Training method of coal mine pressure prediction model and coal mine pressure prediction method

Abstract

The application provides a coal mine pressure prediction model training method and a coal mine pressure prediction method. The coal mine pressure prediction model training method comprises the following steps: obtaining sample coal mine detection data in a target time period in a coal mine scene database; inputting the sample coal mine detection data into a coal mine pressure prediction model, and generating a dynamic adjacency matrix between sensors according to the sample coal mine detection data by a processing layer; inputting the dynamic adjacency matrix and the sample coal mine detection data into a convolution layer for convolution processing to obtain spatial feature information; performing coding and time feature extraction on the sample coal mine detection data by an encoding layer to obtain time feature information; fusing the spatial feature information and the time feature information to obtain fused features, and obtaining a mine pressure prediction result according to the fused features, so as to train the coal mine pressure prediction model, obtain a pre-trained coal mine pressure prediction model, and improve the prediction accuracy of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent prediction of coal mine pressure, and in particular to a training method of a coal mine pressure prediction model and a coal mine pressure prediction method. BACKGROUND

[0002] As a key technology for coal mine safety production, the existing technology mainly faces the challenges of insufficient dynamic spatial correlation modeling and low spatio-temporal feature fusion efficiency. The traditional time series model is limited by the linear modeling framework and is difficult to capture the strong nonlinearity and spatio-temporal coupling characteristics of mine pressure data. Although the early deep learning model enhances the time series modeling ability through the cycle structure, it ignores the dynamic correlation of the spatial distribution of sensors and often treats multi-source data as independent sequences, resulting in distortion of spatial feature expression. The spatio-temporal convolution network and static graph convolution model developed in recent years attempt to fuse spatio-temporal features, but rely on fixed spatial connections and cannot adapt to the dynamic evolution of sensor correlation caused by changes in geological conditions during mining. In addition, the fixed position encoding of the standard Transformer model significantly reduces the long-time sequence dependence modeling ability when processing non-equidistant mine pressure data and time-varying working conditions. Moreover, the existing cross-attention mechanism is mostly one-way interaction, lacks reverse modeling of "spatial structure affecting time sequence changes", and does not dynamically adjust the fusion strategy combined with mining working conditions, resulting in insufficient modeling accuracy of complex spatio-temporal coupling relationships. SUMMARY

[0003] The present application aims to at least partially solve one of the technical problems in the related art.

[0004] To this end, the first object of the present application is to propose a training method of a coal mine pressure prediction model to realize accurate construction of the coal mine pressure prediction model and improve the prediction accuracy of the mine pressure prediction model.

[0005] The second object of the present application is to propose a coal mine pressure prediction method.

[0006] The third object of the present application is to propose a training device of a coal mine pressure prediction model.

[0007] The fourth object of the present application is to propose a coal mine pressure prediction device.

[0008] The fifth object of the present application is to propose an electronic device.

[0009] The sixth object of the present application is to propose a computer-readable storage medium.

[0010] The seventh object of the present application is to propose a computer program product.

[0011] To achieve the above object, the first aspect of the present application proposes a coal mine pressure prediction model training method, comprising:

[0012] Obtain sample coal mine detection data in a target time period in a coal mine scene database;

[0013] Input the sample coal mine detection data into the coal mine pressure prediction model, and generate a dynamic adjacency matrix between sensors according to the sample coal mine detection data by the processing layer of the coal mine pressure prediction model;

[0014] Input the dynamic adjacency matrix and the sample coal mine detection data into the convolution layer for convolution processing to obtain spatial feature information;

[0015] Encode and extract time features of the sample coal mine detection data through the encoding layer to obtain time feature information;

[0016] Input the spatial feature information and the time feature information into the fusion layer for fusion processing to obtain fusion features, and obtain a mine pressure prediction result according to the fusion features;

[0017] Train the coal mine pressure prediction model according to the mine pressure prediction result and the actual result to obtain a pre-trained coal mine pressure prediction model.

[0018] To achieve the above object, the second aspect of the present application proposes a coal mine pressure prediction method, comprising:

[0019] Collect coal mine detection data in a coal mine environment to be analyzed;

[0020] Input the coal mine detection data into the pre-trained coal mine pressure prediction model to output a mine pressure prediction result;

[0021] The coal mine pressure prediction model is a model obtained by the coal mine pressure prediction model training method of the first aspect.

[0022] To achieve the above object, the third aspect of the present application proposes a coal mine pressure prediction model training device, comprising:

[0023] An acquisition module is configured to obtain sample coal mine detection data in a target time period in a coal mine scene database;

[0024] A processing module is configured to input the sample coal mine detection data into the coal mine pressure prediction model, and generate a dynamic adjacency matrix between sensors according to the sample coal mine detection data by the processing layer of the coal mine pressure prediction model;

[0025] A convolution module is configured to input the dynamic adjacency matrix and the sample coal mine detection data into the convolution layer for convolution processing to obtain spatial feature information;

[0026] The encoding module is configured to encode and extract time features of the sample coal mine detection data through an encoding layer to obtain time feature information.

[0027] The fusion module is configured to input the spatial feature information and the time feature information into a fusion layer for fusion processing to obtain fusion features, and obtain a mine pressure prediction result according to the fusion features.

[0028] The training module is configured to train the coal mine pressure prediction model according to the mine pressure prediction result and an actual result to obtain a pre-trained coal mine pressure prediction model.

[0029] To achieve the above object, the fourth aspect of the present application proposes a coal mine pressure prediction device, comprising:

[0030] The acquisition module is configured to collect coal mine detection data in a coal mine to be analyzed.

[0031] The prediction module is configured to input the coal mine detection data into the pre-trained coal mine pressure prediction model to output a mine pressure prediction result. The coal mine pressure prediction model is a model obtained by using the training method of the coal mine pressure prediction model according to the first aspect of the present application.

[0032] To achieve the above object, the fifth aspect of the present application proposes an electronic device, comprising a processor and a memory in communication connection with the processor.

[0033] The memory stores computer execution instructions.

[0034] The processor executes the computer execution instructions stored in the memory to implement the method according to the first aspect of the present application or the second aspect of the present application.

[0035] To achieve the above object, the sixth aspect of the present application proposes a computer readable storage medium, wherein the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by a processor to implement the method according to the first aspect of the present application or the second aspect of the present application.

