Dynamic prediction method for variable scale space-time area pressure events of intelligent mining face

Abstract

The present application relates to a kind of intelligent mining face variable scale space-time area coming pressure event dynamic prediction method, belong to coal mine intelligent technology field.The present application includes: the regional block of constructing the hydraulic support cluster of intelligent mining face;Calculate the coming pressure discriminant index value of each regional block;Extract the space-time area pressure feature of each regional block in T mining work cycle, and construct the space-time area pressure feature matrix of each regional block, then input the regional block cycle end resistance feature prediction model trained in advance, the cycle end resistance feature value of each regional block in future m mining work cycle is predicted by model;Calculate the local moran index of each regional block;Screen out the regional block of cycle end resistance feature value greater than its discriminant index value;According to the local moran index of all regional blocks and the aggregation area of each screened regional block predicted coming pressure event.The present application is relatively high in the prediction accuracy of coming pressure event.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent coal mine, and particularly relates to a dynamic prediction method for variable scale space-time area pressure event of intelligent mining working face. BACKGROUND

[0002] With the continuous exploitation of coal resources and the continuous increase of mining depth, the changes in geological conditions, the increase of mine pressure and the challenges of support technology brought by deep mining have greatly increased the risk of roof instability and accidents. According to statistics, accidents caused by roof problems account for 22.6% of coal mine accidents, which has become one of the main safety hazards in coal production, and has seriously endangered the safety of miners. Therefore, the prevention and control of roof disasters is particularly urgent. In fully mechanized working face, the working state of hydraulic support is an important representation of mine pressure characteristics, which directly reflects the stress condition and deformation characteristics of the roof. By detecting and analyzing the working state of the hydraulic support, the appearance characteristics of the mine pressure in the fully mechanized working face can be revealed, and the periodic changes of the mine pressure can be predicted, which provides an important basis for the prevention of roof disasters.

[0003] The periodic pressure of fully mechanized working face refers to the periodic fracture and subsidence of overburden rock (especially immediate roof and main roof) with the advance of coal mining working face, which leads to the periodic increase of roof pressure in fully mechanized working face. Timely identification and early warning of the occurrence of periodic pressure event can not only effectively reduce the probability of roof disaster, but also provide strong guarantee for the safety production of mine. The column pressure of hydraulic support is one of the effective features for periodic pressure event discrimination. With the advancement of intelligent mining working face (intelligent fully mechanized working face) construction, mine pressure monitoring has developed from only arranging pressure sensors on part of the hydraulic support columns to installing high-precision pressure sensors on each hydraulic support column, and realizing real-time acquisition and uploading of pressure values. The massive pressure data greatly enhances the fine monitoring ability of mine pressure in intelligent mining working face, provides strong data support for in-depth analysis of mine pressure law, and also poses new challenges for mine pressure prediction method.

[0004] The traditional prediction of the pressure event of the intelligent mining working face adopts a single hydraulic support pressure discrimination index, which is used to represent whether the upper support position of the hydraulic support is under pressure by judging whether the cycle end resistance of each hydraulic support column pressure is higher than the pressure discrimination index. However, for a one-time pressure event, the fracture of the roof of the intelligent mining working face is a whole dynamic process, and the process has obvious spatial and temporal differences. This difference is not only reflected in the changes of the roof stress and deformation at different time nodes, but also reflects the spatial and temporal correlation and continuity of the pressure of the intelligent mining working face. Therefore, although the traditional pressure event prediction method can predict the pressure of the whole intelligent mining working face or a fixed area to a certain extent, it often cannot fully consider the complexity of the local area and time evolution in the pressure process, resulting in low accuracy of the prediction result. SUMMARY

[0005] To solve the above technical problems, the present application provides a variable scale space-time region pressure event dynamic prediction method for an intelligent mining working face. The technical scheme of the present application is as follows:

[0006] A variable scale space-time region pressure event dynamic prediction method for an intelligent mining working face, comprising:

[0007] S1, constructing a regional block of the hydraulic support cluster of the intelligent mining working face;

[0008] S2, calculating the pressure discrimination index value of each regional block;

[0009] S3, extracting the space-time region pressure characteristics of each regional block in the T mining working cycles completed by the intelligent mining working face before the current mining working cycle, and constructing a space-time region pressure characteristic matrix of each regional block according to the space-time region pressure characteristics of each regional block;

[0010] S4, inputting the space-time region pressure characteristic matrix of each regional block into a pre-trained regional block cycle end resistance feature prediction model, and predicting the cycle end resistance feature value of each regional block in the future m mining working cycles by the regional block cycle end resistance feature prediction model;

[0011] S5, comparing the cycle end resistance feature value of each regional block in the future m mining working cycles with the pressure discrimination index value thereof, and screening the regional blocks whose cycle end resistance feature values in the future m mining working cycles are greater than the pressure discrimination index values thereof from all the regional blocks;

[0012] S6, calculating the local Morlet index of each regional block;

[0013] S7, according to the local Morlet index of all the regional blocks and the predicted pressure event aggregation area of each screened regional block.

[0014] Optionally, the S2 is used to calculate the area block i When the pressure discrimination index value is obtained, it is realized by formula (1):

[0015]

[0016] In formula (1), For regional blocks i The pressure discrimination index value; n For regional blocks i The number of hydraulic supports included; r is the total number of mining cycles currently completed in the intelligent mining working face; k is the variance coefficient, which is 0.8 to 1; For regional blocks i The resistance at the end of the j-th mining cycle;

[0017]

[0018] In formula (2), For regional blocks i The end-of-cycle resistance of the x-th hydraulic support in the j-th mining working cycle.

