Fully-mechanized mining hydraulic support electrohydraulic control system based on deep learning

Abstract

The invention discloses a fully mechanized coal mining hydraulic support electrohydraulic control system based on deep learning, and the system comprises a data collection and topology construction module which is used for collecting multi-source real-time data and constructing a support topology adjacency matrix; the data preprocessing and dynamic topology updating module is used for preprocessing the original data set and dynamically updating the initial topology adjacency matrix; the space-time diagram convolution depth autoregression modeling module is used for outputting a future multi-step prediction result set based on a space-time diagram convolution depth autoregression network model; the underground safety constraint judgment module is used for performing underground safety constraint judgment; the control instruction generating and issuing module is used for generating an electro-hydraulic control instruction set; and the execution feedback and model closed-loop updating module is used for collecting real-time state data, updating the preprocessed data set and enhancing the spatial-temporal characteristic sequence. According to the invention, the space-time diagram convolution depth autoregressive network is adopted to realize intelligent control of the fully-mechanized mining hydraulic support.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent mining of coal mines, and particularly relates to a fully mechanized hydraulic support electro-hydraulic control system based on deep learning. BACKGROUND

[0002] In the fully mechanized working face of a coal mine, a hydraulic support plays a key role in supporting a roof and ensuring safety of operation. Existing electro-hydraulic control methods mostly rely on fixed threshold logic and manually set control strategies. After collecting pressure, displacement, hydraulic oil temperature, and roof load data through sensors, the methods mainly use threshold alarms or rule control to complete lifting, telescoping, and following-machine actions. Such methods are insufficient in prediction and regulation when facing complex geological conditions, changes in advancing speed, and spatial coupling among multiple supports, and are difficult to handle high-dimensional dynamic characteristics of multi-source data. Although existing research has introduced time series prediction or a single deep learning model for trend analysis, it is mostly limited to independent point prediction and fails to effectively combine the spatial topological structure among supports and lacks systematic modeling of the collaborative relationship among multiple supports.

[0003] In recent years, some schemes have attempted to use graph convolution networks to evaluate the health status of mine equipment or use gray Markov and long short-term memory network methods to predict support parameters, but these methods mostly focus on static structures or single-channel time features and fail to achieve deep fusion of spatial relationships and time series. For the feature that the topological relationship of a hydraulic support changes constantly during coal mining and advancing, existing technologies cannot update structure information in real time, resulting in deviation between the prediction result and the actual state. At the same time, existing control strategies are usually in a serial mode of point prediction and threshold comparison, lack a mechanism of directly combining the probability distribution of the prediction result with underground safety constraints and forming a closed-loop control, and are difficult to quantify uncertainty and provide reliable basis for safety control.

[0004] Therefore, how to provide a fully mechanized hydraulic support electro-hydraulic control system based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One purpose of the present application is to provide a fully mechanized hydraulic support electro-hydraulic control system based on deep learning. The present application uses a spatio-temporal graph convolution deep autoregressive network model to dynamically model the pressure, displacement, hydraulic oil temperature, roof load, and production environment data of a hydraulic support, enhance spatio-temporal feature extraction, and multi-step joint conditional probability prediction, and directly couples the prediction result with underground safety constraints to generate executable electro-hydraulic control instructions, realizes a continuous prediction, control, and correction closed loop, and has the advantages of high prediction accuracy, strong real-time safety, good structure adaptive capability, and fast control response.

[0006] According to the fully mechanized hydraulic support electro-hydraulic control system based on deep learning of the present application, the system comprises:

[0007] a data collection and topology construction module, configured to collect multi-source real-time data of hydraulic supports of the fully-mechanized coal mining face, and construct a support topology adjacency matrix, to obtain an original data set and an initial topology adjacency matrix;

[0008] a data preprocessing and dynamic topology updating module, configured to preprocess the original data set, and dynamically update the initial topology adjacency matrix, to obtain a preprocessed data set and a dynamic topology adjacency matrix;

[0009] a spatio-temporal graph convolution deep autoregressive modeling module, configured to input the preprocessed data set and the dynamic topology adjacency matrix into a spatio-temporal graph convolution deep autoregressive network model, and output a future multi-step prediction result set;

[0010] a downhole safety constraint judgment module, configured to perform downhole safety constraint judgment based on the future multi-step prediction result set, and obtain a safety judgment result;

[0011] a control instruction generation and delivery module, configured to generate an electro-hydraulic control instruction set based on the safety judgment result, and deliver the electro-hydraulic control instruction set to an electro-hydraulic servo control module to perform corresponding actions;

[0012] an execution feedback and model closed-loop updating module, configured to collect real-time state data after the execution of the electro-hydraulic control instruction set, obtain a feedback data set, update the preprocessed data set and enhance a spatio-temporal feature sequence.

[0013] Optionally, the modules are realized through the following methods:

[0014] multi-source real-time data of hydraulic supports of the fully-mechanized coal mining face are collected, and a support topology adjacency matrix is constructed, to obtain an original data set and an initial topology adjacency matrix;

[0015] the original data set is preprocessed, and the initial topology adjacency matrix is dynamically updated, to obtain a preprocessed data set and a dynamic topology adjacency matrix;

[0016] the preprocessed data set and the dynamic topology adjacency matrix are input into a spatio-temporal graph convolution deep autoregressive network model, a spatio-temporal graph convolution operation is performed, and an enhanced spatio-temporal feature sequence is obtained;

[0017] the enhanced spatio-temporal feature sequence is input into an autoregressive sequence modeling layer, multi-step conditional probability modeling is performed, and a future multi-step prediction result set is obtained;

[0018] downhole safety constraint judgment is performed based on the future multi-step prediction result set, and a safety judgment result is obtained;

[0019] an electro-hydraulic control instruction set is generated based on the safety judgment result, and is input into an electro-hydraulic servo control module and executed;

[0020] Collect real-time state data after executing the electro-hydraulic control instruction set to obtain a feedback data set, update the pre-processed data set and enhance the time-space feature sequence.

