control software for the filling body replacement coal pillar constructed in step S6, which is used to realize real-time control of the filling body replacement coal pillar.

[0079] The execution driver module is responsible for interacting with the field hardware, completing data input, command output, and communication protocol adaptation.

[0080] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0081] The advantages of this application over the prior art are as follows:

[0082] First, intelligent data acquisition and fusion analysis. A multi-sensor collaborative sensing system is constructed to achieve unified acquisition and time synchronization of multi-dimensional information such as pressure, flow rate, acoustic emission, geological conditions, and safety conditions, greatly improving the accuracy, stability, and timeliness of the data.

[0083] Secondly, intelligent control algorithms enable high-precision automatic decision-making. By introducing a collaborative representation and mutual information constraint mechanism for filling state-execution, the model can automatically identify state changes under different geological conditions and generate optimal control commands, significantly improving the accuracy and safety of coal pillar replacement.

[0084] Furthermore, the model's self-learning and interpretability are enhanced. The control equipment possesses continuous learning and model self-adaptation capabilities, enabling it to dynamically optimize parameters through data feedback. Moreover, key control factors can be traced through a collaborative representation mechanism, thereby enhancing the interpretability and engineering controllability of the model results.

[0085] Finally, closed-loop automatic control is achieved through the collaboration of software and hardware. By deeply coupling embedded intelligent control software with field control equipment, a closed-loop control chain of "data acquisition - intelligent decision-making - execution feedback" is formed, realizing efficient, stable and unmanned operation of the entire coal pillar replacement process. Attached Figure Description

[0086] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0087] Figure 1 The overall flowchart for implementing the method provided in the embodiments of this application is shown.

[0088] Figure 2 This is a schematic diagram of the overall equipment layout provided in an embodiment of this application.

[0089] Figure 3 A schematic diagram of a trained and optimized multi-frame set provided for an embodiment of this application.

[0090] Figure 4 A schematic diagram of the module composition of the filling state-execution collaborative representation trainer provided in the embodiments of this application.

[0091] Figure 5 This is a schematic diagram of the module composition of the device instruction predictor provided in the embodiments of this application.

[0092] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0093] like Figures 1 to 6 As shown, this application provides an intelligent control method for replacing coal pillars with filling bodies, based on a coal mine under a smart mine system. The equipment in this coal mine includes, but is not limited to: a smart coal mine data platform, underground edge computing equipment, sensors, underground gateways, etc. Figure 1 The flowchart illustrating the method implementation of this application is shown in the document. Specifically, Figure 1 The multiple cloud-like diagrams and cubes shown in the upper left corner represent a large amount of historical scenarios and data regarding the replacement of coal pillars with backfill. This application will establish a model and mapping method based on this data from the "environment and state of coal pillar replacement with backfill" to the "implementation of equipment instructions" (e.g., Figure 1 (Blue dashed box in the middle). After the modeling and training process in steps S1-S4, the intelligent control software can be developed through steps S5-S6. Figure 2 The diagram shows the overall equipment layout for applying this method to a coal mine, which can... Figure 1 The intelligent control software obtained is deployed on the operating system and requires supporting driver software (i.e., execution driver module) for the filling body replacement-related equipment. In this specific implementation scenario (such as...), Figure 2 As shown in the cloud-like diagram and cube diagram, the intelligent equipment for replacing coal pillars with filling materials can obtain multi-dimensional sensing data and hydraulic actuator parameters from the field through the layer-by-layer processing of the communication network module and the data access module. Next, the intelligent control software of the hardware support software part of the calculation control module calculates and forms drive control code, which is delivered to the execution drive module to realize the actual equipment control.

[0094] The implementation methods of this application will now be described in detail.

[0095] First, we introduce the intelligent control method for coal pillar replacement by filling bodies. The specific steps are as follows:

[0096] S1: Multi-dimensional sensing data integration during the filling process: Unifying different types of sensor data (including pressure, flow rate, displacement, acoustic emission, video, geological parameters, etc.) into a multi-dimensional data set that can be used for intelligent control and analysis. This multidimensional dataset Includes multidimensional sensing set of filling area Hydraulic actuator parameter set Simultaneously, a set of instruction tags for filling equipment was prepared. It should be noted that , as well as The data entries correspond one-to-one, representing the environment at a certain moment. Execute Device parameters at the time of instruction .

[0097] S2: For multidimensional data sets Feature extraction: Utilizing a multi-framework set of multidimensional data sets Multi-source heterogeneous data are projected in a unified dimension to form a feature set. . This includes a multi-dimensional sensing set for the filling area. Feature extraction results And parameter sets for hydraulic actuators Feature extraction results ,in It is a set of environmental state characteristics. This is the set of execution parameter features. Multi-framework sets can be optimized after this step.