[0036] To achieve the above object, the seventh aspect of the present application proposes a computer program product, comprising a computer program, which is executed by a processor to implement the method according to the first aspect of the present application or the second aspect of the present application.

[0037] The training method of the coal mine pressure prediction model and the coal mine pressure prediction method provided by the application, by obtaining sample coal mine detection data, and obtaining the correlation between sensors according to the sample coal mine detection data, a dynamic adjacency matrix is constructed, the spatial feature information is obtained based on the dynamic adjacency matrix, and further based on the sample coal mine detection data, coding and time feature extraction are performed to obtain time feature information, the time feature information and the time feature information are fused, the interaction of the time feature and the space feature is more sufficient, the fusion feature is used to obtain the prediction result, and the complex space-time coupling relationship in the mine pressure data is better captured, and the coal mine pressure prediction model with more accurate prediction result is trained.

[0038] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0039] The above and / or additional aspects and advantages of the application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0040] Figure 1 A schematic diagram of a coal mine pressure prediction model training method provided by an embodiment of the application;

[0041] Figure 1A A logic diagram of model training provided by an embodiment of the application;

[0042] Figure 2 A flowchart of a method for generating a dynamic adjacency matrix provided by an embodiment of the application;

[0043] Figure 2A A logic flowchart of obtaining a dynamic adjacency matrix provided by an embodiment of the application;

[0044] Figure 3 A flowchart of a method for obtaining spatial feature information provided by an embodiment of the application;

[0045] Figure 4 A flowchart of a method for obtaining time feature information provided by an embodiment of the application;

[0046] Figure 5 A schematic diagram of a method for obtaining fusion features provided by an embodiment of the application;

[0047] Figure 5A A logic flowchart of obtaining fusion features provided by an embodiment of the application;

[0048] Figure 6 A logic flowchart of training of a coal mine pressure prediction model provided by an embodiment of the application;

[0049] Figure 7 A schematic diagram of a coal mine pressure prediction method provided by an embodiment of the present application;

[0050] Figure 8 A structural schematic diagram of a training device of a coal mine pressure prediction model provided by an embodiment of the present application;

[0051] Figure 9 A structural schematic diagram of a coal mine pressure prediction device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0052] The embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, in which the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0053] The training method of a coal mine pressure prediction model and the coal mine pressure prediction method of an embodiment of the present application are described below with reference to the accompanying drawings.

[0054] Figure 1 A schematic diagram of a training method of a coal mine pressure prediction model provided by an embodiment of the present application. As shown in the figure, the method comprises the following steps: Figure 1

[0055] S101, obtaining sample coal mine detection data in a target time period in a coal mine scene database.

[0056] In some embodiments, the sample coal mine detection data can be three-dimensional data, including time dimension, space dimension and feature dimension. Specifically, the time dimension is different sampling time, the space dimension is a sensor deployed at different positions, and the feature dimension is different feature data.

[0057] In some embodiments, the time dimension can determine the number of specific sampling points. For example, in the present embodiment, the total number of sampling time steps (sampling points) can be T, ; wherein each sampling point is represented as t; the time interval between each two sampling points is , , is the sampling frequency, , and the corresponding time interval is .

[0058] In some implementations, the space dimension is a sensor arranged at different positions, and the space dimension N is the number of sensors. In the present embodiment, each sensor node ​, which are deployed at key positions in the coal mine roadway, such as working face support and fault monitoring point.

[0059] In some embodiments, the feature dimension is different data features collected by the sensor. In the present embodiment, the number of features collected by each sensor is For example, the feature dimension can be divided into mechanical features and environmental features. The mechanical features can include stress, displacement, acceleration and the like, and the environmental features can include temperature, humidity and the like.

[0060] Optionally, the target time period can be a preset time range required for actual scene prediction, for example, one month as the target time period, and the detection data of different dimensions of each sensor at different sampling time points in the target time period is obtained to obtain all sample coal mine detection data.

[0061] In some embodiments, the range of sampling time points in the target time period in the present embodiment is , wherein L is the length of the historical time window H is the prediction time step, T is the total sampling time step; for example, if the length of the historical time window L is 24 and the total sampling time step T is 100, that is, the number of continuous sampling time points in the historical database to be extracted is set to 24 and the total sampling time point is 100, t is the index of each sampling time point in the 100 sampling time points, the starting target time period is , that is, 24 continuous sampling time points, when the sampling time point t is 36, the target time period is 24 continuous sampling time points, when the sampling time point t is the maximum 100-6=94, the target time period is 24 continuous sampling time points.

[0062] It can be understood that the actual target time period can be determined according to the sampling frequency and the range of sampling time points, for example, the sampling frequency is 1 time / 10 minutes, L=24 sampling time points, that is, 24*10 minutes=4 hours, the prediction sampling time point H=6, that is, 6*10 minutes=1 hour, and the total time T=100, that is, 100*10 minutes 16.7 hours.

[0063] S102, input the sample coal mine detection data into the coal mine pressure prediction model, and generate a dynamic adjacency matrix between the sensors according to the sample coal mine detection data by the processing layer of the coal mine pressure prediction model.

[0064] Optionally, the information vector of each sensor can be obtained according to the sample coal mine detection data collected by each sensor, and each sensor can be taken as a node, and the mutual information between each two nodes can be calculated according to the feature information vectors of the two nodes.

[0065] Further, the mutual information between the nodes can be taken as an element in a dynamic adjacency matrix, and the rows and columns of the dynamic adjacency matrix are arranged in the same order of the sensors, that is, the row sensor and the column sensor corresponding to the diagonal element of the dynamic adjacency matrix are the same sensor, and the mutual information between each two sensors can be taken as an element in the dynamic adjacency matrix, and the diagonal element is 0, so that the dynamic adjacency matrix between the sensors can be obtained.

[0066] In S103, the dynamic adjacency matrix and the sample coal mine detection data are input into a convolution layer for convolution processing, and spatial feature information is obtained.

[0067] Optionally, for each sampling time, a feature matrix at the sampling time can be obtained, and the dimension of the feature matrix is N*M, where N is the number of sensors, and M is the number of features collected by each sensor, that is, the feature matrix is constructed by different sensors and the feature data collected by the sensors.