[0021] Optionally, the multi-source real-time data of the hydraulic support of the fully mechanized coal mining face is collected, and a support topology adjacency matrix is constructed to obtain an original data set and an initial topology adjacency matrix, and the method specifically comprises the following steps:

[0022] Real-time measurement data of pressure, displacement, hydraulic oil temperature and roof load of the corresponding support are synchronously collected at each sampling time point through the pressure sensor, displacement sensor, hydraulic oil temperature sensor and roof load sensor installed on each hydraulic support of the fully mechanized coal mining face, and the pressure, displacement, hydraulic oil temperature and roof load data obtained by all supports at the same time point are uniformly collected to form a sensor original measurement data set;

[0023] Production shift information and working face advancing speed data are obtained at the corresponding sampling time through the production management system of the fully mechanized coal mining face, and geological parameter data related to the working face are obtained at the same time through a geological exploration device, and the collected production shift, advancing speed and geological parameter are uniformly arranged and collected in time sequence to form a production and geological information set;

[0024] The sensor original measurement data set and the production and geological information set are correspondingly merged at the same sampling time point, the data at each time point is sequentially arranged and uniformly stored to generate an original data set;

[0025] According to the hydraulic pipeline connection relationship between the hydraulic supports and the adjacent relationship of the spatial positions, the direct association between the supports is determined, and a topology structure is constructed with the numbers of all supports as nodes, a direct connection or spatial adjacent relationship between any two supports is marked as associated, and a direct connection and spatial adjacent relationship is marked as unassociated to generate an initial topology adjacency matrix.

[0026] Optionally, the original data set is pre-processed, and the initial topology adjacency matrix is dynamically updated to obtain a pre-processed data set and a dynamic topology adjacency matrix, and the method specifically comprises the following steps:

[0027] The original data set is timestamped and aligned on a unified sampling time axis, the pressure, displacement, hydraulic oil temperature, roof load, production shift, advancing speed and geological parameter are rearranged in the same time sequence, each time point corresponds to complete and synchronous multi-source data records, and an aligned data sequence is formed;

[0028] In the data sequence after completing timestamp alignment, abnormal detection is performed on each type of measurement value and state data item by item, the average level and fluctuation range of each type of data in the same time period are calculated, observation values exceeding the preset threshold range are identified and removed, and all valid data of the support at each time point after removing the abnormal values are re-arranged into a continuous aligned data sequence;

[0029] In the data sequence after completing timestamp alignment, observation values of each type of data are subjected to abnormal detection, the mean and standard deviation of the same type of data are calculated first, then the observation value at each time point is compared with the mean value, when the deviation of an observation value from the mean value exceeds the preset threshold multiple of the standard deviation, the observation value is determined as an abnormal value, and is removed from the aligned data sequence to obtain an effective time sequence without abnormal data;

[0030] In the effective time sequence after removing the abnormal values, interpolation is performed on the missing data generated due to sensor failure or communication interruption, the previous valid time point and the next valid time point of the missing point are determined in time sequence, and the estimated value of the missing time point is calculated according to the numerical change ratio of the two, the estimated value is filled in the corresponding position, and a continuous and complete multi-source data sequence is generated;

[0031] After obtaining the continuous and complete multi-source data sequence, normalization processing is performed on each type of data, the mean and standard deviation of each type of data are calculated first, then the observation value at each time point is subtracted from the mean value of each type of data and divided by the standard deviation, so that all data are converted into standardized values with a mean value of zero and a variance of one, forming a pre-processing data set;

[0032] After obtaining the normalized pre-processing data set, the relative position change between any two supports and whether the connection relationship changes are calculated in real time according to the displacement data of each support in the sensor original measurement data set and the connection relationship of the initial topology adjacency matrix, and these changes are mapped to the corresponding elements of the adjacency matrix, when the relative position change reaches a preset threshold or the connection relationship state changes, the corresponding adjacency relationship value is adjusted, and a dynamic topology adjacency matrix is generated.

[0033] Optionally, the inputting the pre-processing data set and the dynamic topology adjacency matrix into the spatio-temporal graph convolution deep autoregressive network model, performing spatio-temporal graph convolution operation, and obtaining an enhanced spatio-temporal feature sequence specifically includes:

[0034] Inputting the pre-processing data set and the dynamic topology adjacency matrix into the graph convolution input layer of the spatio-temporal graph convolution deep autoregressive network model;

[0035] In the graph convolution input layer, for each sampling time point, the support adjacency relationship represented by the dynamic topology adjacency matrix is combined with the preprocessed data set at the corresponding time point, the feature weighted sum of each support node and all adjacent support nodes is calculated, and after linear transformation by the weight matrix, a nonlinear activation function is applied to obtain a spatial correlation feature vector;

[0036] After the calculation of the spatial correlation feature vector is completed, the spatial correlation feature vector obtained by each support at the same time point is matched with the corresponding preprocessed time series feature vector according to the same time index, and the two types of features are merged into a single fusion feature vector in the form of vector splicing;

[0037] The fusion feature vector generated at each time point is sequentially input into the autoregressive modeling layer of the spatio-temporal graph convolution deep autoregressive network model, and the hidden state at the previous time point is combined and updated in the order of the time series, and the time series dependent information is continuously transmitted and accumulated through the loop calculation to obtain the time series hidden state;

[0038] All support time series hidden states output by the autoregressive modeling layer at each time point are collected and integrated in time sequence to generate an enhanced spatio-temporal feature sequence.

[0039] Optionally, the enhanced spatio-temporal feature sequence is input into the autoregressive sequence modeling layer to perform multi-step conditional probability modeling to obtain a future multi-step prediction result set, which specifically includes:

[0040] The enhanced spatio-temporal feature sequence and the dynamic topology adjacency matrix are input into the autoregressive sequence modeling layer, and the corresponding relationship and time sequence consistency of the two are maintained within the same spatio-temporal graph convolution deep autoregressive network model;

[0041] The enhanced spatio-temporal feature sequence and the dynamic topology adjacency matrix at the current time point are used as conditional inputs to initialize the prediction hidden state of each hydraulic support, and the initialization value is set to the time series hidden state of the support at the current time point, and the prediction step is also predicted;

[0042] According to the preset prediction step, recursive operation is performed from the first step to the Kth step, and the hidden state at the previous prediction time point and the dynamic topology adjacency matrix at the current time point are used to calculate a new prediction hidden state for each hydraulic support at each prediction time point, forming a multi-step prediction hidden state sequence;

[0043] After the recursive calculation of the multi-step prediction hidden state sequence is completed, the prediction hidden state of each hydraulic support at each prediction time point is input into the joint mapping function together with the dynamic topology adjacency matrix at the corresponding time point, and the two are comprehensively calculated and mapped to generate a multi-step conditional probability distribution parameter;

[0044] Based on the pressure prediction distribution mean, the pressure prediction distribution standard deviation, the displacement prediction distribution mean and the displacement prediction distribution standard deviation obtained for each hydraulic support at each prediction time point, the pressure variable and the displacement variable in the range of all prediction steps are jointly modeled, a joint conditional probability distribution capable of simultaneously describing the overall changes of future multiple steps of pressure and displacement and their mutual relationship is constructed in the form of multivariate normal distribution, and in the joint conditional probability distribution, a vector composed of the pressure and displacement prediction means at each prediction time point is taken as the overall mean, and a covariance matrix covering the range of prediction steps is taken as the overall covariance, to obtain a future multiple-step joint conditional probability distribution set;

[0045] According to the specific value selected from the numerical range between zero and one of the set confidence level parameter, and combined with the quantile coefficient of the standard normal distribution, the interval calculation of the pressure and displacement joint conditional probability distribution of each hydraulic support in the future multiple-step prediction time range is performed, the overall mean vector at each prediction time point is multiplied by the quantile coefficient after the square root of the diagonal elements of the corresponding covariance matrix to form the upper and lower offsets, and the upper and lower offsets are added and subtracted on the overall mean vector respectively to obtain a future multiple-step joint confidence interval set.

