CN-122020488-A - Method for predicting gangue rate of fully-mechanized caving face based on vibration signal information fusion

CN122020488ACN 122020488 ACN122020488 ACN 122020488ACN-122020488-A

Abstract

The invention relates to the technical field of intelligent coal exploitation, in particular to a method for predicting gangue rate of a fully mechanized caving face based on vibration signal information fusion, which comprises the steps of firstly obtaining vibration signal data of a tail beam of a coal gangue caving impact hydraulic support from a test bed, and respectively carrying out time domain segmentation, discrete Fourier transformation and graph modeling method pretreatment based on cosine similarity combined k-NNG on the data of various data sets on the basis; the preprocessing data is input into a multichannel fusion model based on a graph neural network to generate fusion characteristics based on a topological graph, the characteristics are input into a space-time graph neural network prediction module to finally output gangue rate, and the model is jointly optimized by adopting an MSE loss function and a classification auxiliary loss function. The method solves the problems that the existing fully-mechanized caving face waste rock rate detection/prediction method is low in precision, poor in robustness and incapable of being continuously quantized in real time, and achieves end-to-end and high-precision fully-mechanized caving face waste rock rate prediction.

Inventors

YANG SHANGUO
LI JIANG
LIU HOUGUANG
WANG YAO
Jiao Binglong
QIU HAIFENG
CHENG XINYU
Liu Ruze

Assignees

中国矿业大学

Dates

Publication Date: 20260512
Application Date: 20260401

Claims (10)