[0019] Optionally, the spatiotemporal regional pressure characteristics of each block include the end-of-cycle resistance, time-weighted average working resistance and terminal maximum resistance increase rate of all hydraulic supports of each block in each mining working cycle.

[0020] Alternatively, the time-weighted average working resistance of a hydraulic support in any area block during a mining working cycle is calculated using formula (3):

[0021]

[0022] In formula (3), is the time-weighted average work resistance; The number of sections of hydraulic support column pressure within a mining working cycle; For the a The pressure of the hydraulic support column within a time period; For the a The time interval corresponding to the time period;

[0023] The maximum resistance increase rate of a hydraulic support in any area block at the end of a mining working cycle is calculated by formula (4):

[0024]

[0025] In formula (4), is the maximum resistance increase rate at the end; It is the maximum increment of pressure of adjacent segmented hydraulic support columns within the last 5 minutes of a mining working cycle; for the corresponding time period.

[0026] Optionally, before the S4, further comprising:

[0027] S41, acquiring historical hydraulic support column pressure data of a historical mining process of the intelligent mining face;

[0028] S42, extracting cycle-end resistance, time-weighted average working resistance, and end maximum resistance increasing rate of all hydraulic supports of each regional block in the last T mining work cycles from the historical hydraulic support column pressure data, and combining these features in chronological order to form a model input feature matrix with a size of (T, 3n), n being the number of hydraulic supports included in each regional block;

[0029] S43, extracting cycle-end resistance feature values of each regional block in the future m mining work cycles from the historical hydraulic support column pressure data as the output of the regional block cycle-end resistance feature prediction model;

[0030] S44, training the regional block cycle-end resistance feature prediction model through the model input feature matrix and the corresponding output.

[0031] Optionally, in the S43, the cycle-end resistance feature values of the regional block in the future m mining work cycles are calculated through formulas (6) to (8): i

[0032]

[0033]

[0034]

[0035] In formulas (6) to (8), represents the cycle-end resistance of the regional block i in the future m mining work cycles, that is, the output of the regional block cycle-end resistance feature prediction model; represents the cycle-end resistance of the regional block i in the jth mining work cycle; represents the normalized weight value of the jth mining work cycle; represents the sum of the weight values of all work cycles; represents the weight value corresponding to the jth mining work cycle; is a decay factor used to control the rate of decay, and is taken as ; t is the time step of each mining work cycle.

[0036] ​Optionally, the regional block cycle end resistance feature prediction model is a CNN-BiLSTM-Attention network formed by sequentially connecting CNN, Bi-LSTM and Attention.

[0037] Optionally, the S6 is used to calculate the area block i When the local Moran index is calculated, it is realized by formula (9) to formula (11):

[0038]

[0039]

[0040]

[0041] In formula (9), is the local Moran index of region block i; is the total number of regional blocks; is the sum of the weights between all regional blocks; For regional blocks and The spatial weight between

[0042] In formula (10), is the normalized end-of-cycle resistance characteristic of block i; For regional blocks i The characteristic value of the end-of-cycle resistance; is the mean of the end-of-cycle resistance characteristic values ​​of all regional blocks; is the standard deviation of the end-of-cycle resistance characteristic values ​​of all regional blocks;

[0043] In formula (11), W represents the spatial weight matrix.

[0044] Optionally, if the region block i For the filtered area block, S7 includes:

[0045] S71, if the area block i The local Moran index Greater than or equal to the local Moran index threshold , then determine the area block i Pressure accumulation occurs around the area;

[0046] S72, according to the area block i The local Moran index and the local Moran index threshold of the front and back area blocks The size relationship determines the area block i The surrounding area is the gathering area of ​​pressure incidents.

[0047] Optionally, the S72 includes:

[0048] S721, block the area i The previous adjacent area block The local Moran index and Compare, if the area block The local Moran index and , then continue to compare the region blocks The local Moran index of the previous adjacent area block is until the local Moran index is less than The regional block, as the regional block i The starting area block of the surrounding pressure event;

[0049] S722, block the area i The adjacent area block The local Moran index and Compare, if the area block The local Moran index and , d is the total number of regional blocks, then determine the regional blocks i The termination area block of the surrounding pressure event is the area block ;if Continue comparing region blocks The local Moran index of the adjacent area block is until the local Moran index is less than The regional block, as the regional block i The termination region block surrounding incoming pressure events.

[0050] All the above optional technical solutions can be combined arbitrarily, and the present invention does not provide detailed descriptions of the structures after each combination.

[0051] By means of the above solution, the beneficial effects of the present invention are as follows:

[0052] By constructing the regional blocks of the hydraulic support cluster of the intelligent mining face, and calculating the weighting pressure discrimination index value of the regional block as the discrimination index of the weighting pressure event aggregation area, the complexity of the local area and the time evolution in the weighting pressure process is fully considered, so that the accuracy of the subsequent weighting pressure event aggregation area obtained based on the index is relatively high. By constructing the space-time regional pressure feature matrix of each regional block, and inputting it into the pre-trained regional block cycle-end resistance feature prediction model, after the model predicts the cycle-end resistance feature value of each regional block in the next m mining working cycles, the regional blocks that may occur weighting pressure events are screened based on the prediction result, and then the aggregation area of the roof weighting pressure event is determined according to each screened regional block and the local Moran index of each regional block, so that the variable-scale space-time regional weighting pressure event is realized. Real-time and automatic prediction, and a more accurate and real-time monitoring and early warning means is provided for the response to the weighting pressure event of the intelligent mining face.