[0068] In some embodiments, a weight factor can be obtained based on the dynamic adjacency matrix, and a convolution processing can be performed based on the weight factor and the feature matrix at different sampling times, and an output feature at the corresponding sampling time can be obtained, for example, a weighted result can be obtained by multiplying the weight factor and the feature matrix, a summation result can be obtained by adding the weighted result and the feature matrix, and a linear activation can be performed to obtain the output feature.

[0069] In some embodiments, the convolution layer can further include multiple layers of convolution, and the weight factor of each layer of convolution is the dynamic adjacency matrix, the input of the initial convolution is the feature matrix, the input of the second layer of convolution is the output of the first layer of convolution, and so on, until the output matrix of the last layer of convolution is output.

[0070] Further, the output matrices at all sampling times can be spliced, and the spliced matrix can be taken as the spatial feature information.

[0071] In S104, the sample coal mine detection data is encoded and time feature extraction is performed by an encoding layer, and time feature information is obtained.

[0072] Optionally, position encoding can be performed based on the actual working condition parameters, and the encoding data of each position can be obtained, and the encoding data of each position can be embedded into the sample coal mine detection data of the position to obtain encoded detection data.

[0073] Further, time feature extraction is performed on the encoding detection data, in this embodiment, multi-head self-attention heads are used for attention processing, and the output features of each attention head are spliced and fused to determine the final time feature information.

[0074] In S105, the spatial feature information and the time feature information are input into a fusion layer for fusion processing to obtain fusion features, and a mine pressure prediction result is obtained according to the fusion features.

[0075] In some embodiments, a fusion weight factor can also be obtained according to the fusion layer, and the spatial feature information and the time feature information are weighted and fused based on the fusion weight factor to obtain the fusion features.

[0076] Optionally, inference prediction can be performed according to the fusion features to obtain the mine pressure prediction result.

[0077] In S106, a coal mine pressure prediction model is trained according to the mine pressure prediction result and an actual result to obtain a pre-trained coal mine pressure prediction model.

[0078] In some embodiments, the fusion features, the working condition parameter vector and the historical prediction error can be input into an inference layer for inference prediction to obtain the mine pressure prediction result.

[0079] In this embodiment, the state space is: with a dimension of After integrating the three information, the final output is an input vector of the model inference layer, wherein is the historical prediction error, quantifying the prediction accuracy at the previous time, , is the predicted value, is the true value; is the working condition parameter vector, providing a decision basis for the inference layer, improving the accuracy, early warning performance and inference efficiency of the mine pressure prediction.

[0080] Action space: H is the sampling time of the prediction, used to adjust the time range of the model prediction to adapt to the early warning needs under different mining speeds; is the learning rate of the model training, controlling the parameter update amplitude, and increasing to accelerate the model adaptation when the working condition suddenly changes (such as fault exposure); is the number of attention heads, used to adjust the multi-scale capture ability of the time sequence features.

[0081] Reward function: ; wherein RMSE For the current prediction root mean square error, EarlyWarning(t) is the effective warning time (hours) before the disaster occurs, and 0 is taken when there is no warning; InferenceTime(t) is the time consumed by a single inference, which is dynamically adjusted and optimized through model parameters.

[0082] Further, network training is performed, that is, the intelligent learning body is trained; in the embodiment, the model can be a deep Q network (DQN), and the training process includes: experience replay: storing experience tuples into an experience pool with a capacity of , and training is performed by using uniform random sampling.

[0083] Target network update: , the main network parameters are synchronized every 500 steps.

[0084] Loss function: , , wherein is a discount factor, the main network parameters are updated by an Adam optimizer, the learning rate is linearly decayed with the training round, and the model is approximated to the optimal policy by minimizing the loss , so that the trained coal mine pressure prediction model is obtained, and the prediction performance of the coal mine pressure prediction model is improved; for example, the training process of the model in the embodiment is shown in Figure 1A .

[0085] In the embodiment, sample coal mine detection data is obtained, and the correlation between sensors is obtained according to the sample coal mine detection data to construct a dynamic adjacency matrix. The spatial feature information is obtained based on the dynamic adjacency matrix, and the encoding and time feature extraction are further performed based on the sample coal mine detection data to obtain the time feature information. The time feature information and the time feature information are fused, so that the interaction between the time feature and the space feature is more sufficient. The fusion feature after fusion is used to obtain the prediction result, the features are fully mined from the two dimensions of space and time, the long-term change trend of the mine pressure is more accurately captured, the accuracy of the result prediction is effectively improved, the complex space-time coupling relationship in the mine pressure data is better captured, and the coal mine pressure prediction model with more accurate prediction result is trained.

[0086] On the basis of the above embodiment, Figure 2 is a flowchart of a method for generating a dynamic adjacency matrix provided by the embodiment of the application. As shown in Figure 2 , the method comprises:

[0087] S201, according to sample coal mine detection data, determining the joint probability density between each two sensors and the marginal probability density of each sensor.

[0088] In some embodiments, for the sensor and , the time series of the sensor and in the target time period can be represented as and , the joint probability density between the sensor and is obtained based on and is:

[0089]

[0090] wherein, is the joint probability density between the sensor and ; L is the length of the historical time window, in the present embodiment ; h is the kernel bandwidth, which is adaptively adjusted according to the feature standard deviation to ; is the corresponding time series of the sensor ; is the corresponding time series of the sensor ; and are the kth data in the corresponding time series of the sensor and the sensor , respectively.

[0091] It can be understood that, for the feature standard deviation , the global average value of each feature dimension m is obtained by calculation, and specifically, the acquisition of the global average value and the standard deviation can be respectively represented as:

[0092]

[0093]

[0094] wherein, is the global average value of the feature dimension m; is the standard deviation of the feature dimension m; is the feature m data collected by the sensor n at the sampling moment t; T is the total sampling moment, and N is the number of sensors.

[0095] In some implementations, for the calculation of the marginal probability density of the sensor and the marginal probability density of the sensor , they are respectively:

[0096]

[0097]

[0098] wherein, is the marginal probability density of the sensor ; is the marginal probability density of the sensor .

[0099] In some embodiments, before the calculation based on the sample coal mine detection data, data validity detection can also be performed, such as judging whether there is abnormal value or null value and other abnormal data in the sample coal mine detection data, and invalid data is excluded to improve the subsequent calculation accuracy.

[0100] S202, based on the joint probability density and the marginal probability density, determining mutual information between the two sensors corresponding.