[0046] The future multiple-step joint conditional probability distribution set and the future multiple-step joint confidence interval set are combined to generate a future multiple-step prediction result set.

[0047] Optionally, the safety constraint judgment is performed based on the future multiple-step prediction result set, and a safety judgment result is obtained, which specifically includes:

[0048] From the future multiple-step prediction result set, the upper and lower limits of the confidence interval of the support pressure prediction component and the upper and lower limits of the confidence interval of the displacement prediction component are extracted for each hydraulic support and each prediction time step in turn, and the pressure confidence interval and the displacement confidence interval at each time step are classified and arranged in time sequence to generate a component confidence interval set for each support.

[0049] After forming the component confidence interval set of each hydraulic support, the safety constraint parameters are set, and the maximum support force threshold, the hydraulic safety pressure threshold and the displacement limit threshold are uniformly arranged into a safety constraint parameter set.

[0050] The safety constraint judgment operation is performed in combination with the component confidence interval set and the safety constraint parameter set of each hydraulic support, and when the upper limit of the pressure confidence interval exceeds the hydraulic safety pressure threshold, or the upper limit of the displacement confidence interval exceeds the displacement limit threshold, or the support force prediction value exceeds the maximum support force threshold, the safety state of the support at the corresponding prediction time step is marked as not meeting the safety requirement, and if none of the above conditions is triggered, the safety state of the support at the corresponding prediction time step is marked as meeting the safety requirement, to form a safety state marking result.

[0051] After the safety state label of each hydraulic support at all predicted time steps is completed, the safety state label results of all supports at each predicted time step are summarized one by one according to the support number and time sequence, and are uniformly arranged to form a safety judgment result set covering all supports and all prediction steps.

[0052] Optionally, the electro-hydraulic control instruction set is generated based on the safety judgment result, input into the electro-hydraulic servo control module and specifically includes:

[0053] According to the safety judgment result set and the future multi-step prediction result set, the safety state label, the pressure prediction mean value and the displacement prediction mean value of each hydraulic support at each predicted time step are extracted, and the three data are matched and combined according to the support number and time sequence to generate a control decision input vector;

[0054] According to the control decision input vector formed by each hydraulic support at each predicted time step, a pre-set control rule function is called for operation, and the operation result is comprehensively converted into electro-hydraulic control instructions including lifting, telescoping and following the machine with the machine three types of actions;

[0055] The electro-hydraulic control instructions generated by each hydraulic support at all predicted time steps are classified and summarized according to the support number and time sequence, and the lifting, telescoping and following the machine with the machine action instructions of each support at each predicted time point are included in a unified data structure one by one to form an electro-hydraulic control instruction set;

[0056] The electro-hydraulic control instruction set generated in the previous step is input into the electro-hydraulic servo control module according to the time sequence, and the electro-hydraulic servo control module analyzes the lifting, telescoping and following the machine with the machine action instructions corresponding to each support at each predicted time step in the set in turn, and executes the actions matched with the instructions through the drive proportional valve and the hydraulic actuator.

[0057] Optionally, the real-time state data after executing the electro-hydraulic control instruction set is collected to obtain a feedback data set, and the pre-processing data set is updated and the time-space feature sequence is enhanced, specifically including:

[0058] After executing the electro-hydraulic control instruction set, the latest observation data of each support is collected in real time through the sensor installed on the hydraulic support, and the production shift information, the advancing speed information and the geological parameters are recorded synchronously at the same sampling time point to form a real-time feedback data set;

[0059] The real-time feedback data set is time-stamped, aligned, outlier removed, missing completed and normalized, and the dynamic topological adjacency matrix is updated according to the support displacement information in the feedback data to obtain an updated pre-processing feedback data set;

[0060] The updated pre-processed feedback data set and the updated dynamic topology adjacency matrix are input into a graph convolution input layer of the spatio-temporal graph convolution deep auto-regressive network model, a spatio-temporal graph convolution operation is performed, new spatial correlation features are extracted, and the new spatial correlation features are spliced with corresponding feedback time sequence features to form an enhanced feedback spatio-temporal feature sequence;

[0061] Based on the enhanced feedback spatio-temporal feature sequence, the parameter weights and hidden states in the spatio-temporal graph convolution deep auto-regressive network model are updated, and the enhanced spatio-temporal feature sequence is continuously iteratively corrected to obtain a corrected enhanced spatio-temporal feature sequence.

[0062] The corrected enhanced spatio-temporal feature sequence is replaced and updated to the original enhanced spatio-temporal feature sequence, and the updated pre-processed feedback data set is replaced and updated to the original pre-processed data set, so that the continuous closed-loop operation of prediction, control and correction is realized.

[0063] The beneficial effects of the present application are:

[0064] The present application realizes dynamic topology modeling and joint conditional probability prediction of the hydraulic support multi-source real-time data by constructing a spatio-temporal graph convolution deep auto-regressive network, and outputs the probability distribution and confidence interval of the future multi-step pressure and displacement in the prediction stage, so that the spatial correlation and time dependence relationship between the supports are uniformly modeled and updated in real time. Compared with the traditional method which depends on fixed threshold and single time sequence prediction, the present application can timely reflect the change of support topology structure in the coal mining process, and ensure the matching degree of the prediction result and the actual running state, thereby providing complete and reliable spatio-temporal feature input for subsequent control instruction generation.

[0065] By coupling the prediction result with the underground safety constraints constituted by the maximum support force, the hydraulic safety pressure and the displacement limit, the present application realizes the continuous closed loop of prediction, judgment and instruction generation in the model, so that the uncertainty can be quantified and the potential risk can be evaluated before the electro-hydraulic control instruction is issued, thereby improving the safety and response speed of the control process. In combination with the continuous updating of the pre-processed data set and the enhanced spatio-temporal feature sequence based on the real-time feedback data, the whole system can maintain long-term stable operation and adapt to the dynamic change of the coal mining environment and production parameters, and ensure that the hydraulic support group realizes intelligent control with high precision and low delay under complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0067] Fig. 1 A method flowchart of the electro-hydraulic control system of the fully mechanized hydraulic support based on deep learning is provided.