[0098] S3: Filling State - Execution Co-representation Training: Constructing the Filling State - Execution Co-representation Trainer By filling the state-execution co-representation trainer Set of environmental state characteristics With the set of execution parameters Perform joint training to obtain a filled state with cooperative properties and minimized redundant information - execute cooperative representation. Furthermore, the optimal filling state-execution co-representation trainer was obtained through training. .

[0099] S4: Device instruction prediction training based on collaborative representation: filling state - executing collaborative representation As input, and with the filling equipment instruction label set P as the label, a regression model is trained to predict future filling equipment instructions. This regression model serves as the equipment instruction predictor. Ultimately, the optimal device instruction predictor can be obtained. .

[0100] S5: Construct the intelligent control kernel for coal pillar replacement with filling body: This involves using the model framework (Frameworks) obtained in steps S1-S4. and The sequentially stacked organization forms the intelligent filling body replacement coal pillar control core. The core receives data from various sensors at the coal pillar replacement site, which are then processed sequentially through Frameworks-> -> Afterwards, instructions regarding the hydraulic actuators can be obtained. .

[0101] S6: Develop embedded intelligent control software for filling body replacement coal pillars: The software encapsulates the kernel built in step S5 and includes control interfaces for hydraulic actuators, enabling direct automated control of various devices involved in the filling body replacement coal pillar process.

[0102] Step S1 specifically includes:

[0103] S1.1: Data Sensing and Acquisition Related to Coal Pillar Replacement by Backfilling: Through an industrial bus or industrial IoT platform, synchronously acquire data on the filling environment at different replacement points (including multi-dimensional sensing data such as filling medium and pipeline parameters, surrounding rock and ground pressure information, environmental and safety operating conditions, etc.) and data on hydraulic actuators (including actuator components, power source supply and return oil parameters, control valves and pipeline control parameters, etc.). Simultaneously, acquire or record corresponding filling equipment instruction information.

[0104] because and The same strategy is used for processing; for simplicity, the operations in S1 are represented by multidimensional data sets. For example, let's describe it.

[0105] S1.1.1: Sensor deployment in the filling area and coal pillar area: For the coal pillar replacement operation area of ​​the filling body, pressure sensors are deployed in the grouting pipeline and the surrounding rock pressure area, flow sensors are deployed at the slurry flow rate monitoring point, ultrasonic densitometers are deployed at the slurry concentration monitoring point, triaxial strain gauges are deployed in the surrounding rock deformation area, and temperature, humidity, gas and dust sensors are configured in the environmental safety area to achieve comprehensive perception of multi-dimensional physical quantities in the filling area and coal pillar area.

[0106] S1.1.2: Signal Acquisition Synchronization Settings: Assuming the first... The clock of each sensor is The calibration time relationship between the sensor and the gateway is then: , among which is Frequency offset parameter, It is an offset parameter. This is the gateway reference time parameter. Then, it can be obtained using bidirectional packets. and The initial estimate of the channel delay is used as the initial value for both. Then, multiple samples are taken over a period of time T. and The optimal solution is obtained using the moving least squares method based on the data. and Its formula is: ,in This indicates the subscript for the number of samples. Finally, use the formula... Calculate calibration gateway time ,in , The optimal solution after solving and This value (i.e., the calibration gateway time) That is, the final number. The synchronization time of each sensor data is determined by setting the corresponding data upload time each time data is uploaded. Data transmission begins after constructing a timestamp based on the current moment.

[0107] S1.1.3: Data Packaging: The signal data from each sensor is aggregated by the gateway and then packaged. The data format is defined as follows: ,in A data identifier for device identity. The timestamp obtained in step S1.1.2 This represents the signal data from each sensor, where the subscript indicates the sensor number. This indicates the total number of sensors assigned to the current gateway.

[0108] S1.1.4: Data transmission: The packaged data is transmitted to the data platform using an industrial bus or wireless industrial IoT protocol.

[0109] S1.2: Loading of data related to coal pillar replacement of filling body: Load the collected multidimensional data and filling equipment instruction information into the data lake warehouse system of the data platform for persistent storage and management.

[0110] S1.2.1: Data Reception and Verification: Receive the packaged data output from step S1.1, parse the message fields, and perform integrity verification and format conversion to ensure that the data structure conforms to the data entry specifications.