[0053] The above description is only a summary of the technical scheme of the present application. In order to more clearly understand the technical means of the present application, and to implement the content of the description, the following will be described in detail with the preferred embodiments of the present application and with the help of the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is the flow chart of the variable-scale space-time regional weighting pressure event dynamic prediction method of the intelligent mining face provided by the embodiment of the present application.

[0055] Figure 2 It is the regional weighting pressure discrimination state scatter plot in the embodiment of the present application.

[0056] Figure 3 It is a segmented schematic diagram of the column pressure of a hydraulic support in one mining working cycle in the embodiment of the present application.

[0057] Figure 4 It is the composition structure schematic diagram of the CNN-BiLSTM-Attention network in the embodiment of the present application.

[0058] Figure 5 It is an example diagram of one training sample in the embodiment of the present application.

[0059] Figure 6 It is the flow chart of determining the surrounding weighting pressure event aggregation area of the regional block in the embodiment of the present application. i

[0060] Figure 7 It is the overall flow chart of the embodiment of the present application. DETAILED DESCRIPTION

[0061] The specific embodiments of the present application will be further described in detail below in combination with the drawings and examples. The following examples are used to illustrate the present application, but not to limit the scope of the present application.​

[0062] The method for dynamic prediction of variable-scale spatiotemporal regional pressure events of the intelligent mining working face provided by the embodiment of the present invention can be implemented by any electronic device with computing function, such as PC, mobile terminal or server. By utilizing the massive hydraulic support column pressure data of the intelligent mining working face and through the discrimination analysis of the end resistance area of ​​the hydraulic support cycle, the pressure manifestation characteristics of the intelligent mining working face cycle can be effectively characterized. Therefore, the embodiment of the present invention proposes a method for dynamic prediction of variable-scale spatiotemporal regional pressure events based on the massive hydraulic support column pressure data of the working face, which extracts the spatiotemporal characteristics of the historical hydraulic support column pressure of the regional block of the intelligent mining working face, and predicts the regional block cycle end resistance characteristics of the future mining working cycle by constructing a regional block cycle end resistance characteristic prediction model, and compares it with the regional block cycle end resistance characteristic to form a pressure discrimination state, and then constructs a variable-scale pressure regional autocorrelation aggregation model to realize the dynamic prediction of the future mining working cycle pressure events of the intelligent mining working face. In addition, during the mining process, the intelligent mining working face will carry out multiple mining work cycles. The regional block end-of-cycle resistance characteristic prediction model is trained through the historical hydraulic support column pressure data collected in the previous mining work cycle, so that the trained regional block end-of-cycle resistance characteristic prediction model can predict the end-of-cycle resistance characteristic value of the subsequent mining work cycle of the intelligent mining working face.

[0063] Combining the above content, such as Figure 1 As shown, the method for dynamically predicting pressure events in variable-scale spatiotemporal regions of intelligent mining working faces provided by the embodiment of the present invention can be implemented through the following steps S1 to S7:

[0064] S1, construct the regional block of hydraulic support cluster of intelligent mining working face.

[0065] Specifically, for the hydraulic support cluster of the intelligent mining working face, this step can sequentially construct regional blocks using a sliding window method. For example, if the hydraulic support cluster includes 100 hydraulic supports and each regional block includes 3 hydraulic supports, the first to third hydraulic supports are regional block 1, the second to fourth hydraulic supports are regional block 2, the third to fifth hydraulic supports are regional block 3, and so on.

[0066] The present embodiment does not impose a specific limit on the number of hydraulic supports included in each block; it can be set as needed during implementation. To ensure more accurate prediction of the area where pressure events are concentrated, the present embodiment configures each block to include 3-5 hydraulic supports, and the sliding window movement step size is one hydraulic support.

[0067] S2: Calculate the pressure discrimination index value of each area block.

[0068] In a specific embodiment, the S2 is in the calculation area block iThe coming pressure discrimination index value of the region block is obtained by formula (1) when the coming pressure discrimination index value of the region block is obtained.

[0069]

[0070] In formula (1), is the coming pressure discrimination index value of the region block i . n is the number of hydraulic supports included in the region block i . is the total number of mining work cycles that have been completed by the intelligent mining working face at present; is a variance coefficient, and is 0.8-1; is the resistance at the end of the jth mining work cycle of the xth hydraulic support in the region block i .

[0071]

[0072] In formula (2), is the resistance at the end of the jth mining work cycle of the xth hydraulic support in the region block i .