[0068] Fig. 2 A structure schematic diagram of a spatiotemporal graph convolution deep autoregressive network model in a fully mechanized hydraulic support electro-hydraulic control system based on deep learning is provided for the present application.

[0069] Fig. 3 A structure schematic diagram of a fully mechanized hydraulic support electro-hydraulic control system based on deep learning is provided for the present application. DETAILED DESCRIPTION

[0070] The present application will now be further described in detail with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the present application in a schematic manner, and thus only show the components related to the present application.

[0071] REFERENCE Figs. 1-3 A fully mechanized hydraulic support electro-hydraulic control system based on deep learning, comprising:

[0072] A data acquisition and topology construction module is configured to acquire multi-source real-time data of hydraulic supports in a fully mechanized working face and construct a support topology adjacency matrix, thereby obtaining an original data set and an initial topology adjacency matrix.

[0073] A data preprocessing and dynamic topology updating module is configured to preprocess the original data set and dynamically update the initial topology adjacency matrix, thereby obtaining a preprocessed data set and a dynamic topology adjacency matrix.

[0074] A spatiotemporal graph convolution deep autoregressive modeling module is configured to input the preprocessed data set and the dynamic topology adjacency matrix into a spatiotemporal graph convolution deep autoregressive network model and output a future multi-step prediction result set.

[0075] A downhole safety constraint judgment module is configured to perform downhole safety constraint judgment based on the future multi-step prediction result set and obtain a safety judgment result.

[0076] A control instruction generation and delivery module is configured to generate an electro-hydraulic control instruction set based on the safety judgment result and deliver it to an electro-hydraulic servo control module for execution of corresponding actions.

[0077] An execution feedback and model closed-loop updating module is configured to acquire real-time state data after execution of the electro-hydraulic control instruction set, obtain a feedback data set, update the preprocessed data set, and enhance the spatiotemporal feature sequence.

[0078] In this embodiment, the modules are realized through the following methods:

[0079] Multi-source real-time data of hydraulic supports in a fully mechanized working face are acquired, and a support topology adjacency matrix is constructed, thereby obtaining an original data set and an initial topology adjacency matrix.

[0080] The original data set is preprocessed, and the initial topology adjacency matrix is dynamically updated to obtain a preprocessed data set and a dynamic topology adjacency matrix;

[0081] The preprocessed data set and the dynamic topology adjacency matrix are input into a spatio-temporal graph convolution deep autoregressive network model, a spatio-temporal graph convolution operation is performed, and an enhanced spatio-temporal feature sequence is obtained;

[0082] The enhanced spatio-temporal feature sequence is input into an autoregressive sequence modeling layer, a multi-step conditional probability modeling is performed, and a future multi-step prediction result set is obtained;

[0083] Based on the future multi-step prediction result set, a downhole safety constraint judgment is performed to obtain a safety judgment result;

[0084] Based on the safety judgment result, an electro-hydraulic control instruction set is generated, which is input into an electro-hydraulic servo control module and executed;

[0085] After collecting the real-time state data after executing the electro-hydraulic control instruction set, a feedback data set is obtained, the preprocessed data set is updated, and the spatio-temporal feature sequence is enhanced.

[0086] In the embodiment, the multi-source real-time data of the fully mechanized working face hydraulic support is collected, and a support topology adjacency matrix is constructed to obtain an original data set and an initial topology adjacency matrix. Specifically, it includes:

[0087] Through the pressure sensor, displacement sensor, hydraulic oil temperature sensor and roof load sensor installed on each hydraulic support of the fully mechanized working face, the real-time measurement data of the pressure, displacement, hydraulic oil temperature and roof load of the corresponding support are synchronously collected at each sampling time point, and the pressure, displacement, hydraulic oil temperature and roof load data of all supports obtained at the same time point are uniformly collected to form a sensor original measurement data set covering all hydraulic supports, which is used as the basis input for subsequent data processing and modeling;

[0088] Through the production management system of the fully mechanized working face, the production shift information and working face advancing speed data are obtained at the corresponding sampling time, and through the geological exploration device, the geological parameter data related to the working face are obtained at the same time. The collected production shift, advancing speed and geological parameters are uniformly arranged and summarized in time sequence to form a production and geological information set that can reflect the current production state and geological environment, which is used as the input for subsequent original data set construction;

[0089] The sensor raw measurement data set collected and sorted by the sensor and the production and geological information set obtained by the production management system and the geological exploration device are correspondingly merged at the same sampling time point, the data at each time point is sequentially arranged and uniformly stored, and finally a raw data set capable of simultaneously reflecting the running state of the hydraulic support, the production condition and the geological environment is generated as the basic input for subsequent modeling and analysis;

[0090] According to the hydraulic pipeline connection relationship and the spatial position adjacent relationship between the hydraulic supports, the direct association between the supports is determined, and a topology structure is constructed with the numbers of all the supports as nodes. On this basis, the direct connection or spatial adjacent relationship between any two supports is marked as associated, and the relationship without direct connection and spatial adjacent relationship is marked as unassociated. Finally, an initial topology adjacency matrix capable of completely representing the spatial layout and hydraulic connection characteristics of the support group is generated, providing a basic structure for subsequent dynamic topology updating and spatiotemporal feature extraction.

[0091] In the embodiment, the preprocessing of the raw data set and the dynamic updating of the initial topology adjacency matrix specifically include:

[0092] The raw data set is subjected to timestamp alignment processing on a unified sampling time axis, and the pressure, displacement, hydraulic oil temperature, roof load, production shift, advancing speed and geological parameters are rearranged in the same time sequence. Each time point corresponds to complete and synchronous multi-source data records, thereby forming an aligned data sequence organized by a unified time index, which is used as the input basis for subsequent outlier removal, missing completion and normalization processing;

[0093] In the data sequence after the timestamp alignment, each type of measurement value and state data is subjected to abnormality detection, the average level and fluctuation range of each type of data in the same time period are calculated, the observation values exceeding the preset threshold range are identified and removed, and the valid data of all the supports after the removal of the abnormal values at each time point are rearranged into a continuous aligned data sequence, which provides a clean data input for subsequent missing completion and normalization processing;

[0094] In the data sequence after the timestamp alignment, the observation values of each type of data are subjected to abnormality detection. First, the mean value and standard deviation of the same type of data are calculated, and then the observation value at each time point is compared with the mean value. When the deviation of an observation value from the mean value exceeds the standard deviation of the preset threshold multiple, the observation value is determined as an abnormal value and removed from the aligned data sequence, thereby obtaining an effective time sequence without abnormal data;