[0111] S1.2.2: Data Persistence: To address the characteristics of multi-source heterogeneous data, a structured and unstructured database system is constructed for hierarchical storage and unified management of different types of data.

[0112] S1.2.3: Data Indexing and Cache Construction: Based on Timestamps and Devices Establish a multi-level index structure and build a high-speed caching mechanism to support the efficient execution of subsequent high-frequency queries, statistical analysis and modeling operations.

[0113] S1.3: Quality Improvement of Data Related to Coal Pillar Replacement by Filling Body: Preprocess the multi-dimensional sensing data collected in step S1.1 to improve data quality and consistency. Note that this step is not required for filling equipment instruction data.

[0114] S1.3.1: Signal denoising: Mean filtering and Kalman filtering strategies are used for denoising to reduce sensor noise interference and improve signal stability.

[0115] S1.3.2: Handling of parameters exceeding limits: When a parameter is detected to exceed the preset threshold range, the control unit adaptively corrects the corresponding parameter based on PID control or fuzzy control strategy to prevent outliers from interfering with subsequent analysis.

[0116] S1.3.3: Data Standardization Processing: Perform a normalization operation on the processed data. The calculation formula is as follows: ,in Represents the original data. and These represent the maximum and minimum values ​​of the data entry, respectively.

[0117] S1.4: Construction of Multidimensional Perception Dataset for Filling Process: The standardized data is manually labeled and verified based on expert knowledge to form an approved multidimensional perception set for the filling area. This provides high-quality input for subsequent intelligent analysis and control.

[0118] S1.4.1: True value standardization and generation: Obtain key indicator data such as compressive strength, fluidity and setting time of the filling body according to the same standard test, and correct and map the sensor data after processing in step S1.3, so as to establish the correspondence between the field measured data and the experimental data, and form a "true value benchmark" system that integrates the two.

[0119] S1.4.2: Data Labeling: Since the objective of this application is to more accurately predict the correct filling equipment commands based on filling environment data and hydraulic actuation parameter data, the filling equipment command data can be regarded as a multi-dimensional data set. The "labels" are then used. Data annotation tools combined with a human expert review mechanism are used to professionally annotate and verify the above "training-label" pairs to ensure the accuracy and consistency of the data labels.

[0120] Step S2 specifically includes:

[0121] S2.1: Multi-framework-based multi-dimensional sensing set for filling regions Feature Unified Mapping: Due to This dataset is a multi-source heterogeneous dataset, with different data types possessing their own specific structural features and information distribution characteristics. Therefore, a multi-frame neural network architecture is employed to perform feature mapping on various data sources. Through this multi-frame mapping mechanism, perceptual data from different modalities are uniformly projected onto the same low-dimensional feature space, forming a unified set of environmental state features. And the optimal set of multi-frameworks after training, such as Figure 3 As shown.

[0122] S2.1.1: Data type classification: If defined ,in Represents time-series data. Represents spatial displacement type data, Represents image / point cloud type data, This represents discrete data. Therefore, it is necessary to classify the data according to its data type. It is divided into four subsets so that corresponding neural network processing frameworks can be designed accordingly.

[0123] S2.1.2: Multi-framework construction for feature mapping: Corresponding neural network frameworks are constructed to achieve efficient feature mapping based on the feature distribution and structural characteristics of different types of data. The data uses the LSTM framework. The class data uses the GCN framework. The class data uses the ResNet framework. The data types employ an MLP framework. Each framework includes encoder and decoder modules, forming a symmetric input-reconstruction structure. For example, an LSTM network for time-series data can be represented as: .

[0124] S2.1.3: Feature training based on reconstruction loss: using time-series data Taking the corresponding LSTM framework as an example, the reconstruction loss function is defined as follows: ;make By minimizing the aforementioned reconstruction loss, the LSTM framework is iteratively optimized using gradient descent. This process yields feature representations with high information preservation and discriminative power. and optimal framework .

[0125] S2.1.4: Constructing a set of environmental state features Frameworks: Following the method in step S2.1.3, the four types of data and their corresponding network frameworks are trained respectively, and finally summarized into the following two sets: , And the set of environmental state characteristics The data dimensions for each category are completely consistent.

[0126] S2.2: Hydraulic Actuator Parameter Set Feature mapping: targeting Correlated time-series data and discrete parameter data are mapped using a single feature extraction network to obtain the feature set of execution parameters of the hydraulic actuator. .in, The feature dimensions and steps in S2.1 The feature dimensions are consistent to facilitate subsequent cross-modal feature fusion and collaborative representation training.