[0073] Taking n=3 as an example, the region blocks are sequentially constructed in a sliding window manner, the resistance at the end of each mining work cycle of all the region blocks is calculated one by one, the coming pressure discrimination index value of each region block is calculated according to formula (1), K the values are respectively 0.8, 0.9 and 1, whether the resistance at the end of each mining work cycle of each region block in the intelligent mining working face is greater than the coming pressure discrimination index value thereof is judged, the coming pressure characteristic state of the same hydraulic support in different region blocks is superimposed, and a region coming pressure discrimination state scatter plot is formed, as shown in the d graph, the e graph and the f graph in Figure 2 . The resistance at the end of the cycle greater than the discrimination index is displayed as a black scatter point. It can be known from the observation of the region coming pressure discrimination state scatter plot that the roof coming pressure of the intelligent mining working face causes the resistance at the end of each mining work cycle of each region block to be periodically higher than the region discrimination index, and in space, a certain number of hydraulic support region ranges are presented, and in time, the duration is generally 2-8 mining work cycles, which better depicts the overburden strata caving law of the intelligent mining working face. The region coming pressure discrimination index better presents the spatial and temporal periodicity of the intelligent mining working face, the coming pressure step distance is easier to obtain, and there are fewer noise points.

[0074] Therefore, the region coming pressure discrimination method reduces the confusion of the traditional method using a single hydraulic support to discriminate the coming pressure, the same hydraulic support in a continuous region can realize information sharing, the robustness of the discrimination result is enhanced through smoothing data, the spatial and temporal characteristics of the periodic coming pressure can be more effectively captured, and the method is more suitable for the analysis of the massive hydraulic support column pressure data of the intelligent mining working face.

[0075] S3, extract the space-time regional pressure characteristics of each regional block in T mining work cycles completed by the intelligent mining face before the current mining work cycle, and construct a space-time regional pressure characteristic matrix of each regional block according to the space-time regional pressure characteristics of each regional block.

[0076] Specifically, in order to better reveal the periodic weighting law of the intelligent mining face and improve the accuracy of the cycle-end resistance characteristic value predicted by the regional block cycle-end resistance characteristic prediction model, the input feature quantity of the regional block cycle-end resistance characteristic prediction model should be able to capture the working state of each hydraulic support in different mining work cycles, and also reflect the state characteristics of each hydraulic support in the spatial dimension, thereby providing more abundant and accurate input information for the regional block cycle-end resistance characteristic prediction model. Therefore, the space-time regional pressure characteristics of each regional block in the embodiment of the present application include the cycle-end resistance, time-weighted average working resistance and end maximum resistance increase rate of all hydraulic supports of each regional block in each mining work cycle as the input feature quantity of the regional block cycle-end resistance characteristic prediction model, which are fused to form the model input feature matrix.

[0077] The cycle-end resistance refers to the bearing resistance of the hydraulic support at the end of each mining work cycle, which usually shows a significant increase during periodic weighting and is an important characterization index of periodic weighting. The time-weighted average working resistance is obtained by weighting the average of the resistance of the hydraulic support with time, which reflects the comprehensive support condition in the entire mining work cycle and can more comprehensively describe the resistance change and its dynamic response to periodic weighting. The end maximum resistance increase rate is the maximum change rate of the resistance of the hydraulic support in the last 5 minutes of each work cycle, which usually indicates that pressure is accumulating before periodic weighting when the rate continuously increases in consecutive mining work cycles. The length of the mining work cycle of the hydraulic support is the time required for the hydraulic support to complete a complete column lifting, supporting and column lowering process, which reflects the advancing speed of the intelligent mining face, and the advancing speed of the intelligent mining face is one of the main factors affecting the mine pressure behavior.

[0078] In one specific embodiment, the cycle-end resistance is the column pressure value of the hydraulic support obtained at the end of a mining work cycle.

[0079] In one specific embodiment, the time-weighted average working resistance of a hydraulic support of any regional block in a mining work cycle is calculated by formula (3):

[0080]

[0081] In formula (3), is the time-weighted average working resistance; is the number of segments of the column pressure of the hydraulic support in a mining work cycle; is the column pressure of the hydraulic support in the i th segment of the mining work cycle;a The pressure of the hydraulic support column within a time period; For the a The time interval corresponding to the time period;

[0082] like Figure 3 As shown in FIG, it is a segmented schematic diagram of the pressure of a hydraulic support column in a mining working cycle. In the specific implementation process, the pressure sensor installed on each hydraulic support column collects the pressure of the hydraulic support column in real time and reports it to the electronic equipment. When the electronic equipment detects that the pressure of a hydraulic support column changes, it records the changed hydraulic support column pressure and the corresponding time period, forming a data structure as shown in FIG. Figure 3 The schematic diagram of the hydraulic support column pressure segmentation is shown in FIG. Figure 3 The values ​​of the parameters in formula (3) can be obtained from the hydraulic support column pressure segmentation diagram shown.

[0083] The maximum resistance increase rate of a hydraulic support in any area block at the end of a mining working cycle is calculated by formula (4):

[0084]

[0085] In formula (4), is the maximum resistance increase rate at the end; It is the maximum increment of pressure of adjacent segmented hydraulic support columns within the last 5 minutes of a mining working cycle; is the corresponding time period.

[0086] Specifically, Figure 3 Taking the hydraulic support column pressure segmentation diagram shown in the figure as an example, when calculating the maximum resistance increase rate at the end, the number of segments of the hydraulic support in the last 5 minutes of a mining working cycle is obtained, and the maximum increase in the pressure of the hydraulic support columns of the adjacent segments within the 5 minutes is calculated to obtain and .

[0087] S4, the spatiotemporal regional pressure feature matrix of each regional block is input into a pre-trained regional block end-of-cycle resistance feature prediction model, and the regional block end-of-cycle resistance feature prediction model predicts the end-of-cycle resistance feature value of each regional block in the next m mining working cycles.