[0095] In the valid time series after removing outliers, the missing data caused by sensor failure or communication interruption is interpolated and completed, the previous valid time point and the next valid time point of the missing point are determined in time sequence, and the estimated value of the missing time point is calculated according to the numerical change ratio of the two, and the estimated value is filled in the corresponding position, so as to generate a continuous and complete multi-source data sequence on the time axis, and provide complete input for subsequent normalization processing;

[0096] After obtaining the continuous and complete multi-source data sequence, normalization processing is performed on each type of data. First, the mean and standard deviation of each type of data are calculated, then the observation value at each time point is subtracted from the mean of each type of data and divided by the standard deviation, so that all data are converted into standardized values with a mean of zero and a variance of one, thereby eliminating the dimensional difference between different physical quantities and forming a unified scale of preprocessed data set, providing a consistent numerical basis for subsequent dynamic topology update and feature extraction;

[0097] After obtaining the normalized preprocessed data set, the relative position change between any two supports and whether the connection relationship changes are calculated in real time according to the displacement data of each support in the original sensor measurement data set and the connection relationship of the initial topology adjacency matrix, and these changes are mapped to the corresponding elements of the adjacency matrix. When the relative position change reaches a preset threshold or the connection relationship state changes, adjust the corresponding adjacency relationship value, generate a dynamic topology adjacency matrix reflecting the current support group spatial layout and hydraulic connection characteristics, and provide real-time structure input for subsequent spatio-temporal feature extraction and prediction modeling.

[0098] In the embodiment, the input of the preprocessed data set and the dynamic topology adjacency matrix into the spatio-temporal graph convolution deep autoregressive network model includes:

[0099] The preprocessed data set and the dynamic topology adjacency matrix are input into the graph convolution input layer of the spatio-temporal graph convolution deep autoregressive network model;

[0100] In the graph convolution input layer, for each sampling time point, the support adjacency relationship represented by the dynamic topology adjacency matrix is jointly operated with the preprocessed data set at the corresponding time point, the feature weighted sum of each support node and all adjacent support nodes is calculated, and then a nonlinear activation function is applied after linear transformation through a weight matrix, so as to obtain a spatial correlation feature vector that can reflect the spatial correlation degree between the support and its adjacent supports at the corresponding time point;

[0101] After the calculation of the spatial correlation feature vector is completed, the spatial correlation feature vector obtained by each support at the same time point is matched with the corresponding pre-processed time sequence feature vector according to the same time index, and the two types of features are merged into a single fusion feature vector in the form of vector splicing, so that the fusion feature vector at each time point contains the time sequence feature of the support itself and the spatial correlation feature with the surrounding supports, providing complete space-time feature input for subsequent autoregressive modeling;

[0102] The fusion feature vector generated at each time point is sequentially input into the autoregressive modeling layer of the space-time graph convolution deep autoregressive network model, and is recursively combined and updated with the hidden state of the previous time point in the order of time sequence, and the time sequence dependent information is continuously transmitted and accumulated through loop calculation, so that the time sequence hidden state which can comprehensively reflect the relationship between the current fusion feature and the historical state is obtained, providing continuous time sequence modeling results for subsequent generation of enhanced space-time feature sequence;

[0103] All support time sequence hidden states output by the autoregressive modeling layer at each time point are collected and integrated in time sequence to generate an enhanced space-time feature sequence covering all support nodes, so that the enhanced space-time feature sequence completely represents the time variation feature of each support at each time point and its spatial correlation feature with adjacent supports, providing unified and structured input data for the next step of multi-step prediction and conditional probability modeling.

[0104] In the embodiment, the input of the enhanced space-time feature sequence into the autoregressive sequence modeling layer includes:

[0105] The enhanced space-time feature sequence and the dynamic topology adjacency matrix are input into the autoregressive sequence modeling layer, and the corresponding relationship and time sequence consistency of the two are maintained in the same space-time graph convolution deep autoregressive network model, so that the spatial correlation feature and the time sequence feature can participate in the operation at the same time when entering the prediction modeling stage, providing unified space-time coupling input for subsequent multi-step conditional probability modeling.

[0106] The enhanced space-time feature sequence and the dynamic topology adjacency matrix at the current time point are used as conditional input to initialize the prediction hidden state of each hydraulic support, and the initialization value is set as the time sequence hidden state of the support at the current time point, and the prediction step length is also predicted, which is used to support the subsequent autoregressive recursive operation;

[0107] According to a preset prediction step, recursive operation is sequentially performed from the first step to the Kth step, and for each hydraulic support, the hidden state at the previous prediction time point and the dynamic topological adjacency matrix at the current time point are used to jointly calculate a new prediction hidden state at each prediction time point, and through this step-by-step recursive manner, the time series information and the spatial topological relationship are synchronously updated and transmitted, so as to form a complete multi-step prediction hidden state sequence;

[0108] After the recursive calculation of the multi-step prediction hidden state sequence is completed, the prediction hidden state of each hydraulic support at each prediction time point and the dynamic topological adjacency matrix at the prediction time point are input into a joint mapping function, and the two are comprehensively calculated and mapped to generate multi-step conditional probability distribution parameters, including pressure prediction distribution mean, pressure prediction distribution standard deviation, displacement prediction distribution mean and displacement prediction distribution standard deviation, which provide complete probability parameter input for subsequent construction of future multi-step joint conditional probability distribution;

[0109] Based on the pressure prediction distribution mean, the pressure prediction distribution standard deviation, the displacement prediction distribution mean and the displacement prediction distribution standard deviation of each hydraulic support at each prediction time point, the pressure variable and the displacement variable in the range of all prediction steps are jointly modeled, and a joint conditional probability distribution capable of describing the overall change of future multi-step pressure and displacement and their mutual relationship is constructed in the form of multivariate normal distribution, and in the joint conditional probability distribution, a vector composed of the pressure and displacement prediction means at each prediction time point is taken as the overall mean, and a covariance matrix covering the range of prediction steps is taken as the overall covariance, so as to obtain a complete set of future multi-step joint conditional probability distributions representing the time correlation of pressure and displacement;

[0110] According to the value of the set confidence level parameter in the range of zero to one, and combined with the quantile coefficient of the standard normal distribution, the pressure and displacement joint conditional probability distribution of each hydraulic support in the future multi-step prediction time range is calculated, the overall mean vector at each prediction time point is multiplied by the square root of the diagonal element of the corresponding covariance matrix to form the upper and lower offsets, and the upper and lower offsets are added and subtracted on the overall mean vector respectively to obtain a set of pressure and displacement joint confidence intervals covering all prediction time points, which are used to provide quantitative risk boundaries for subsequent safety constraint judgment and control instruction generation;

[0111] The future multi-step joint conditional probability distribution set and the future multi-step joint confidence interval set are combined to generate a future multi-step prediction result set.