[0101] Optionally, the mutual information between the sensor and may be expressed as:

[0102]

[0103] wherein, represents the mutual information between the sensor and , which can be calculated by numerical integration (such as Gaussian quadrature) in the embodiment.

[0104] S203, determining the edge weight between the two sensors corresponding by the mutual information, and obtaining the dynamic adjacency matrix based on the edge weight corresponding between all sensors as the matrix element.

[0105] Optionally, the mutual information can be normalized as the edge weight between the two sensors corresponding ： ; wherein, is the edge weight between the sensor and ; , which is used to avoid the denominator being zero.

[0106] Further, based on the edge weight corresponding between all sensors as the matrix element, the dynamic adjacency matrix is obtained, and the row and column of the dynamic adjacency matrix are the same order of the sensors, that is, the diagonal element is the mutual information of the same sensor itself, and in the embodiment , that is, the diagonal elements in the dynamic adjacency matrix are all 0, wherein the elements in the 2 rows and 3 columns in the dynamic adjacency matrix correspond to the edge weight between the second sensor and the third sensor.

[0107] In some embodiments, the sensor set can also be monitored in real time to determine whether there are changes in the topological relationships between sensors, such as adding or deleting some sensors, so as to dynamically update the edge weights between sensors, thereby responding in real time to determine a more timely dynamic adjacency matrix and improving the accuracy of spatial feature extraction.

[0108] like Figure 2A As shown, after acquiring sample coal mine detection data through a set of sensors, data validity is checked. When the data is valid, mutual information is calculated, and edge weights are obtained by normalization based on the mutual information. Topological changes are detected in real time. When there are topological changes, the mutual information between sensors is updated, thereby updating the edge weights to obtain a new dynamic adjacency matrix. If there are no topological changes, the current edge weights are maintained to determine the dynamic adjacency matrix for subsequent analysis.

[0109] In this embodiment, the joint probability density between sensors and the edge probability density of individual sensors are obtained from the sample coal mine detection data collected by the original sensors. The mutual information between sensors is determined by the joint probability density and the edge probability density to fully reflect the correlation between sensors. The edge weights are determined by normalization based on the mutual information between sensors. A dynamic adjacency matrix is ​​constructed based on the edge weights and sensors. This dynamic adjacency matrix includes the edge weights between different sensors, which can accurately capture the dynamic spatial relationship between sensors. The spatial feature information is extracted by combining the dynamic adjacency matrix to ensure the comprehensiveness of spatial feature mining.

[0110] Based on the above embodiments, the convolutional layer includes at least one sub-convolutional layer. Figure 3 This is a flowchart illustrating a method for obtaining spatial feature information provided in an embodiment of this application. Figure 3 As shown, the method includes:

[0111] S301, for any sub-convolutional layer in the convolutional layer, determine the input data.

[0112] In some embodiments, if any sub-convolutional layer is the first sub-convolutional layer, the input data is the target sample data after preprocessing of the sample coal mine detection data.

[0113] Optionally, preprocessing the sample coal mine detection data can involve standardizing it. In this embodiment, each feature data is standardized based on its global mean and standard deviation to obtain the preprocessed target sample data for each sensor. The standardization process can be specifically represented as follows:

[0114]

[0115] in, This involves standardizing the feature data m collected by sensor n at sampling time t after the standardization transformation, which is the target sample data; ensuring that the standardized target sample data meets the expected value. And variance To improve the stability of model training.

[0116] It is understandable that the target sample data used as output data refers to the different feature data collected by different sensors n at the same sampling time t. That is, the dimension of the target sample data is N*M, where N is the number of sensors and M is the number of features collected by the sensors.

[0117] In some embodiments, if any sub-convolutional layer is a layer other than the first layer, the input data is the output data of the previous sub-convolutional layer.

[0118] S302, based on the addition of the preset matrix and the dynamic adjacency matrix, the augmented adjacency matrix is ​​obtained.

[0119] Optionally, the preset matrix can be an identity matrix.

[0120] Alternatively, the augmented adjacency matrix can be represented as: ,in For augmented adjacency matrix; It is a dynamic adjacency matrix; It is an identity matrix.

[0121] S303 calculates the degree matrix of the augmented adjacency matrix and performs symmetric normalization on the degree matrix to obtain the target matrix.

[0122] Alternatively, the degree matrix calculation can be expressed as: .

[0123] Furthermore, by performing symmetric normalization on the degree matrix, the resulting target matrix can be expressed as: .

[0124] S304 determines the output data of the current sub-convolutional layer based on the target matrix and input data.

[0125] Alternatively, the output data of the sub-convolutional layer can be represented as:

[0126]

[0127] in, Indicates the first Output data of the sub-convolutional layer; For the first The output data of the sub-convolutional layer has a dimension of ; The weight matrix is ​​a learnable matrix; in this embodiment, the dimension of the weight matrix is ​​[dimension value missing]. .

[0128] It can be understood that, if is 1, then , is the target sample data.

[0129] For example, if three sub-convolutional layers are included in the embodiment, for the first sub-convolutional layer, the augmented adjacency matrix is obtained, the degree matrix is obtained according to the augmented adjacency matrix, and the target matrix is obtained by performing symmetric normalization on the degree matrix, and based on the target matrix, the target sample data, and the first layer learnable weight matrix, the output data of the first sub-convolutional layer is determined , and the dimension of the first layer learnable weight matrix is , , the first layer output channel number is in the embodiment; the dimension of the output data is , which can be 64.

[0130] For the second sub-convolutional layer, the target matrix is obtained, and based on the target matrix, and the second layer learnable weight matrix, the output data of the second sub-convolutional layer is determined , wherein the dimension of the second layer learnable weight matrix is , , the second layer output channel number is , the dimension of the output data is , which can be 128 in the embodiment.

[0131] For the third sub-convolutional layer, the target matrix is obtained, and based on the target matrix, and the third layer learnable weight matrix, the output data of the third sub-convolutional layer is determined , wherein the dimension of the third layer learnable weight matrix is , and the dimension of the output data is .

[0132] S305, based on the output data of the last sub-convolutional layer, the spatial feature information is determined.

[0133] In some embodiments, the output data of the last sub-convolutional layer at all sampling times can be stacked to obtain the spatial feature information, and all sampling times are L in the embodiment. Therefore, the output data of the last sub-convolutional layer at L sampling times is spliced, and assuming that the last sub-convolutional layer is the third layer, the spatial feature information formed by stacking the three-dimensional vectors in time sequence can be expressed as:

[0134]

[0135] wherein, is the spatial feature information, and the dimension is .