1. A method for predicting gangue rate of a fully-mechanized caving face based on vibration signal information fusion is characterized by comprising the following steps: S1, acquiring vibration signals of a coal gangue collapse impact hydraulic support, establishing a total data set and dividing a training data set and a test data set; S2, carrying out time domain segmentation and Discrete Fourier Transform (DFT) on the sampled data to extract frequency domain amplitude characteristics, calculating inter-node edge weights of each sampled data by adopting cosine similarity according to the frequency domain amplitude characteristics, sparsifying and constructing a topological graph by a k-NNG method, and establishing a graph structure data set; S3, constructing a gangue rate prediction model based on multi-channel vibration signal information fusion of a graph neural network GNN, inputting graph structure data into the model to generate fusion characteristics based on a topological graph, inputting the fusion characteristics into a space-time graph neural network ST-GNN prediction module, and outputting gangue rate prediction values; S4, optimizing the model by adopting a joint loss function; s5, model training and evaluation.
2. The method for predicting gangue rate on fully-mechanized caving face based on vibration signal information fusion as claimed in claim 1, wherein the time-domain segmentation in step S2 comprises collecting vibration signals at a fixed sampling frequency f s , segmenting the original signals according to a preset time window length T w , setting an overlapping rate delta between adjacent segments, each segment serving as a time-domain segment and expressed as ; The DFT extracting the frequency domain amplitude feature comprises the following steps of carrying out DFT on each time domain segment: ; where j is an imaginary unit as defined in the present claim, k is a frequency component index as defined in the present claim, represents a kth frequency component, For complex conversion result of the ith time domain segment on the kth frequency component, n is the sampling point index in the claim, x n is the input time domain discrete signal sampling value, i.e. the nth sampling point corresponding to the ith time domain segment, L is the length of the signal segment, i.e. the total number of sampling points participating in DFT calculation, and the formula is calculated by a rotation factor Extracting the amplitude characteristics of the signal at different frequency components, and keeping the amplitude of the positive frequency part as the frequency domain characteristic vector corresponding to the segment Obtaining a node characteristic matrix Where M is the number of time domain segments.
3. The method for predicting gangue rate of fully-mechanized caving face based on vibration signal information fusion as claimed in claim 1, wherein calculating the inter-node edge weight of each sampled data by cosine similarity in step S2 comprises calculating the edge weight between cosine similarity nodes of frequency domain feature vectors v i and v j corresponding to any two nodes as I 1 and i 2 are node indexes, and the calculation mode is as follows: ; The sparse construction topological graph through the k-NNG method comprises the steps of reserving only edges between k neighbor nodes with highest cosine similarity with each node, setting the edge weight of the node and other nodes to be zero to obtain a sparse adjacent matrix A, and constructing structural data of the graph, wherein the construction structural data comprises the steps of repeating the calculation flow on all samples of all channels to obtain a characteristic matrix set of each channel node And sparse adjacent matrix set corresponding to each channel node P is the number of channels and a is the channel index described in the present claim.
4. The method for predicting the gangue rate of the fully-mechanized caving face based on vibration signal information fusion according to claim 1 is characterized in that in the step S3, a multi-channel vibration signal information fusion based on GNN comprises a GNN multi-channel information fusion module, an ST-GNN prediction module and an MLP regression output module, wherein the GNN multi-channel information fusion module comprises an MLP dimension reduction module, a global adjacent matrix construction module, a multi-layer GCN message transmission module, a channel attention module and a multi-head attention fusion module.
5. The method for predicting gangue rate of fully-mechanized caving face based on vibration signal information fusion as claimed in claim 4, wherein the sparse adjacent matrix corresponding to each channel node is characterized by Building diagonal block element sets P is the number of channels, a is the channel index described in the present claim, and each diagonal block element satisfies Wherein I M is M-order identity matrix, and the inter-channel connection weight is constructed according to cosine similarity between average feature vectors of all channels, and the inter-channel similarity is obtained when the inter-channel connection weight is equal to the inter-channel similarity , When the threshold value is reached, adding a connection weight at a corresponding position to obtain a global adjacency matrix A global , wherein the construction form is as follows: ; Wherein W ab represents a weight matrix between an a-th channel and a b-th channel, a and b are channel indexes described in the present claim, and the weight matrix satisfies , , And And the average feature vectors are respectively expressed as an a-th channel and a b-th channel, and in the multi-layer GCN message transfer module, each layer of GCN aggregates the feature information of adjacent nodes according to a global adjacency matrix, and the aggregation range comprises the adjacent nodes in the same channel and the adjacent nodes between different channels.
6. The method for predicting the gangue rate of the fully-mechanized caving face based on vibration signal information fusion according to claim 4, wherein the channel attention module calculates the attention weight of each channel after each layer of GCN, multiplies the attention weight by the channel characteristics to realize the self-adaptive weighting of the channel characteristics, and the multi-head attention fusion module adopts a plurality of attention heads to respectively perform linear transformation of Query, key Key and Value, and the calculation mode of each attention head is as follows: ; Where c is the attention head index described in the present claim, h is the total number of attention heads, o c is the output vector of the c-th attention head, representing the weighted fusion feature extracted by the attention head, a is the channel index described in claim 5, softmax a (·) is the normalization function for calculating the a-th channel in the c-th attention head, Q c is the query vector for the c-th attention header, representing the global feature map, The weight matrix may be learned for the c-th attention head query, g avg is a global average feature vector, obtained by global average pooling of all channel features, used to guide attention allocation, K a,c is the key vector corresponding to the a-th channel in the c-th attention head, For the key of the c-th attention head, the weight matrix can be learned, g a is the node characteristic vector of the a-th channel after the processing of the GCN layer, V a,c is the value vector corresponding to the a-th channel in the c-th attention header, For the c-th attention head value, a weight matrix can be learned, d c is a scaling factor, the value is the characteristic dimension of the vector k a,c , gradient disappearance caused by overlarge dot product is prevented, and the output of each attention head is spliced to obtain the fusion characteristic of all channels based on the topological graph: ; Wherein, F fused is a multi-channel fusion feature matrix obtained by final calculation, concat (·) represents a stitching operation of connecting outputs of all attention heads into one long vector, and W o is a projection matrix of mapping the stitched high-dimensional features back to the target dimension.
7. The method for predicting the gangue rate of the fully-mechanized caving face based on vibration signal information fusion according to claim 4, wherein the ST-GNN prediction module comprises a one-dimensional time convolution layer 1D-Conv, a partition space convolution layer and a global pooling layer, the partition space convolution layer divides a space neighborhood into three areas of a self-ring area, a near neighborhood area and a far neighborhood area, and the calculation mode is as follows: ; Wherein H (l+1) and H (l) represent the space-time convolutional layer hidden layer matrix of the first layer and the first layer respectively, wherein H (l) terms form residual connection to prevent deep network degradation, M is the region spatial index described in the claim, and represents three regions respectively, mapped to self-ring domain, near-neighborhood and far-neighborhood respectively, M m is a leachable mask matrix, and the method comprises the steps of Hadamard product Element-by-element weighting is carried out on the space-time convolution result, the self-adaptive fine adjustment of different neighborhood weights is realized, For the space-time adjacency matrix of the m-th area after decoupling, Z t is the input space-time characteristic tensor, represents the time domain sequence information after one-dimensional time convolution processing, And constructing a space-time adjacent matrix of the ST-GNN prediction module according to a global adjacent matrix and a full-connection matrix of adjacent time steps, wherein the space-time adjacent matrix is expressed as: ; Wherein A st represents a space-time adjacency matrix, u is the number of adjacent time step samples, A global spatial adjacency matrix for the u-th sample, located at a diagonal position, for describing the static spatial topological relationship between different channels at the same sampling time, For a P-order all-1 matrix, P is the number of channels of claim 5, namely a fully-connected time-step projection matrix, time-domain correlation of nodes between adjacent time steps is established, and spatial dependence and time dependence are captured.
8. The method for predicting the gangue rate of the fully-mechanized caving face based on vibration signal information fusion according to claim 1, wherein in the step S4, the joint loss function is a sum of MSE loss and classification auxiliary loss: ; Wherein the method comprises the steps of For the regression loss of the gangue rate MSE, And the lambda is the balance coefficient for classifying loss of the coal discharging state.
9. The method for predicting the gangue rate of the fully mechanized caving face based on vibration signal information fusion according to claim 1, wherein the model training and evaluation in the step S5 comprises initializing network parameters, inputting training samples into the model, calculating joint loss and optimizing the model network parameters, judging model effects according to MSE, MAE, R indexes after training of each round, selecting optimal model parameters, and inputting test sets into the trained model evaluation performance.
10. The fully-mechanized caving face gangue rate prediction system based on vibration signal information fusion is applied to the fully-mechanized caving face gangue rate prediction method based on vibration signal information fusion, and is characterized by comprising the following steps: The signal acquisition module is used for acquiring a vibration signal of the coal gangue collapse impact hydraulic support tail beam; The diagram structure construction module is used for carrying out time domain segmentation, DFT, cosine similarity calculation and k-NNG sparsification on the vibration signal to construct diagram structure data; the feature fusion module is used for inputting the graph structure data into a GNN-based multichannel fusion model to generate fusion features based on a topological graph; And the prediction output module is used for inputting the fusion characteristics into the ST-GNN prediction module and outputting a gangue rate predicted value.