[0127] S2.2.1: Feature Mapping Framework Construction: Building a Transformer-based encoder and decoder network architecture: .

[0128] S2.2.2: Feature training and feature construction based on reconstruction loss: Referring to the scheme in step S2.1.3, the optimized parameter set of the hydraulic actuator is obtained. Execution parameter feature set and the optimal framework after training and incorporate it into In the collection. After this step, and composition .

[0129] Step S3 specifically includes:

[0130] S3.1: Constructing the Filling State - Executing the Cooperative Representation Trainer : Construct a collaborative representation trainer specifically for coal pillar replacement operations using backfilling. This trainer is a trainable neural network that uses the environmental state feature set obtained in step S2. With the set of execution parameters As input, a joint encoding structure is designed to simultaneously capture the interaction between environmental state and execution information, ultimately outputting their respective fused reasoning representations. and .like Figure 4 As shown, the filling state – executing the collaborative representation trainer consists of the following modules:

[0131] The perception initialization module consists of an initial connection layer, a batch normalization (BN) layer, and an activation function, and is used to receive a set of environmental state features. With the set of execution parameters The data in the middle.

[0132] Geo-force feature extraction module: contains multiple residual units and downsampling convolutions to capture the spatial gradient features of geological and stress fields, and realize the structured encoding of geo-force signals.

[0133] The filling medium evolution coding module consists of a continuous stack of residual blocks, combined with batch normalization and ReLU activation, to adapt to multi-scale feature representation. It is used to characterize the changes in physical properties of high-water filling materials under different proportions and hydration stages, forming a spatiotemporal coding of material states.

[0134] Hydraulic control response encoding module: Combining multi-channel residual blocks and local attention mechanisms, it is used to model dynamic sequence features such as pressure and flow rate.

[0135] Collaborative fusion inference module: Includes cross-channel fusion unit and nonlinear projection layer to achieve feature alignment in multi-view latent space.

[0136] The architecture of the infill state-execution co-representation trainer is a hierarchical multi-branch fusion architecture: the input first extracts low-level features of the multi-dimensional infill zone environment and execution signals through a perception initialization block; then it is input into a geo-mechanical feature extraction module to further encode geological and stress information. The output is then split into two paths: one enters the infill medium evolution encoding module to model the physical evolution of the high-water infill body; the other enters the hydraulic control response encoding module to learn the dynamic features of the hydraulic execution process. The outputs of the two paths are concatenated and input into the co-fusion inference module for latent space fusion and mutual information coupling, forming a unified infill state-execution co-representation.

[0137] S3.2: Filling State - Execution Intrinsic Information Redundancy Removal: By minimizing the redundant mutual information between the filling region environmental state characteristics and the execution parameter characteristics, invalid correlations are weakened.

[0138] S3.2.1: Calculate the collaborative fusion reasoning representation: using the filling state-execution collaborative representation trainer built in step S3.1. Calculate about and Fusion reasoning representation and The calculation formula is: .

[0139] S3.2.2: Calculate the co-occurrence matrix among fused inference representations The calculation formula is: Then normalize the matrix. Its formula is ,in Representation matrix The sum of all elements in the matrix It is a square matrix with a dimension of D.

[0140] S3.2.3: Calculate the filling status - execute the intrinsic information redundancy. The calculation formula is:

[0141] :

[0142] in The first smoothing parameter can be set. Representation matrix The Line 1 Column elements, Indicates the first The sum of actions: , Indicates the first Sum of columns: .

[0143] S3.2.4: Redundancy Removal Training: To minimize To achieve this, update the filling state using gradient descent and execute the co-representation trainer. The parameters.

[0144] S3.2.5: Repeat steps S3.2.1-S3.2.4 for a total of Second-rate.

[0145] S3.3: Filling State - Execution of Cooperative Action Coupling Matching: Maximizing Coupling Matching under the Same Working Condition and Mutual information is used to align positive example representations and achieve dynamic coordination between states and actions.

[0146] S3.3.1: Computational Coupling Between Fusion Reasoning Representations First, obtain the fusion reasoning representation using the exact same method as in step S3.2.1. and .assumed and The first in The rows are respectively and Calculate the coupling using the following formula : ,in express and The number of rows. Then normalize the matrix in the same way as in step S3.2.2. Finally, by weighting the average, we get: The matrix It is also a square matrix, with the same dimension D as the matrix in step S3.2.2. The dimensions are consistent.

[0147] S3.3.2: Calculating Fill State - Performing Coupled Matching The calculation formula is:

[0148] ;

[0149] in This is a configurable second smoothing parameter. Representation matrix The Line 1 Column elements, Indicates the first The sum of actions: , Indicates the first Sum of columns: .