[0088] It should be noted that before step S4, it is necessary to first train the regional block end-of-cycle resistance feature prediction model. Specifically, when training the regional block end-of-cycle resistance feature prediction model, it can be achieved through the following steps S41 to S43:

[0089] S41, obtaining historical hydraulic support column pressure data of the historical mining process of the intelligent mining working face.

[0090] Specifically, the historical hydraulic support column pressure data is the hydraulic support column pressure reported by each hydraulic support in the historical mining process.

[0091] S42, extract the cycle-end resistance, time-weighted average working resistance and end maximum resistance increasing rate of all hydraulic supports in each regional block in the last T mining working cycles from the historical hydraulic support column pressure data, and combine these features in time and space order to form a model input feature matrix of size (T, 3n), n is the number of hydraulic supports included in each regional block.

[0092] Specifically, the specific way of extracting the cycle-end resistance, time-weighted average working resistance and end maximum resistance increasing rate of all hydraulic supports in each regional block in the last T mining working cycles is the same as the way of calculating the three features described in S3 above, which will not be repeated here.

[0093] Taking the regional block s in the intelligent mining working face as an example, using the calculation method described in S3 above, the cycle-end resistance, time-weighted average working resistance and end maximum resistance increasing rate of all n hydraulic supports in the regional block s in the last T mining working cycles are extracted, and these features are combined in time and space order to form a model input feature matrix of size (T, 3n) s which can be represented as formula (5):

[0094]

[0095] In formula (5), represents the cycle-end resistance of the i-th hydraulic support in the regional block s in the j-th mining cycle; s represents the time-weighted average working resistance of the i-th hydraulic support in the regional block s in the j-th mining cycle; represents the end maximum resistance increasing rate of the i-th hydraulic support in the regional block s in the j-th mining cycle. s s

[0096] S43, extract the cycle-end resistance feature value of each regional block in the next m mining working cycles from the historical hydraulic support column pressure data as the output of the regional block cycle-end resistance feature prediction model.

[0097] ​​​​​​​​​Specifically, the output of the regional block end-of-cycle resistance feature prediction model is the end-of-cycle resistance feature values of the n hydraulic supports of the regional block in the next m mining work cycles, the specific value of which is the average value of the end-of-cycle resistance of the regional block in space and is weighted and summed by using the exponential decay method for the m mining work cycles in time.

[0098] In one specific embodiment, in the S43, the regional block i The end-of-cycle resistance feature values in the next m mining work cycles are calculated by using formulas (6) to (8):

[0099]

[0100]

[0101]

[0102] In formulas (6) to (8), represents the regional block i The end-of-cycle resistance feature values in the next m mining work cycles, that is, the output of the regional block end-of-cycle resistance feature prediction model; is the regional block i The end-of-cycle resistance in the jth mining work cycle; represents the weighted value of the jth mining work cycle after normalization; is the sum of the weighted values of all work cycles; represents the weighted value corresponding to the jth mining work cycle; is a decay factor used to control the rate of decay, and is taken as ; t is the time step of each mining work cycle. When calculating the weighted value, the embodiment of the present application considers that the time interval of each mining work cycle is the same, so the time steps of the first, second, …, and mth mining work cycles are taken as 0, 1, …, (m-1), respectively. Since the weight directly calculated is not 1, the embodiment of the present application normalizes the calculated weight by using formula (7).

[0103] S44, the regional block end-of-cycle resistance feature prediction model is trained by using the model input feature matrix and the corresponding output.

[0104] Specifically, according to the above model input and output extraction method, the column pressure values of all the hydraulic supports of the overall intelligent mining working face are calculated, and the input and output are aligned to form model training samples. The regional block end-of-cycle resistance feature prediction model is trained by using the model training samples. For the specific way of training the regional block end-of-cycle resistance feature prediction model, refer to the existing neural network model training method, and the embodiment of the present application does not elaborate on this.

[0105] Furthermore, regarding the specific type of the regional block cycle end resistance feature prediction model, the embodiment of the present invention can select a neural network model such as CNN as needed. However, since the input of the regional block cycle end resistance feature prediction model in the embodiment of the present invention is a matrix quantity, and has a strong temporal and spatial order, and the variable weights at different spatiotemporal positions are different, the embodiment of the present invention provides a deep learning network architecture that combines a convolutional neural network (CNN), a bidirectional long short-term memory network (Bi-LSTM) and an attention mechanism (Attention). This architecture integrates the advantages of different neural network layers and can fully explore the relationship between spatial and temporal features, thereby effectively capturing the periodic pressure law for modeling. Preferably, the regional block cycle end resistance feature prediction model is a CNN-BiLSTM-Attention network formed by sequentially connecting CNN, Bi-LSTM and Attention, such as Figure 4 shown.