[0112] In the embodiment, the underground safety constraint judgment based on the future multi-step prediction result set specifically includes:

[0113] From the future multi-step prediction result set, for each hydraulic support and each prediction time step, the upper and lower limits of the support pressure prediction component confidence interval and the upper and lower limits of the displacement prediction component confidence interval are extracted, and the pressure confidence interval and the displacement confidence interval of each time step are classified and arranged in chronological order. Finally, a component confidence interval set covering all prediction steps is generated for each support, which records the interval boundaries of pressure and displacement at each time point in the future multi-step prediction, and provides accurate interval input data for subsequent safety constraint judgment;

[0114] After forming the component confidence interval set of each hydraulic support, set the underground safety constraint parameters, unify the maximum support force threshold for limiting the support capacity, the hydraulic safety pressure threshold for limiting the safety range of the hydraulic system, and the displacement limit threshold for restricting the stability of the support structure into a safety constraint parameter set, and determine the safety constraint parameter set as the unified comparison and judgment standard for subsequent safety state judgment;

[0115] Combine the component confidence interval set and the safety constraint parameter set of each hydraulic support to perform safety constraint judgment operation. When the upper limit of the pressure confidence interval exceeds the hydraulic safety pressure threshold, or the upper limit of the displacement confidence interval exceeds the displacement limit threshold, or the support force prediction value exceeds the maximum support force threshold, the safety state of the support at the corresponding prediction time step is marked as not meeting the safety requirement. If none of the above conditions is triggered, the safety state of the support at the corresponding prediction time step is marked as meeting the safety requirement, to form the safety state marking result of each support in the range of all prediction steps;

[0116] After completing the safety state marking of each hydraulic support at all prediction time steps, the safety state marking results of all supports at each prediction time step are summarized one by one according to the support number and time sequence, and a safety judgment result set covering all supports and all prediction steps is formed, which records the state information of each support meeting or not meeting the safety requirement in the future multi-step prediction range.

[0117] In this embodiment, the generation of the electro-hydraulic control instruction set based on the safety judgment result set, input into the electro-hydraulic servo control module and executed specifically includes:

[0118] Combine the safety judgment result set and the future multi-step prediction result set, extract the safety state marking, pressure prediction mean value and displacement prediction mean value of each hydraulic support at each prediction time step, and match and combine these three data according to the support number and time sequence to generate a control decision input vector, so that each control decision input vector contains the safety state information and core prediction parameters of the support at the corresponding prediction time step, providing a complete and consistent data basis for the calculation of subsequent electro-hydraulic control instructions;

[0119] According to the control decision input vector formed by each hydraulic support at each prediction time step, a pre-set control rule function is called for operation. The control rule function determines whether action is needed according to the safety state mark of the support at the prediction time point, and calculates the amplitude and direction of the lifting, telescoping and machine following action according to the difference relationship between the pressure prediction mean value and the displacement prediction mean value and the corresponding threshold value. The operation result is comprehensively converted into electro-hydraulic control instructions containing three types of actions of lifting, telescoping and machine following, and the complete and directly deliverable control instruction result is output for each support at the corresponding prediction time step;

[0120] The electro-hydraulic control instructions generated by each hydraulic support at all prediction time steps are classified and summarized according to the support number and time sequence, and the lifting, telescoping and machine following action instructions of each support at each prediction time point are sequentially included in a unified data structure, and finally an electro-hydraulic control instruction set covering all supports and all prediction steps is formed, providing a complete and orderly control instruction set for subsequent input into the electro-hydraulic servo control module and driving the proportional valve and hydraulic actuator;

[0121] The electro-hydraulic control instruction set generated in the previous step is input into the electro-hydraulic servo control module in time sequence, and the electro-hydraulic servo control module sequentially analyzes the lifting, telescoping and machine following action instructions of each support in the set at each prediction time step, and executes the actions matched with the instructions through driving the proportional valve and hydraulic actuator, thereby completing the automatic support control process based on the safety determination result within the entire prediction time range.

[0122] In the embodiment, the real-time state data after executing the electro-hydraulic control instruction set is collected to obtain a feedback data set, and the pre-processing data set is updated and the spatio-temporal feature sequence is enhanced, which specifically includes:

[0123] After executing the electro-hydraulic control instruction set, the latest observation data of each support is collected in real time through the sensors installed on the hydraulic support, and the production shift information, the advancing speed information and the geological parameters are recorded synchronously at the same sampling time point to form a real-time feedback data set;

[0124] The real-time feedback data set is subjected to timestamp alignment, outlier rejection, missing completion and normalization processing, and the dynamic topology adjacency matrix is updated according to the support displacement information in the feedback data to obtain an updated pre-processing feedback data set;

[0125] The updated pre-processing feedback data set and the updated dynamic topology adjacency matrix are input into the graph convolution input layer of the spatio-temporal graph convolution deep autoregressive network model, the spatio-temporal graph convolution operation is performed, the new spatial correlation features are extracted, and the corresponding feedback time series features are spliced to form enhanced feedback spatio-temporal feature sequences;

[0126] Based on the enhanced feedback spatiotemporal feature sequence, the parameter weights and hidden states in the spatiotemporal graph convolutional deep autoregressive network model are updated, the enhanced spatiotemporal feature sequence is continuously iteratively corrected, and a corrected enhanced spatiotemporal feature sequence is obtained.

[0127] The corrected enhanced spatiotemporal feature sequence is replaced and updated to the original enhanced spatiotemporal feature sequence, and the updated preprocessed feedback data set is replaced and updated to the original preprocessed data set, so as to realize continuous closed-loop operation of prediction, control and correction.

[0128] Embodiment 1:

[0129] In order to verify the feasibility of the application in implementation, the application is applied to the intelligent control scene of the hydraulic support group of the fully mechanized working face of a large mine. The working face is equipped with hundreds of hydraulic supports. Influenced by complex geological conditions and changes in the advancing speed of the coal mining machine, the stress relationship between the supports, the hydraulic pressure and the displacement state may change at any time. The traditional electro-hydraulic control mode with fixed threshold and manual control as the core often has difficulty in making high-precision prediction of future multi-step operation state when facing multi-source heterogeneous data and real-time topology structure changes, and the quantification of uncertainty is insufficient, which limits the ability to discover risks in advance and adjust in time.

[0130] In this embodiment, by arranging pressure, displacement, hydraulic oil temperature, roof load sensors on each hydraulic support and connecting production shifts, advancing speed, geological parameters and other external information, high-frequency acquisition and dynamic topology construction of multi-source real-time data of the whole working face are realized. The collected raw data are aligned by time stamp, outliers are removed, missing values are completed and normalized, and combined with support displacement data and topology relationship changes to generate a dynamic topology adjacency matrix, ensuring real-time updating of data integrity and structure. Then, the preprocessed data and dynamic topology adjacency matrix are input into the spatiotemporal graph convolutional deep autoregressive network model, spatial correlation feature extraction and time series recursive modeling are completed in the same network, and the joint conditional probability distribution and confidence interval of pressure and displacement of each support within the next thirty minutes are directly output, providing a quantitative risk boundary for safety control.