[0150] S3.3.3: Matching Enhancement Training: To Maximize To achieve this, update the filling state using gradient descent and execute the co-representation trainer. The parameters.

[0151] S3.3.4: Repeat steps S3.3.1-S3.3.3 (total). Second-rate.

[0152] S3.4: Filling State - Execution Co-representation Construction: Based on the training results of steps S3.1-S3.3, obtain the filling state - execution co-representation. The calculation formula is: = .

[0153] Steps S3.1-S3.3 constitute the complete deep learning training process. Steps S3.2 and S3.3 define the training objective, impose constraints on the trainer constructed in step S3.1, and optimize its parameters. After training, step S3.4 can be executed to obtain the final filling state—executing collaborative representation. and the optimal trainer .

[0154] Step S4 specifically includes:

[0155] S4.1: Building an Equipment Command Predictor: Constructing an equipment command predictor specifically for coal pillar replacement operations using backfill material. The trainer is a trainable neural network that performs a cooperative representation based on the filling state obtained in step S3. Input, output prediction instruction .like Figure 5 As shown, device command predictor It is composed of the following levels stacked in the order of introduction:

[0156] Collaborative information normalization layer: Contains initial convolution, batch normalization, and ReLU activation units. Used to perform collaborative representation on the filling state. Scale standardization and feature enhancement are performed to eliminate statistical differences between different physical quantities.

[0157] Device state resolution block: Consists of two residual units, including convolution, batch normalization, and shorting structures. Used to extract filling state and perform collaborative representation. The underlying information reflects the operating status, stability, and health characteristics of the hydraulic actuator.

[0158] Working condition feature analysis block: Stacked residual units with downsampling convolution to expand the receptive field. Used to identify macroscopic working condition patterns under different geological and backfilling environments and capture the changing patterns of backfilling state.

[0159] Control signal sensing block: This block combines multi-channel convolution with batch normalization and adds attention modules to some channels. It is used to enhance the latent feature dimensions most relevant to hydraulic control commands and learn key control factors.

[0160] Predictive inference layer: Composed of high-order residual units and nonlinear projection layers. Used to synthesize multi-scale features from the previous stage to form a nonlinear mapping capability to the target command.

[0161] Hydraulic control predictive head: a multi-layer fully connected network. It outputs continuous hydraulic command vectors to achieve accurate regression prediction of control quantities such as pump speed, valve position, and pressure.

[0162] S4.2: Based on Filling State - Execution Cooperative Representation The equipment command prediction training is conducted using a regression prediction scheme.

[0163] S4.2.1: Calculation of equipment instruction regression error, the calculation formula is as follows: ,in and These represent the instruction tag sets for filling equipment. and filling state - execution co-representation The Instructions and collaborative representations, and Indicates the number of data points. Instructions indicating a prediction.

[0164] S4.2.2: Device Command Prediction Training: To Maximize To achieve this, the collaborative representation trainer is updated using gradient descent. The parameters.

[0165] S4.3.3: Repeat steps S4.3.1-S4.3.2 for a total of This process is repeated to obtain the final optimized model. .

[0166] This application also proposes an intelligent control device for coal pillar replacement with filling material, comprising two parts: hardware and software.

[0167] The hardware of the control device consists of a computing and control module, a data access module, an execution driver module, a communication network module, a power protection module, and a human-machine interaction module.

[0168] The core of the equipment is the computing and control module, which runs the kernel of the intelligent control software for the filling body and performs collaborative representation reasoning and hydraulic command generation. The data access module is responsible for acquiring multi-dimensional sensor and equipment status data through the field gateway. The execution drive module converts the calculated control commands into real-time control signals for hydraulic pumps, valves, and related actuators to achieve precise operation of the coal pillar replacement process. The communication network module provides a high-speed data interaction channel between the equipment and the field gateway, the upper-level monitoring system, and the industrial internet platform, supporting Ethernet and industrial bus communication. The power protection module provides stable power supply for the equipment and has lightning, surge, overcurrent, and explosion protection to ensure safe and reliable operation of the equipment in complex mining environments. The human-machine interaction submodule is used for on-site information display, parameter adjustment, and emergency control operations.

[0169] The software portion of the control device mainly consists of an operating system module, intelligent control software, and an execution driver module.