[0106] The CNN-BiLSTM-Attention network contains 8 layers in total, and the input layer shape is a time series data of (32, T, 3n), where 32 represents that the batch size used for training is 32, T represents the time step (T mining working cycles), and 3n represents that 3n features are contained in each time step. The second layer (convolutional layer 1) uses a one-dimensional convolutional layer (Conv1D) for feature extraction. The convolutional layer uses 128 convolutional kernels with a kernel size of 5 to capture global features, and uses a ReLU activation function and L2 regularization to prevent overfitting. The third layer (convolutional layer 2) further extracts more fine-grained features through a one-dimensional convolutional layer with 64 convolutional kernels and a kernel size of 3. This layer also uses a ReLU activation function and L2 regularization. The fourth layer (BiLSTM layer) uses a bidirectional LSTM (Bidirectional LSTM) to capture long-term dependencies in the time series, and the number of hidden units of the LSTM is 256. The fifth layer (multi-head attention layer) further enhances the attention to key time steps in the sequence through a multi-head attention mechanism (Multi-Head Attention). This layer uses 4 attention heads and a key-value dimension of 32 for each head, which can learn different feature relationships from multiple subspaces. Batch normalization (BatchNormalization) and dropout (0.2) are used between each of the second to fifth layers to further improve the training effect of the model and prevent overfitting of the model. The sixth layer is a global pooling layer. This layer performs average pooling on the output features of each time step to convert the data from a three-dimensional tensor to a two-dimensional tensor, thereby converting the output data into a form suitable for regression tasks. The seventh layer is a fully connected layer. The model maps the features to a hidden layer with a size of 128 through a dense layer (Dense), uses a ReLU activation function, and additionally adds dropout (0.2) to further avoid overfitting. The eighth layer is the output layer, which outputs a value in the range of [0, 1] to represent the predicted regional block resistance feature value at the end of the future m mining working cycles.

[0107] In one specific example, the embodiment of the present application uses field data to perform modeling instantiation analysis. According to the analysis of the space-time characteristics of the working face during the mining process, the model prediction effect is considered, and the minimum regional block is taken as an example, i.e., n=3 and m=3, to construct the model input and output samples. In addition, the T values of this example are selected as 9, 10, 11, 12 and 13 for comparative analysis, and the CNN-BiLSTM-Attention network proposed in the embodiment of the present application is used to model different T values and obtain model evaluation indexes, as shown in Table 1.

[0108]

[0109] Table 1 shows that as the time step T increases, the constructed feature matrix effectively improves the model's fitting ability. In particular, when T = 12, the model achieves optimal performance across all evaluation metrics. However, when T increases to 13, all evaluation metrics show a slight decrease compared to T = 12. This suggests that excessively large T values ​​can lead to over-complexity of the model, causing interference from redundant information, which in turn affects its learning effectiveness and ultimately leads to degradation of model performance. Therefore, the following instantiation process uses T = 12.

[0110] Based on the above analysis, T=12 is selected as the time step to construct the characteristic matrix M of the sample. The field data is the hydraulic support column pressure data of 200 working cycles from May 22 to July 12, 2024. A total of 8118 training samples are constructed, among which a training sample example is selected as follows Figure 5 All training samples were divided into training and test sets in a ratio of 8:2. Different models were used to train the training set and validated with the test set. The final evaluation indicators of each model are shown in Table 2.

[0111]

[0112] Comparing the evaluation metrics of the various models reveals that the LSTM model performs relatively poorly, with large prediction errors and low fit. In contrast, the addition of a bidirectional LSTM (BiLSTM) improves model performance. This demonstrates that the BiLSTM model can better capture bidirectional dependencies in time series, thereby improving the model's fit and prediction accuracy. Further incorporating a convolutional neural network (CNN) as a feature extraction layer, combined with an LSTM (CNN-LSTM) and a BiLSTM (CNN-BiLSTM), significantly improves model performance. The results demonstrate that combining a CNN with either an LSTM or a BiLSTM helps extract richer local features, thereby enhancing the model's expressiveness and generalization capabilities.

[0113] Building on this foundation, the introduction of a multi-head attention mechanism further improved model performance. The CNN-BiLSTM-Attention network achieved top performance across all evaluation metrics, demonstrating its advantage in capturing information at critical moments. The multi-head attention mechanism effectively focuses on key parts of the input sequence, enhancing the model's feature representation capabilities and further improving model fitting and prediction accuracy.

[0114] Therefore, the embodiment of the present application preferably takes the CNN-BiLSTM-Attention network as the basis, and predicts the future cycle-end resistance characteristic value of the regional block by constructing a spatio-temporal zone pressure feature matrix (including cycle-end resistance, time-weighted working resistance, and end maximum resistance increasing rate) based on the history of 12 mining working cycles in the regional block, so as to provide a basis for subsequent variable-scale compression zone autocorrelation aggregation.

[0115] S5, comparing the cycle-end resistance characteristic value of each regional block in the future m mining working cycles with the compression discrimination index value thereof, and screening the regional block whose cycle-end resistance characteristic value in the future m mining working cycles is greater than the compression discrimination index value thereof from all the regional blocks.

[0116] Specifically, the overburden movement of the intelligent mining working face has certain variability in the space influence range during the mining process, and there is close correlation between adjacent regional blocks when the compression appears, and the cycle-end resistance of the adjacent regional blocks can be aggregated by autocorrelation to reveal the compression event of the variable space scale. Specifically, if , it is considered that the regional block i will have a compression event in the future m mining working cycles. By comparison, it can be screened which regional blocks will have a compression event. Therefore, the embodiment of the present application screens the regional block whose cycle-end resistance characteristic value in the future m mining working cycles is greater than the compression discrimination index value thereof from all the regional blocks.

[0117] S6, calculating the local Moran's index of each regional block.

[0118] The local Moran's index is a spatial autocorrelation analysis tool and is widely used to study the spatial aggregation and spatial heterogeneity in geographical data. By calculating the local Moran's index of each regional block, the aggregation region with a higher or lower index value and the dispersion region with high-low or low-high can be identified.