[0131] Based on the future multi-step prediction results of the model output, the system performs real-time safety constraint judgment of the underground, including three key parameters of maximum support force, hydraulic safety pressure and displacement limit. When the prediction results show that the upper limit of the pressure of some supports in the future time period may approach or exceed the safety threshold, the system immediately generates corresponding electro-hydraulic control instructions, including pressure reduction, reduction of support elongation or adjustment of the following machine speed, and automatically sends them to the electro-hydraulic servo control module to complete accurate operation through proportional valves and hydraulic actuators. At the same time, the real-time state data after execution is collected again and fed back to the model to update the pre-processing data and enhance the spatio-temporal feature sequence, thereby forming a continuous prediction-control-correction closed loop, enabling the system to continuously maintain prediction accuracy and control stability in a dynamic environment.

[0132] In the continuous running verification process, the system works continuously for 120 hours with a sampling period of 1 minute, collecting more than 40 million data records. Compared with the existing traditional threshold control method in the mine, the invention has obvious advantages in key indicators: the average absolute error of support pressure prediction is reduced from 0.82 MPa of the traditional method to 0.28 MPa, and the average absolute error of displacement prediction is reduced from 7.5 mm to 2.1 mm; the 95% confidence interval coverage rate of multi-step prediction reaches more than 96%, improving the reliability of risk warning. In terms of control response, the average time delay from risk discovery to instruction issuance is shortened from about 15 seconds of the traditional method to less than 3 seconds, and the execution accuracy of control instructions is more than 98%, effectively avoiding the safety hazards of roof subsidence and support instability. The system has not appeared prediction deviation accumulation or control failure in a long time running, verifying the ability of the invention to realize high-precision, low-delay and sustainable intelligent control of hydraulic supports in complex geological conditions.

[0133] Table 1 Comparison of key data of implementation effect of the invention

[0134]

[0135] As can be seen from the above table, the invention has improved compared with the traditional threshold control method in multiple key performance indicators of prediction and control of fully mechanized hydraulic supports. First, in terms of prediction accuracy, the average absolute error of support pressure prediction is reduced from 0.82 MPa to 0.28 MPa, with a decrease of about 66%; the average absolute error of displacement prediction is reduced from 7.5 mm to 2.1 mm, with a decrease of about 72%. This result shows that the modeling method based on spatio-temporal graph convolution deep autoregressive network can more fully utilize the spatial correlation and temporal dependence relationship between supports, improving the accuracy and stability of multi-step prediction of pressure and displacement.

[0136] Secondly, in terms of risk warning and uncertainty quantification, the 30-minute prediction 95% confidence interval coverage of the invention reaches 96%, which is better than the traditional method of 78%, an increase of about 23 percentage points; the risk warning advance time increases from 1.2 minutes to 5.5 minutes, more than 4 times. This shows that by directly generating conditional probability distribution and confidence interval, the invention can provide reliable risk prompt in an earlier time range, providing sufficient decision window for regulation.

[0137] In the control response and execution link, the average response time of the control instruction of the invention is shortened from 15 seconds to 3 seconds, a reduction of about 80%, the control instruction execution accuracy is improved from 90% to 98%, an increase of about 8 percentage points; at the same time, the cumulative rate of prediction deviation is reduced from 5% to less than 1%, and the continuous stable operation time of the system is extended from 72 hours to more than 120 hours. These data show that the invention greatly surpasses the traditional threshold control method in real-time and execution reliability, and can continuously maintain high efficiency and stability under complex working conditions.

[0138] In summary, based on the prediction modeling of spatio-temporal graph convolution deep autoregressive network, multi-step conditional probability output and closed-loop feedback update mechanism, the invention has realized the improvement in prediction accuracy, risk warning timeliness, control response speed and execution reliability, fully verifying its engineering application value and safety guarantee ability in the intelligent control scene of coal mine fully mechanized hydraulic support.

[0139] The above is only the preferred specific embodiment of the invention, but the protection scope of the invention is not limited thereto, any person skilled in the art within the technical scope disclosed by the invention, according to the technical scheme and inventive concept of the invention, equivalent replacement or change, should be covered within the protection scope of the invention.

Claims

1. A deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports, characterized in that, The method comprises the following steps: a data acquisition and topology construction module is used to acquire multi-source real-time data of hydraulic supports in a fully mechanized coal mining face and construct a support topology adjacency matrix, so as to obtain an original data set and an initial topology adjacency matrix; a data preprocessing and dynamic topology updating module is used to preprocess the original data set and dynamically update the initial topology adjacency matrix, so as to obtain a preprocessed data set and a dynamic topology adjacency matrix; a spatio-temporal graph convolution deep autoregressive modeling module is used to input the preprocessed data set and the dynamic topology adjacency matrix into a spatio-temporal graph convolution deep autoregressive network model, and output a future multi-step prediction result set; a downhole safety constraint judgment module is used to perform downhole safety constraint judgment based on the future multi-step prediction result set, and obtain a safety judgment result; a control instruction generation and delivery module is used to generate an electro-hydraulic control instruction set based on the safety judgment result, and deliver the electro-hydraulic control instruction set to an electro-hydraulic servo control module for execution of corresponding actions; an execution feedback and model closed-loop updating module is used to acquire real-time state data after execution of the electro-hydraulic control instruction set, obtain a feedback data set, update the preprocessed data set and enhance a spatio-temporal feature sequence.

2. The deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports according to claim 1, characterized in that, The modules are connected through the following methods: multi-source real-time data of hydraulic supports in a fully mechanized coal mining face are acquired, and a support topology adjacency matrix is constructed, so as to obtain an original data set and an initial topology adjacency matrix; the original data set is preprocessed, and the initial topology adjacency matrix is dynamically updated, so as to obtain a preprocessed data set and a dynamic topology adjacency matrix; the preprocessed data set and the dynamic topology adjacency matrix are input into a spatio-temporal graph convolution deep autoregressive network model, and a spatio-temporal graph convolution operation is performed, so as to obtain an enhanced spatio-temporal feature sequence; the enhanced spatio-temporal feature sequence is input into an autoregressive sequence modeling layer, a multi-step conditional probability modeling is performed, and a future multi-step prediction result set is obtained; downhole safety constraint judgment is performed based on the future multi-step prediction result set, and a safety judgment result is obtained; an electro-hydraulic control instruction set is generated based on the safety judgment result, and is input into an electro-hydraulic servo control module and executed; real-time state data after execution of the electro-hydraulic control instruction set are acquired, a feedback data set is obtained, the preprocessed data set is updated, and a spatio-temporal feature sequence is enhanced.