[0170] The operating system module provides the basic environment and system services for equipment operation, including task scheduling, file management, network communication and security management; the intelligent control software is the intelligent control software for filling body replacement coal pillars built in step S6 above, which is used to realize real-time control of filling body replacement coal pillars; the execution drive module is responsible for interacting with the field hardware, completing data input, command output and communication protocol adaptation, so that control signals can be directly applied to hydraulic pumps, valves and other actuators.

[0171] Figure 6 A structural block diagram of a computer device according to a specific embodiment of this application is shown. Figure 6 As shown, the computer device includes a memory and a processor, the memory storing instructions executable on the processor. When the processor executes the instructions, it implements the methods described in the above embodiments. The number of memories and processors can be one or more. This computer device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The computer device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0172] The computer device may also include a communication interface for communicating with external devices and exchanging data. The devices are interconnected using different buses and can be mounted on a common motherboard or otherwise as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as a display device coupled to the interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple electronic devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). The bus can be divided into address buses, data buses, control buses, etc. For ease of illustration, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0173] Optionally, in a specific implementation, if the memory, processor, and communication interface are integrated on a single chip, then the memory, processor, and communication interface can communicate with each other through an internal interface.

[0174] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors or any conventional processor. It is worth noting that the processor can be a processor supporting advanced RISC machines (ARM) architecture.

[0175] This application provides a computer-readable storage medium (such as the memory described above) storing computer instructions that, when executed by a processor, implement the method provided in this application.

[0176] Optionally, the memory may include a stored program area and a stored data area, wherein the stored program area may store the operating system and application programs required for at least one function; the stored data area may store data created based on the use of the computer device for mapping. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the computer device for mapping via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligent control of coal pillar replacement by filling material, based on smart mines, characterized in that: Includes the following steps: S1: Multi-dimensional sensing data integration during the filling process: Unifying and integrating data from different types of sensors into a multi-dimensional data set that can be used for intelligent control and analysis. This multidimensional dataset Includes multidimensional sensing set of filling area Hydraulic actuator parameter set Simultaneously, a set of instruction tags for filling equipment was prepared. ; Step S1 specifically includes: S1.1: Data sensing and acquisition related to coal pillar replacement by filling body: Through industrial bus or industrial IoT platform, synchronously collect filling environment data and hydraulic actuator data at different replacement points, and at the same time, collect or record the corresponding filling equipment instruction information; The filling environment data includes multi-dimensional sensing data such as filling medium and pipeline parameters, surrounding rock and ground pressure information, environmental and safety conditions data, and hydraulic actuator data such as actuator elements, power source supply and return oil parameters, control valves and pipeline control parameters. S1.2: Loading of data related to coal pillar replacement by filling body: Load the collected multi-dimensional sensing data and filling equipment instruction information into the data lake warehouse system of the data platform for persistent storage and management; S1.3: Preprocess the multidimensional sensing data collected in step S1.1; S1.4: Construction of Multidimensional Perception Dataset for Filling Process: The preprocessed data is manually labeled and verified based on expert knowledge to form an approved multidimensional perception set for the filling area. ; S2: For multidimensional data sets Feature extraction: Utilizing a multi-framework set of multidimensional data sets Multi-source heterogeneous data are projected in a unified dimension to form a feature set. Feature set Including multi-dimensional sensing sets for filling areas Feature extraction results And parameter sets for hydraulic actuators Feature extraction results ,in It is a set of environmental state characteristics. For the set of execution parameter features; S3: Filling State - Execution Co-representation Training: Constructing the Filling State - Execution Co-representation Trainer By filling the state-execution co-representation trainer Set of environmental state characteristics With the set of execution parameters Perform joint training to obtain a filled state with cooperative properties and minimized redundant information - execute cooperative representation. And the optimal filling state - execute the co-representation trainer ; S4: Device Command Prediction Training Based on Cooperative Representation: Constructing a Device Command Predictor To fill the state - perform collaborative representation As input, fill the device instruction label set P with labels, and train the device instruction predictor. Ultimately, the optimal device instruction predictor is obtained. ; S5: Construct the intelligent control kernel for coal pillar replacement with filling body: This involves using the model framework (Frameworks) obtained in steps S1-S4. and The sequentially stacked organization forms the intelligent filling body replacement coal pillar control core. The core receives data from various sensors at the coal pillar replacement site, and the sensor data is sequentially processed through Frameworks-> -> Subsequently, instructions regarding the hydraulic actuators were received. ; S6: Develop embedded intelligent control software for filling body replacement coal pillars: The software encapsulates the kernel built in step S5 and includes control interfaces for hydraulic actuators, directly automating the control of various devices involved in the filling body replacement coal pillar process.