[0119] In one specific embodiment, the S6 realizes the local Moran's index of the regional block i by formula (9) to formula (11):

[0120]

[0121]

[0122]

[0123] In formula (9), is the local Moran's index of the regional block i; is the total number of regional blocks; is the sum of the weights between all regional blocks; is the spatial weight between region blocks i and j;

[0124] In formula (10), is the cycle-end resistance feature of region block i after normalization; is the cycle-end resistance feature value of region block i; i is the mean of cycle-end resistance feature values of all region blocks; is the standard deviation of cycle-end resistance feature values of all region blocks;

[0125] In formula (11), denotes a spatial weight matrix. An element in the matrix is is the spatial weight between region blocks i and j.

[0126] Specifically, before calculating the local Moran index, the cycle-end resistance feature values of the region blocks need to be normalized, and then a corresponding spatial weight matrix is constructed. Since the region blocks in the Ziqie working face present a one-dimensional linear adjacency relationship, the first region block is adjacent to only the next region block, the last region block is adjacent to only the previous region block, and the remaining region blocks are adjacent to both the previous region block and the next region block. Therefore, based on this adjacency relationship, the embodiment of the present application constructs the spatial weight matrix shown in formula (11) W .

[0127] It should be noted that this step S6 can be executed after step S4, and the execution order of S6 is not limited by the embodiment of the present application.

[0128] S7, according to the local Moran index of all region blocks and each screened region block, predicting the gathering area of the coming pressure event.

[0129] Specifically, this step is implemented by the variable-scale coming pressure region autocorrelation aggregation model provided by the embodiment of the present application.

[0130] In one specific embodiment, if the region block i is a screened region block, then the S7 includes:

[0131] S71, if the local Moran index i of the region block is greater than or equal to the local Moran index threshold , it is determined that the coming pressure gathering phenomenon occurs around the region block i .

[0132] Further, if the local Moran index i of the region block is less than the local Moran index threshold , the coming pressure event gathering area is determined to be the region block i .​

[0133] wherein, is set according to experience.

[0134] S72, determining the regional block i according to the size relationship between the local Moran's I of the front and rear regional blocks of the regional block and the local Moran's I threshold value i to determine the regional block around which the aggregation area of the inrush event occurs.

[0135] In one specific embodiment, the S72 comprises:

[0136] i S721, comparing the local Moran's I of the front adjacent regional block of the regional block with if the local Moran's I of the regional block is greater than , then continue to compare the size relationship between the local Moran's I of the front adjacent regional block of the regional block and until a regional block whose local Moran's I is less than i is found as the starting regional block of the inrush event around the regional block .

[0137] It should be noted here that if i is equal to 1, step S721 does not need to be executed, and step S722 is directly executed.

[0138] S722, comparing the local Moran's I of the rear adjacent regional block i of the regional block with if the local Moran's I of the regional block is greater than and , and d is the total number of regional blocks, then the termination regional block of the inrush event around the regional block i is the regional block ; if continue to compare the size relationship between the local Moran's I of the rear adjacent regional block of the regional block and until a regional block whose local Moran's I is less than i is found as the termination regional block of the inrush event around the regional block .

[0139] Figure 7As shown, the method provided by the embodiment of the present application takes n hydraulic support area blocks as a spatial scale and T mining working cycles as a time scale, constructs a two-dimensional space-time area pressure feature matrix, inputs the matrix into an area block cycle-end resistance feature prediction model based on a CNN-BiLSTM-Attention network, outputs cycle-end resistance feature values of area blocks in m future mining working cycles, calculates area block pressure coming discrimination index values according to formulas (3) and (4), compares and judges whether the pressure coming features are met, automatically connects and fuses the pressure coming features of all area blocks of the intelligent mining working face based on a variable-scale pressure coming area autocorrelation aggregation model based on the Moran index algorithm, obtains different scale area pressure coming states in m future mining working cycles, realizes real-time and automatic prediction of variable-scale space-time area pressure coming events, and provides a more accurate and real-time monitoring and early warning method for coping with the pressure coming events of the intelligent mining working face. The cycle-end resistance feature values of all area blocks of the predicted intelligent mining working face are subjected to variable-scale area autocorrelation aggregation, so that possible pressure coming events and their area ranges can be automatically and dynamically predicted.

[0140] The above description is only the preferred embodiments of the present application and is not used to limit the present application. It should be pointed out that, for ordinary skilled persons in the art, several improvements and modifications can be made without departing from the technical principles of the present application, and these improvements and modifications should be considered as the protection scope of the present application.

Claims

1. A method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces, characterized by: include: S1, the regional block for constructing the hydraulic support cluster of the intelligent mining working face; S2, calculate the pressure discrimination index value of each area block; S3, extracting the spatiotemporal regional pressure characteristics of each regional block in T mining work cycles completed by the intelligent mining working face before the current mining work cycle, and constructing a spatiotemporal regional pressure characteristic matrix of each regional block based on the spatiotemporal regional pressure characteristics of each regional block; S4, inputting the spatiotemporal regional pressure characteristic matrix of each regional block into a pre-trained regional block end-of-cycle resistance characteristic prediction model, and using the regional block end-of-cycle resistance characteristic prediction model to predict the end-of-cycle resistance characteristic value of each regional block in the next m mining working cycles; S5, comparing the end-of-cycle resistance characteristic value of each block in the next m mining cycles with its incoming pressure discrimination index value, and selecting from all the blocks the block whose end-of-cycle resistance characteristic value in the next m mining cycles is greater than its incoming pressure discrimination index value; S6, calculate the local Moran index of each regional block; S7, compressing the clustering area of ​​events based on the local Moran index of all regional blocks and the local Moran index prediction of each filtered regional block; If the area block For the filtered area block, S7 includes: S71, if the area block The local Moran index Greater than or equal to the local Moran index threshold , then determine the area block i Pressure accumulation occurs around the area; S72, according to the area block The local Moran index and the local Moran index threshold of the front and back area blocks The size relationship determines the area block The surrounding area is the gathering area of ​​pressure incidents.