[0136] In some embodiments, the convolutional layer can be an adaptive graph convolutional neural network (AGCN), that is, a plurality of sub-convolutional layers are directly executed by the AGCN to obtain the final spatial feature information.

[0137] In this embodiment, the target sample data is processed by multi-layer convolution and a dynamic adjacency matrix. In each layer of convolution processing, the target matrix is obtained by the dynamic adjacency matrix, and the output data of each sub-convolutional layer is obtained by calculation based on the target matrix and the input data. After multi-layer convolution of all sub-convolutional layers, the output data at each sampling time is obtained. The output data at all sampling times is stacked to obtain the spatial feature information. Based on the dynamic adjacency matrix, the real dependency relationship between sensors at different sampling times is responded more timely, the spatial characteristics of the data are fully mined for analysis, and the analysis accuracy of the subsequent coal mine pressure prediction model is improved.

[0138] On the basis of the above-described embodiments, Figure 4 is a flowchart of a method for obtaining time feature information provided by the embodiments of the present application. As Figure 4 shown, the method comprises:

[0139] S401, obtaining a working condition parameter vector and a mining speed at a current sampling time.

[0140] In some embodiments, the working condition parameters at least include geological parameters, mining process parameters and environmental parameters. The geological parameters can include mining depth, rock hardness coefficient, fault distance (such as the distance from the working face ), process parameters can include support strength and daily advance rate, and environmental parameters can include roof pressure grade and underground water content.

[0141] In some embodiments, all working condition parameters can also be normalized to ensure that the value range of each element is [0, 1], and the working condition parameter vector is obtained to avoid the influence of dimensional differences on calculation.

[0142] S402, position encoding based on the working condition parameter vector and the mining speed to obtain an encoding result.

[0143] Optionally, the process of position encoding based on the working condition parameter vector and the mining speed to obtain an encoding result adopts the following formula:

[0144]

[0145] wherein, is the encoding result; is the sample point index; is the feature dimension index; is the total feature dimension; is the mining speed; is the working condition parameter vector; is the weight coefficient.

[0146] In the embodiment , may be 128; The core objective is to integrate the actual working conditions (such as mining speed, rock hardness, etc.) in the coal mining process into the time sequence feature extraction process, thereby enhancing the model's ability to capture long-time sequence dependence and time-varying characteristics, jointly encoding “time position” and “working condition state”, enabling the model to perceive real-time working conditions in the mining process, and improving the representation accuracy of mine pressure time sequence features.

[0147] S403, the encoding result and the target sample data after preprocessing of the sample coal mine detection data are subjected to multi-head self-attention processing to obtain time feature information.

[0148] Optionally, the target sample data is combined with the encoding data as an initial input, which can be represented as: with a corresponding dimension of , subjected to attention processing by attention heads to obtain time feature information.

[0149] Optionally, the multi-head self-attention processing can be represented as:

[0150]

[0151] wherein, the calculation of a single attention head is:

[0152]

[0153] wherein, , that is, the total number of channels is evenly distributed to h attention heads, ensuring the dimensional consistency of multi-head parallel calculation; , and are obtained by linear transformation , , , wherein is a global query matrix, a global key matrix, and a global value matrix, all with a dimension of ; is a learnable weight matrix, is the first input of layer attention; for initial input dimension of dimension adjustment to , for example , , each head dimension is 8, ensuring the subspace independence of multi-head parallel calculation.

[0154] It can be understood that for the initial input , the position encoding of each sampling time pos is no longer a fixed value, but a coding result dynamically calculated according to the working condition parameter vector and the mining speed, that is, the coding result is embedded into the features of each sampling time, so that the model can perceive the current working condition when processing each sampling time, and then extract the time sequence features containing the working condition information.

[0155] Further, based on multi-head self-attention processing, time feature information is obtained, with a dimension of , and is consistent with the dimension of the spatial feature information.

[0156] In some embodiments, the encoding layer can be implemented based on a Transformer encoder to more effectively capture time sequence dependent features.

[0157] In this embodiment, in the time feature extraction, working condition parameters such as mining speed and rock hardness are introduced, and dynamic position encoding is performed based on the working condition parameter vector to form joint encoding of time position + working condition state. When the mining speed increases, the model can automatically enhance the attention to short-term stress fluctuation features through dynamic position encoding, and when hard rock layers are exposed, the modeling of long-term stress accumulation features will be strengthened, so that each position no longer uses the original fixed encoding, and the encoding data is embedded into the target sample data for multi-head self-attention processing, which can effectively capture long-range time sequence dependent features containing working condition information, and finally output time feature information to more accurately predict the change trend of mine pressure.

[0158] On the basis of the above embodiments, Figure 5 is a schematic diagram of a method for obtaining fused features provided by an embodiment of the present application. As Figure 5 shown, the method comprises:

[0159] ​​​​​S501, respectively, the spatial feature information and the time feature information are subjected to forward cross-attention processing and reverse cross-attention processing to obtain first feature information and second feature information.

[0160] Optionally, a first vector set can be determined according to the spatial feature information, the time feature information and the first parameter matrix set; a second vector set can be determined according to the spatial feature information, the time feature information and the second parameter matrix set; a first attention matrix and a second attention matrix can be respectively determined according to the first vector set and the second vector set; the first feature information can be determined according to the first attention matrix and a first Value vector; and the second feature information can be determined according to the second attention matrix and a second Value vector.

[0161] Optionally, the spatial feature information can be unfolded according to a sampling time as , and the dimension is , and the multi-head attention modeling node association is ; correspondingly, the time sequence feature information can be unfolded according to a sampling time as , and the dimension is , and the time dependence is captured .

[0162] In some embodiments, the first parameter matrix set includes a first learnable Query weight matrix , a first Key weight matrix and a first Value weight matrix , the first vector set includes a first Query vector , a first Key vector and a first Value vector ; and the first vector set is obtained as follows: , is the time feature information, is the spatial feature information.

[0163] Correspondingly, the second parameter matrix set includes a second learnable Query weight matrix , a second Key weight matrix and a second Value weight matrix , the second vector set includes a second Query vector , a second Key vector and a second Value vector ; and the second vector set is obtained as follows: .