2. The intelligent control method for replacing coal pillars with filling bodies according to claim 1, characterized in that: Step S2 specifically includes: S2.1: Multi-framework-based multi-dimensional sensing set for filling regions Unified Feature Mapping: A multi-framework neural network architecture is used to perform feature mapping on various data sources to form a unified set of environmental state features. And the optimal set of multiple frameworks after training; S2.2: Hydraulic Actuator Parameter Set Feature mapping: targeting Correlated time-series data and discrete parameter data are mapped using a single feature extraction network to obtain the feature set of execution parameters of the hydraulic actuator. ;in, Feature dimensions and environmental state feature set in step S2.1 The feature dimensions are consistent.

3. The intelligent control method for replacing coal pillars with filling bodies according to claim 2, characterized in that: Step S2.1 specifically includes: S2.1.1: Data Type Classification: Classify data types according to their data type. If divided into four subsets, then ,in Represents time-series data. Represents spatial displacement type data, Represents image / point cloud type data, Represents discrete data types; S2.1.2: Multi-framework construction for feature mapping: Corresponding neural network frameworks are constructed to achieve efficient feature mapping based on the feature distribution and structural characteristics of different types of data. The data uses the LSTM framework. The class data uses the GCN framework. The class data uses the ResNet framework. The data adopts the MLP framework, and each neural network framework contains an encoder module and a decoder module, forming a symmetrical input-reconstruction structure; S2.1.3: Feature Training Based on Reconstruction Loss: Define the reconstruction loss function as: ； make ; in, express , Decoders representing neural network frameworks corresponding to different types of data. Encoders that represent the neural network frameworks corresponding to different types of data; By minimizing the aforementioned reconstruction loss, gradient descent iterative optimization is performed on the four neural network frameworks. This process yields feature representations with high information preservation and discriminative power. and optimal framework ; S2.1.4: Constructing a set of environmental state features Frameworks: Following the method in step S2.1.3, the four types of data and their corresponding network frameworks are trained respectively, and finally summarized into the following two sets: , .

4. The intelligent control method for replacing coal pillars with filling bodies according to claim 1, characterized in that: Filling State - Execution Co-representation Trainer It is a hierarchical, multi-branch fusion architecture. Includes the following modules: The perception initialization module consists of an initial connection layer, a batch normalization (BN) layer, and an activation function, and is used to receive a set of environmental state features. With the set of execution parameters Data in the middle; Geo-force feature extraction module: contains multiple residual units and downsampling convolutions to capture the spatial gradient features of geological and stress fields, and realize the structured encoding of geo-force signals; The filling medium evolution coding module consists of a continuous stack of residual blocks, combined with batch normalization and ReLU activation, to adapt to multi-scale feature expression. It is used to characterize the changes in physical properties of high-water filling bodies under different ratios and hydration stages, forming a spatiotemporal coding of material states. Hydraulic control response encoding module: Combining multi-channel residual blocks and local attention mechanisms to model dynamic sequence features; Collaborative fusion inference module: Includes cross-channel fusion unit and nonlinear projection layer to achieve feature alignment in multi-view latent space.

5. The intelligent control method for replacing coal pillars with filling bodies according to claim 4, characterized in that: Step S3 specifically includes: S3.1: Constructing the Filling State - Executing the Cooperative Representation Trainer ; S3.2: Filling State - Execution Intrinsic Information Redundancy Reduction: By minimizing the redundant mutual information between the filling region environmental state characteristics and the execution parameter characteristics, invalid correlations are reduced, including: S3.2.1: Calculate the collaborative fusion reasoning representation: using the filling state-execution collaborative representation trainer built in step S3.

1. Calculate about and Fusion reasoning representation and The calculation formula is: ; S3.2.2: Calculate the co-occurrence matrix among fused inference representations The calculation formula is: Then normalize the matrix. Its formula is ,in Representation matrix The sum of all elements in the matrix It is a square matrix with dimension D; S3.2.3: Calculate the filling status - execute the intrinsic information redundancy. The calculation formula is: ： in The first smoothing parameter is configurable. Representation matrix The Line number Column elements, Indicates the first The sum of actions: , Indicates the first Sum of columns: ; S3.2.4: Redundancy Removal Training: To minimize To achieve this, update the filling state using gradient descent and execute the co-representation trainer. Parameters; S3.2.5: Repeat steps S3.2.1-S3.2.4 for a total of Second-rate; S3.3: Filling State - Execution of Cooperative Action Coupling Matching: Maximizing Coupling Matching under the Same Working Condition and Mutual information, aligning positive example representations, and achieving dynamic coordination between states and actions, including: S3.3.1: Computational coupling between fusion inference representations First, obtain the fusion reasoning representation using the exact same method as in step S3.2.