3. The deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports according to claim 2, characterized in that, The acquisition of multi-source real-time data of hydraulic supports in a fully mechanized coal mining face and the construction of a support topology adjacency matrix to obtain an original data set and an initial topology adjacency matrix specifically comprise the following steps: real-time measurement data of pressure, displacement, hydraulic oil temperature and roof load of each hydraulic support in the fully mechanized coal mining face are acquired through various sensors installed on each hydraulic support, so as to form a sensor original measurement data set; production shift information and working face advancing speed data are acquired, and geological parameter data related to the working face are acquired through a geological exploration device, so as to form a production and geological information set; the sensor original measurement data set and the production and geological information set are correspondingly merged at the same sampling time point, so as to generate an original data set; an initial topology adjacency matrix is constructed according to the pipeline connection relationship and the spatial proximity relationship between the hydraulic supports.

4. The deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports according to claim 2, characterized in that, The preprocessing comprises timestamp alignment, outlier removal, missing value completion and normalization processing.

5. The deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports according to claim 2, characterized in that, The pre-processed data set and the dynamic topology adjacency matrix are input into the space-time graph convolution deep autoregressive network model to perform space-time graph convolution operation to obtain an enhanced space-time feature sequence. The pre-processed data set and the dynamic topology adjacency matrix are input into the space-time graph convolution deep autoregressive network model to perform space-time graph convolution operation to obtain an enhanced space-time feature sequence. In the graph convolution input layer, for each sampling time point, the feature weighted sum of each support node and all adjacent support nodes is calculated, and after linear transformation through a weight matrix and application of a nonlinear activation function, a spatial correlation feature vector is obtained. The spatial correlation feature vector and the corresponding pre-processed time series feature vector are fused to obtain a fusion feature vector of the support through feature splicing. The fusion feature vector is input into the autoregressive modeling layer of the space-time graph convolution deep autoregressive network model to perform recursive update operation in the time dimension to obtain a time series hidden state. All support time series hidden states output by the autoregressive modeling layer at each time point are collected and integrated in time sequence to generate an enhanced space-time feature sequence.

6. The deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports according to claim 2, characterized in that, The enhanced space-time feature sequence and the dynamic topology adjacency matrix are input into the autoregressive sequence modeling layer, and their corresponding relationship and time sequence consistency are maintained within the same space-time graph convolution deep autoregressive network model. The enhanced space-time feature sequence and the dynamic topology adjacency matrix at the current time point are used as conditions to initialize the prediction hidden state, and the prediction step is set. Recursive operation is performed in sequence according to the preset prediction step to obtain a complete multi-step prediction hidden state sequence. Based on the multi-step prediction hidden state sequence, the multi-step conditional probability distribution parameters of pressure and displacement are generated through a joint mapping function. Based on the multi-step conditional probability distribution parameters, the pressure variable and the displacement variable within all prediction steps are jointly modeled to obtain a future multi-step joint conditional probability distribution set. Based on the confidence level parameter, the joint conditional probability distribution of pressure and displacement of each hydraulic support within the future multi-step prediction time range is calculated by combining the quantile coefficient of the standard normal distribution to obtain a future multi-step joint confidence interval set. The future multi-step joint conditional probability distribution set and the future multi-step joint confidence interval set are combined to generate a future multi-step prediction result set. Based on the future multi-step prediction result set, the downhole safety constraint judgment is performed to obtain a safety judgment result.

7. The deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports according to claim 2, characterized in that, Based on the future multi-step prediction result set, the upper and lower limits of the pressure confidence interval and the upper and lower limits of the displacement confidence interval are extracted from the joint confidence interval set for each hydraulic support and each prediction time step. The downhole safety constraint parameters are set, including the maximum support force threshold, the hydraulic safety pressure threshold, and the displacement limit threshold, and a safety constraint parameter set is constructed. ​ Based on the component confidence interval set and the safety constraint parameter set, a safety constraint judgment operation is performed, and when the upper limit of the pressure confidence interval exceeds the hydraulic safety pressure threshold, or the upper limit of the displacement confidence interval exceeds the displacement limit threshold, or the support force prediction value exceeds the maximum support force threshold, the safety state of the support is marked as not meeting the safety requirement; If none of the above conditions is triggered, the safety state of the support at the corresponding prediction time step is marked as meeting the safety requirement, and a safety state marking result is obtained; The safety state marking results of all supports at each prediction time step are summarized according to the support number and time sequence to form a safety judgment result set.

8. The deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports according to claim 2, characterized in that, The safety judgment result set is used to generate an electro-hydraulic control instruction set, which is input into an electro-hydraulic servo control module and specifically includes: Based on the safety judgment result set and the future multi-step prediction result set, a control decision input vector is generated for each hydraulic support at each prediction time step; Based on the control decision input vector and a preset control rule function, an electro-hydraulic control instruction is calculated for each support at the corresponding prediction time point; All electro-hydraulic control instructions are summarized according to the support number and time sequence to form an electro-hydraulic control instruction set; The electro-hydraulic control instruction set is input into the electro-hydraulic servo control module in time sequence, and actions matching the instructions are executed through the drive proportional valve and the hydraulic actuator.

9. The deep learning-based electro-hydraulic control system for fully mechanized hydraulic supports according to claim 2, characterized in that, The real-time state data after executing the electro-hydraulic control instruction set is collected to obtain a feedback data set, and the pre-processing data set is updated and the spatio-temporal feature sequence is enhanced, specifically including: After executing the electro-hydraulic control instruction set, the latest observation data of each support is collected in real time through the sensor, and production shift information, advance speed information and geological parameters are recorded synchronously at the same sampling time point to form a real-time feedback data set; The real-time feedback data set is pre-processed to obtain an updated pre-processed feedback data set, and the dynamic topology adjacency matrix is updated according to the support displacement information in the real-time feedback data set; The updated pre-processed feedback data set and the updated dynamic topology adjacency matrix are input into the spatio-temporal graph convolution deep autoregressive network model to form an enhanced feedback spatio-temporal feature sequence; Based on the enhanced feedback spatio-temporal feature sequence, the parameter weights and hidden states in the spatio-temporal graph convolution deep autoregressive network model are updated, and the enhanced spatio-temporal feature sequence is continuously iteratively corrected to obtain a corrected enhanced spatio-temporal feature sequence; The corrected enhanced spatio-temporal feature sequence replaces the original enhanced spatio-temporal feature sequence, and the updated pre-processed feedback data set replaces the original pre-processed data set.

Images