[0164] Further, the first attention matrix and the second attention matrix are respectively represented as:

[0165]

[0166]

[0167] wherein, is the first attention matrix; is the second attention matrix.

[0168] The first feature information and the second feature information are respectively represented as:

[0169]

[0170]

[0171] wherein, is the first feature information; is the second feature information.

[0172] S502, the first feature information and the second feature information are weighted and fused to obtain a fusion feature.

[0173] In some embodiments, a working condition parameter vector of a current sampling time can be obtained, and the working condition parameter at least includes: geological parameters, mining process parameters and environmental parameters; based on a full connection layer, a fusion weight is generated according to the working condition parameter vector.

[0174] Optionally, the fusion weight can be represented as:

[0175]

[0176]

[0177] wherein, is the fusion weight; is the working condition parameter vector; the dimension of is , the dimension of is , are all learnable parameter matrices; is a bias.

[0178] The embodiment obtains the fusion weight through two full connection layers, wherein and are parameters of the first full connection layer, and are parameters of the second full connection layer, which are obtained through a back propagation algorithm (such as Adam optimizer) and trained cooperatively with other parameters of the model.

[0179] Further, the first feature information and the second feature information are weighted and fused based on the fusion weight to obtain a fusion feature, and the fusion feature is represented as:

[0180]

[0181] wherein, is a fusion feature; represents an element-wise weighted fusion, thereby realizing dynamic coupling of time-space features.

[0182] As Figure 5A shown, the spatial features and the temporal features are combined to perform forward cross-attention and reverse cross-attention respectively, in which the forward cross-attention is time-guided spatial enhancement, and the reverse cross-attention is spatially constrained temporal enhancement feature, thereby obtaining first feature information and second feature information, obtaining a fusion weight, and performing weighted fusion according to the fusion weight to obtain a final spatio-temporal fused fusion feature.

[0183] In this embodiment, the spatial feature information and the temporal feature information are preliminarily fused based on the forward cross-attention and the reverse cross-attention to obtain the first feature information and the second feature information. The forward cross-attention guides the spatial features with the temporal features, so that the model can focus on the spatial features related to the temporal changes. The reverse cross-attention constrains the temporal features with the spatial features, focusing on the influence of the spatial structure on the temporal changes. Further, a fusion weight is obtained, and the first feature information and the second feature information are weighted fused with the fusion weight factor, so that the spatio-temporal feature interaction is more sufficient, and a fusion feature with better fusion effect is obtained.

[0184] On the basis of the above embodiment, Figure 6 is a logical flowchart of training of a coal mine pressure prediction model provided by an embodiment of the present application. As Figure 6 shown, data acquisition is performed through a sensor set, data standardization change is determined for a target time period to obtain target sample data, mutual information calculation is performed on the originally collected data to obtain a dynamic adjacency matrix, the dynamic adjacency matrix is input into an adaptive graph convolution layer for spatial feature information extraction, position coding is performed on the target sample data, and the position coding is embedded for multi-head self-attention analysis to obtain temporal feature information. The spatial feature information and the temporal feature information are fused by bidirectional cross-attention to obtain a fusion feature. Based on the fusion feature, a prediction result is obtained, and a coal mine pressure prediction model is trained according to the prediction result to obtain a trained coal mine pressure prediction model.

[0185] On the basis of the above embodiment, Figure 7 is a schematic diagram of a coal mine pressure prediction method provided by an embodiment of the present application. As Figure 7 shown, the method comprises:

[0186] S701, collecting coal mine detection data in a coal mine environment to be analyzed.

[0187] In some embodiments, one or more sensors can be included in the coal mine environment for data collection, and each sensor can collect one or more feature data.

[0188] In some embodiments, the coal mine detection data can be data within a preset time period, that is, the coal mine detection data is data of three dimensions of sensors, features, and time.

[0189] S702, input the coal mine detection data into the pre-trained coal mine pressure prediction model, and output the mine pressure prediction result.

[0190] The coal mine pressure prediction model is a model obtained by using the training method of the coal mine pressure prediction model in the above-mentioned embodiments of the present application. The coal mine detection data is input into the coal mine pressure prediction model, and the coal mine pressure prediction model directly outputs the mine pressure prediction result, which improves the prediction efficiency while ensuring the prediction accuracy.

[0191] In this embodiment, after obtaining the coal mine detection data in the analysis environment, the pre-trained coal mine pressure prediction model is called for mine pressure prediction. The coal mine pressure prediction model obtains more accurate mine pressure prediction results by fully extracting and fusing time features and space features, which improves the overall efficiency of mine pressure prediction and improves the accuracy of mine pressure prediction results.

[0192] To implement the above-mentioned embodiments, the present application further provides a coal mine pressure prediction model training device.

[0193] Figure 8 A structural schematic diagram of a coal mine pressure prediction model training device provided by an embodiment of the present application is shown in FIG. 8. Figure 8 As shown in FIG. 8, the coal mine pressure prediction model training device 800 includes:

[0194] The acquisition module 801 is configured to acquire sample coal mine detection data in a target time period in a coal mine scene database.

[0195] The processing module 802 is configured to input the sample coal mine detection data into the coal mine pressure prediction model, and generate a dynamic adjacency matrix between sensors according to the sample coal mine detection data by a processing layer of the coal mine pressure prediction model.

[0196] The convolution module 803 is configured to input the dynamic adjacency matrix and the sample coal mine detection data into a convolution layer for convolution processing to obtain space feature information.

[0197] The encoding module 804 is configured to encode and extract time features of the sample coal mine detection data by an encoding layer to obtain time feature information.

[0198] The fusion module 805 is configured to fuse the spatial feature information and the time feature information in a fusion layer to obtain fused features, and obtain a mine pressure prediction result according to the fused features.

[0199] The training module 806 is configured to train the coal mine pressure prediction model according to the mine pressure prediction result and an actual result to obtain a pre-trained coal mine pressure prediction model.

[0200] Further, in a possible implementation manner of the embodiment of the present application, the processing module 802 comprises:

[0201] According to the sample coal mine detection data, the joint probability density between each two sensors and the marginal probability density of each sensor are determined;

[0202] Based on the joint probability density and the marginal probability density, mutual information between the corresponding two sensors is determined.

[0203] The edge weight between the corresponding two sensors is determined according to the mutual information, and a dynamic adjacency matrix is obtained based on the corresponding edge weights between all the sensors as matrix elements.