2. The method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces according to claim 1 is characterized in that: The S2 block in the calculation area i When the pressure discrimination index value is obtained, it is realized by formula (1): In formula (1), For regional blocks i The pressure discrimination index value; n For regional blocks i The number of hydraulic supports included; The total number of mining cycles currently completed in the intelligent mining working face; is the variance coefficient, ranging from 0.8 to 1; For regional blocks i The resistance at the end of the j-th mining cycle; In formula (2), For regional blocks i Neidi The end-of-cycle resistance of a hydraulic support in the jth mining working cycle.

3. The method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces according to claim 1 is characterized in that: The spatiotemporal regional pressure characteristics of each block include the end-of-cycle resistance, time-weighted average working resistance and terminal maximum resistance increase rate of all hydraulic supports in each block in each mining working cycle.

4. The method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces according to claim 3 is characterized in that: The time-weighted average working resistance of a hydraulic support in any area block during a mining working cycle is calculated by formula (3): In formula (3), is the time-weighted average work resistance; The number of sections of hydraulic support column pressure within a mining working cycle; For the a The pressure of the hydraulic support column within a time period; For the a The time interval corresponding to the time period; The maximum resistance increase rate of a hydraulic support in any area block at the end of a mining working cycle is calculated by formula (4): In formula (4), is the maximum resistance increase rate at the end; It is the maximum increment of pressure of adjacent segmented hydraulic support columns within the last 5 minutes of a mining working cycle; is the corresponding time period.

5. The method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces according to claim 1, 3 or 4, characterized in that: Before S4, it also includes: S41, obtaining historical hydraulic support column pressure data of the historical mining process of the intelligent mining working face; S42, extracting the end-of-cycle resistance, time-weighted average working resistance, and terminal maximum resistance increase rate of all hydraulic supports in each block during the first T mining cycles from the historical hydraulic support column pressure data, and combining these features in a temporal and spatial order to form a model input feature matrix of size (T, 3n), where n is the number of hydraulic supports included in each block; S43, extracting the end-of-cycle resistance characteristic value of each block in the next m mining cycles from the historical hydraulic support column pressure data as the output of the end-of-cycle resistance characteristic prediction model of the block; S44, a regional block end-of-cycle resistance feature prediction model is trained through the model input feature matrix and its corresponding output.

6. The method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces according to claim 5 is characterized in that: In the above S43, the area block i The end-of-cycle resistance characteristic values ​​of the next m mining cycles are calculated using formulas (6) to (8): In formula (6) to formula (8), Indicates area block The end-of-cycle resistance characteristic value of the next m mining cycles, i.e., the output of the regional block end-of-cycle resistance characteristic prediction model; For regional blocks i The resistance at the end of the j-th mining cycle; represents the normalized weighted value of the jth mining cycle; is the sum of the weighted values ​​of all working cycles; represents the weighted value corresponding to the jth mining working cycle; is the attenuation factor, which is used to control the attenuation rate. The time step for each mining work cycle.

7. The method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces according to claim 1 is characterized in that: The regional block cycle end resistance feature prediction model is a CNN-BiLSTM-Attention network formed by sequentially connecting CNN, Bi-LSTM and Attention.

8. The method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces according to claim 1 is characterized in that: The S6 is in the calculation area block i When the local Moran index is calculated, it is realized by formula (9) to formula (11): In formula (9), is the local Moran index of region block i; is the total number of regional blocks; is the sum of the weights between all regional blocks; For regional blocks and The spatial weight between In formula (10), is the normalized end-of-cycle resistance characteristic of block i; For regional blocks i The characteristic value of the end-of-cycle resistance; is the mean of the end-of-cycle resistance characteristic values ​​of all regional blocks; is the standard deviation of the end-of-cycle resistance characteristic values ​​of all regional blocks; In formula (11), W represents the spatial weight matrix.

9. The method for dynamic prediction of variable-scale spatiotemporal regional pressure events in intelligent mining working faces according to claim 1 is characterized in that: The S72 includes: S721, block the area i The previous adjacent area block The local Moran index and Compare, if the area block The local Moran index and , then continue to compare the region blocks The local Moran index of the previous adjacent area block is until the local Moran index is less than The regional block, as the regional block i The starting area block of the surrounding pressure event; S722, block the area i The adjacent area block The local Moran index and Compare, if the area block The local Moran index and , d is the total number of regional blocks, then determine the regional blocks i The termination area block of the surrounding pressure event is the area block ;if Continue comparing region blocks The local Moran index of the adjacent area block is until the local Moran index is less than The regional block, as the regional block i The termination region block surrounding incoming pressure events.

Images