1. and ,assumed and The first in The rows are respectively and Calculate the coupling using the following formula : ,in express and The number of rows; then normalize the matrix in the same way as in step S3.2.

2. Finally, we calculate the weighted average to get: The matrix It is also a square matrix, with the same dimension D as the matrix in step S3.2.

2. The dimensions are consistent; S3.3.2: Calculating Fill State - Performing Coupled Matching The calculation formula is: ； in This is a configurable second smoothing parameter. Representation matrix The Line number Column elements, Indicates the first The sum of actions: , Indicates the first Sum of columns: ; S3.3.3: Matching Enhancement Training: To Maximize To achieve this, update the filling state using gradient descent and execute the co-representation trainer. Parameters; S3.3.4: Repeat steps S3.3.1-S3.3.3 (total). Second-rate; S3.4: Filling State - Execution Co-representation Construction: Based on the training results of steps S3.1-S3.3, obtain the filling state - execution co-representation. The calculation formula is: = .

6. The intelligent control method for replacing coal pillars with filling bodies according to claim 1, characterized in that: Device command predictor It is composed of the following levels stacked in the order of introduction: Collaborative information normalization layer: Contains initial convolution, batch normalization, and ReLU activation units, used to perform collaborative representation on the filling state. Scale standardization and feature enhancement are performed to eliminate statistical differences between different physical quantities; Device state parsing block: Consists of two residual units, including convolution, batch normalization, and shorting structures, used to extract filling state - perform collaborative representation. The underlying information reflects the operating status, stability, and health characteristics of the hydraulic actuator; Working condition feature analysis block: stacked residual units with downsampling convolution, used to identify macroscopic working condition patterns under different geological and backfilling environments and capture the changing patterns of backfilling state; Control signal sensing block: By combining multi-channel convolution with batch normalization and adding attention modules to some channels, it is used to enhance the potential feature dimensions most relevant to hydraulic control commands and learn key control factors. Predictive inference layer: Composed of high-order residual units and nonlinear projection layers, it is used to synthesize multi-scale features from the previous stage to form a nonlinear mapping capability for the target command; Hydraulic control prediction head: a multi-layer fully connected network used to output continuous hydraulic command vectors.

7. The intelligent control method for replacing coal pillars with filling bodies according to claim 6, characterized in that: Step S4 specifically includes: S4.1: Building a Device Command Predictor ; S4.2: Based on Filling State - Execution Cooperative Representation The device instruction prediction training includes: S4.2.1: Calculation of equipment instruction regression error, the calculation formula is as follows: ,in and These represent the instruction tag sets for filling equipment. and filling state - execution co-representation The Instructions and collaborative representations, Indicates the number of data points. Instructions indicating a prediction; S4.2.2: Device Command Prediction Training: To Maximize To achieve this, the collaborative representation trainer is updated using gradient descent. Parameters; S4.3.3: Repeat steps S4.3.1-S4.3.2 for a total of This process is repeated to obtain the final optimized model. .

8. A filling material replacement coal pillar intelligent control device for executing control commands output by the filling material replacement coal pillar intelligent control method as described in any one of claims 1-7, characterized in that: It includes two parts: hardware and software. The hardware consists of a computing and control module, a data access module, an execution driver module, a communication network module, a power protection module, and a human-computer interaction module. Among them, the computing and control module is the core of the equipment, which is used to run the kernel of the intelligent control software for the filling body and to perform collaborative representation reasoning and hydraulic command generation. The data access module is responsible for acquiring multi-dimensional sensor and equipment status data through the field gateway; The execution drive module converts the calculated control commands into real-time control signals for the hydraulic pump, valves and related actuators, enabling precise operation of the coal pillar replacement process; The communication network module provides a high-speed data interaction channel between the device and the field gateway, the upper-level monitoring system and the industrial Internet platform, and supports Ethernet and industrial bus communication; The power protection module provides a stable power supply for the equipment; The human-computer interaction submodule is used for on-site information display, parameter adjustment, and emergency control operations; The software component consists of an operating system module, intelligent control software, and an execution driver module. The operating system module provides the basic environment and system services for device operation. The intelligent control software is the intelligent control software for the filling body replacement coal pillar constructed in step S6, which is used to realize real-time control of the filling body replacement coal pillar. The execution driver module is responsible for interacting with the field hardware, completing data input, command output, and communication protocol adaptation.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

Images