[0204] Further, in a possible implementation manner of the embodiment of the present application, the convolution layer comprises at least one sub-convolution layer, and the convolution module 803 comprises:

[0205] For any sub-convolution layer in the convolution layer, input data is determined; when the any sub-convolution layer is a first layer, the input data is target sample data obtained by preprocessing the sample coal mine detection data, and when the any sub-convolution layer is a layer other than the first layer, the input data is output data of a previous sub-convolution layer;

[0206] The augmented adjacency matrix is obtained by adding the preset matrix and the dynamic adjacency matrix;

[0207] Degree matrix calculation is performed on the augmented adjacency matrix, and symmetric normalization processing is performed on the degree matrix to obtain a target matrix;

[0208] Based on the target matrix and the input data, output data of the current sub-convolution layer is determined.

[0209] Based on the output data of the last sub-convolution layer, the spatial feature information is determined.

[0210] Further, in a possible implementation manner of the embodiment of the present application, the encoding module 804 comprises:

[0211] The working condition parameter vector and the mining speed at the current sampling time are obtained, and the working condition parameters at least include geological parameters, mining process parameters and environmental parameters;

[0212] The position is encoded based on the working condition parameter vector and the mining speed to obtain an encoding result.

[0213] The encoding result and the target sample data obtained by preprocessing the sample coal mine detection data are subjected to multi-head self-attention processing to obtain time feature information.

[0214] Further, in a possible implementation of the embodiment of the present application, the encoding module 804 comprises:

[0215] The process of obtaining the encoding result adopts the following formula:

[0216]

[0217] wherein, is the encoding result; is a sampling point index; is a feature dimension index; is a total feature dimension; is a mining speed; is a working condition parameter vector; is a weight coefficient.

[0218] Further, in a possible implementation of the embodiment of the present application, the fusion module 805 comprises:

[0219] The spatial feature information and the time feature information are subjected to forward cross-attention processing and reverse cross-attention processing respectively to obtain first feature information and second feature information;

[0220] The first feature information and the second feature information are subjected to weighted fusion to obtain fusion features.

[0221] Further, in a possible implementation of the embodiment of the present application, the fusion module 805 comprises:

[0222] According to the spatial feature information, the time feature information and a first parameter matrix set, a first vector set is determined, the first parameter matrix set comprising a learnable first Query weight matrix, a first Key weight matrix and a first Value weight matrix, and the first vector set comprising a first Query vector, a first Key vector and a first Value vector;

[0223] According to the spatial feature information, the time feature information and a second parameter matrix set, a second vector set is determined, the second parameter matrix set comprising a learnable second Query weight matrix, a second Key weight matrix and a second Value weight matrix, and the second vector set comprising a second Query vector, a second Key vector and a second Value vector;

[0224] According to the first vector set and the second vector set, a first attention matrix and a second attention matrix are respectively determined;

[0225] According to the first attention matrix and the first Value vector, first feature information is determined;

[0226] According to the second attention matrix and the second Value vector, second feature information is determined.

[0227] Further, in a possible implementation manner of the embodiment of the present application, the fusion module 805 comprises:

[0228] Obtaining a working condition parameter vector at a current sampling time, the working condition parameter at least comprising: geological parameters, mining process parameters and environmental parameters;

[0229] Generating a fusion weight based on the full connection layer according to the working condition parameter vector;

[0230] Weighted fusion of the first feature information and the second feature information based on the fusion weight to obtain fusion features.

[0231] Further, in a possible implementation manner of the embodiment of the present application, the fusion module 805 comprises:

[0232] Inputting the fusion features, the working condition parameter vector and the historical prediction error into an inference layer for inference prediction to obtain a mine pressure prediction result.

[0233] It should be noted that the aforementioned explanation and description of the training method embodiment of the coal mine pressure prediction model also applies to the training device of the coal mine pressure prediction model of this embodiment, which will not be described here.

[0234] In the embodiment, sample coal mine detection data is obtained, and the association between sensors is obtained according to the sample coal mine detection data to construct a dynamic adjacency matrix. The spatial feature information is obtained based on the dynamic adjacency matrix. Further, the sample coal mine detection data is encoded and time feature is extracted to obtain time feature information. The time feature information and the time feature information are fused to make the interaction of time feature and space feature more sufficient. The fusion features after fusion are used to obtain the prediction result. The features are fully mined from two dimensions of space and time to more accurately capture the long-term trend of mine pressure, effectively improve the accuracy of result prediction, better capture the complex space-time coupling relationship in mine pressure data, and train a coal mine pressure prediction model with more accurate prediction result.

[0235] In order to realize the above-mentioned embodiments, the present application further provides a coal mine pressure prediction device.

[0236] Figure 9 A structural schematic diagram of a coal mine pressure prediction device provided by the embodiment of the present application is shown in FIG. 8. As shown in FIG. 8, the coal mine pressure prediction device comprises a data preprocessing module 801, a feature extraction module 802, a fusion module 805, an inference layer 806 and a result output module 807.Figure 9 As shown, the coal mine pressure prediction device 900 includes:

[0237] The acquisition module 901 is configured to collect coal mine detection data in a coal mine to be analyzed.

[0238] The prediction module 902 is configured to input the coal mine detection data into a pre-trained coal mine pressure prediction model and output a mine pressure prediction result. The coal mine pressure prediction model is a model obtained by using the training method of the coal mine pressure prediction model of the first aspect.

[0239] It should be noted that the foregoing explanation and description of the coal mine pressure prediction method embodiment also applies to the coal mine pressure prediction device of this embodiment, which will not be described here.

[0240] In this embodiment, after obtaining the coal mine detection data in the analysis environment, the pre-trained coal mine pressure prediction model is called to predict the mine pressure. The coal mine pressure prediction model obtains more accurate mine pressure prediction results by fully extracting and fusing the time characteristics and the spatial characteristics, thereby improving the overall efficiency of the mine pressure prediction and the accuracy of the mine pressure prediction results.

[0241] To implement the above-mentioned embodiments, the present application further provides an electronic device, comprising a processor and a memory connected with the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0242] To implement the above-mentioned embodiments, the present application further provides a computer readable storage medium, which stores computer execution instructions, and the computer execution instructions are executed by a processor to implement the method provided in the foregoing embodiments.

[0243] To implement the above-mentioned embodiments, the present application further provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the method provided in the foregoing embodiments.

[0244] The collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in the present application comply with relevant laws and regulations and do not violate public order and good customs.

[0245] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy policies and procedures.

[0246] This application is intended to provide an implementation scheme for users to selectively prevent the use or access to their personal information data. Specifically, this disclosure is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information is de-identified to protect user privacy.

[0247] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0248] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0249] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0250] The logic and / or steps represented in flow diagrams or otherwise described herein, for example, can be considered as a sequence of executable instructions, and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For purposes of this specification, a "computer-readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be a product of the manufacturing and / or processing. The computer-readable medium can include, but is not limited to, the following: an electronic connection (an electronic device having one or more wires), a portable computer diskette (a magnetic device), a RAM (random access memory), a ROM (read-only memory), an EPROM (erasable programmable ROM) or Flash memory, an optical fiber device, and a portable CD ROM. Additionally, the computer-readable medium can be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via, for example, an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in order to be executed.

[0251] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. As such, in some embodiments, specifically configured hardware can be used to implement at least some of the functionality described herein. For example, if implemented in hardware, the hardware can include any or a combination of the following: a discrete logic circuit having logic gates for implementing logic functions upon data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0252] Those of skill in the art would understand that information and signals can be represented using any of a variety of technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that can be referenced throughout the above description can be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0253] In addition, each of the functional units in the various embodiments of the present application can be integrated in one processing module, or each of the units can be physically present separately, or two or more units can be integrated in one module. The integrated module can be implemented in the form of hardware or in the form of a software functional module. When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0254] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for training a coal mine pressure prediction model, characterized in that, The method comprises: obtaining sample coal mine detection data in a target time period in a coal mine scene database; inputting the sample coal mine detection data into a coal mine pressure prediction model, and generating a dynamic adjacency matrix between sensors according to the sample coal mine detection data by a processing layer of the coal mine pressure prediction model; performing convolution processing on the dynamic adjacency matrix and the sample coal mine detection data through a convolution layer to obtain spatial feature information; performing coding and time feature extraction on the sample coal mine detection data through an encoding layer to obtain time feature information; inputting the spatial feature information and the time feature information into a fusion layer for fusion processing to obtain fusion features, and obtaining a mine pressure prediction result according to the fusion features; training the coal mine pressure prediction model according to the mine pressure prediction result and an actual result to obtain a pre-trained coal mine pressure prediction model; wherein the generating of the dynamic adjacency matrix between sensors according to the sample coal mine detection data by the processing layer of the coal mine pressure prediction model comprises: determining a joint probability density between each two sensors and an edge probability density of each sensor according to the sample coal mine detection data; determining mutual information between the corresponding two sensors based on the joint probability density and the edge probability density; determining an edge weight between the corresponding two sensors based on the mutual information, and obtaining the dynamic adjacency matrix based on the corresponding edge weights between all sensors as matrix elements; the inputting of the spatial feature information and the time feature information into the fusion layer for fusion processing to obtain fusion features comprises: determining a first vector set according to the spatial feature information, the time feature information and a first parameter matrix set, wherein the first parameter matrix set includes a first query (Query) weight matrix, a first key (Key) weight matrix and a first value (Value) weight matrix that can be learned, and the first vector set includes a first Query vector, a first Key vector and a first Value vector; determining a second vector set according to the spatial feature information, the time feature information and a second parameter matrix set, wherein the second parameter matrix set includes a second Query weight matrix, a second Key weight matrix and a second Value weight matrix that can be learned, and the second vector set includes a second Query vector, a second Key vector and a second Value vector; determining a first attention matrix and a second attention matrix respectively according to the first vector set and the second vector set; determining first feature information according to the first attention matrix and the first Value vector; determining second feature information according to the second attention matrix and the second Value vector; obtaining a working condition parameter vector at a current sampling time, wherein the working condition parameter at least includes a geological parameter, a mining process parameter and an environmental parameter; generating a fusion weight based on a full connection layer according to the working condition parameter vector; performing weighted fusion on the first feature information and the second feature information based on the fusion weight to obtain fusion features.

2. The method of claim 1, wherein, The convolution layer includes at least one sub-convolution layer, and the dynamic adjacency matrix and the sample coal mine detection data are input into the convolution layer for convolution processing to obtain spatial feature information, including: For any sub-convolution layer in the convolution layer, input data is determined; when the any sub-convolution layer is the first layer, the input data is target sample data obtained by preprocessing the sample coal mine detection data, and when the any sub-convolution layer is other layers except the first layer, the input data is output data of the previous sub-convolution layer; Based on the preset matrix and the dynamic adjacency matrix, an augmented adjacency matrix is obtained by addition; Degree matrix calculation is performed on the augmented adjacency matrix, and symmetric normalization processing is performed on the degree matrix to obtain a target matrix; Based on the target matrix and the input data, output data of the current sub-convolution layer is determined; Based on the output data of the last layer of sub-convolution layer, the spatial feature information is determined.

3. The method of claim 1, wherein, The sample coal mine detection data is encoded and time feature information is extracted through the encoding layer, including: Obtain the working condition parameter vector and the mining speed of the current sampling time, and the working condition parameter at least includes: geological parameter, mining technology parameter and environmental parameter; Based on the working condition parameter vector and the mining speed, position encoding is performed to obtain an encoding result; The encoding result and the target sample data obtained by preprocessing the sample coal mine detection data are subjected to multi-head self-attention processing to obtain the time feature information.

4. The method of claim 3, wherein, The process of position encoding based on the working condition parameter vector and the mining speed to obtain the encoding result adopts the following formula: wherein, is the encoding result; is the sample point index; is the feature dimension index; is the total feature dimension; is the exploitation speed; is the working condition parameter vector; is the weight coefficient.

5. The method of claim 1, wherein, According to the fusion feature, a mine pressure prediction result is obtained, including: The fusion feature, the working condition parameter vector and the historical prediction error are input into an inference layer for inference prediction to obtain a mine pressure prediction result.

6. A method of coal mine pressure prediction, characterised by, The method includes: Collecting coal mine detection data in a coal mine environment to be analyzed; The coal mine detection data is input into a pre-trained coal mine pressure prediction model to output a mine pressure prediction result; The coal mine pressure prediction model is a model obtained by using the training method of the coal mine pressure prediction model in any one of claims 1-5.

Images