Description

Method for predicting gangue rate of fully-mechanized caving face based on vibration signal information fusion Technical Field The invention relates to the technical field of intelligent coal mining, in particular to a method for realizing real-time prediction of gangue content in the coal caving process of a fully-mechanized caving face by utilizing multichannel vibration signals, graph signal processing and graph neural network for information fusion. Background Fully mechanized caving mining (comprehensive mechanized caving roof mining) is a mainstream high-yield and high-efficiency mining mode of thick and extra-thick coal beds in China. The 'see gangue closing door' is an operation principle of a coal caving process, namely, when top coal is completely caving and top board gangue begins to collapse in a large amount, a coal caving opening is timely closed, so that the quality of coal is ensured, and the resource recovery rate is improved. In the prior art, gangue rate judgment mainly depends on manual experience observation or manual sampling test, and has the problems of strong hysteresis, strong subjectivity, incapability of realizing real-time closed-loop control and the like. The partial automation attempt is mainly based on a single sensor vibration signal, an acoustic signal or an image, but is influenced by complex working conditions such as underground strong noise, continuous impact of multiple coal blocks, random change of coal gangue block and the like, and the single source signal identification precision and the robustness are insufficient. In recent years, although small sample coal gangue recognition methods based on time-frequency analysis and machine learning appear, the problems still exist that the strong non-stationary, nonlinear and multi-scale characteristics of vibration signals are difficult to effectively describe, the spatial topological relation and the inter-channel complementarity between sensors are ignored by single-channel or simple multi-channel fusion, an end-to-end prediction model from the vibration signals to gangue content (continuous numerical value) is lacking, most of the methods stay at the level of coal/gangue two classification or discrete state classification, and the method for realizing high-precision gangue content continuous prediction in fully-mechanized exploitation is lacking at present. Therefore, a method for realizing high-precision continuous prediction of gangue content by comprehensively utilizing the space-time correlation characteristics of multichannel vibration signals is needed to support intelligent gangue-seeing closing decision of a fully-mechanized caving face. Disclosure of Invention The invention aims to provide a method for predicting gangue rate of a fully-mechanized caving face based on vibration signal information fusion, so as to solve the problems in the background technology. In order to achieve the above purpose, the invention provides the following technical scheme that in a first aspect, the invention provides a method for predicting the gangue rate of a fully mechanized caving face based on vibration signal information fusion, which comprises the following steps: S1, acquiring vibration signals of a coal gangue collapse impact hydraulic support, establishing a total data set and dividing a training data set and a test data set; s2, carrying out time domain segmentation and Discrete Fourier Transform (DFT) on the sampled data to extract frequency domain amplitude characteristics, calculating inter-node edge weights of each sampled data by adopting cosine similarity according to the frequency domain amplitude characteristics, sparsifying and constructing a topological graph by a k-NNG method, and establishing a graph structure data set; S3, constructing a gangue rate prediction model based on multi-channel vibration signal information fusion of a Graph Neural Network (GNN), inputting graph structure data into the model to generate fusion characteristics based on a topological graph, inputting the fusion characteristics into a space-time graph neural network (ST-GNN) prediction module, and outputting gangue rate prediction values; S4, optimizing the model by adopting a joint loss function; s5, model training and evaluation. As a further improvement of the invention, the time domain segmentation in step S2 comprises collecting the vibration signal at a fixed sampling frequency f s, segmenting the original signal according to a preset time window length T w, setting an overlap rate delta between adjacent segments, each segment being a time domain segment and expressed as; The DFT extracting the frequency domain amplitude feature comprises the following steps of carrying out DFT on each time domain segment: ; where j is an imaginary unit as defined in the present claim, k is a frequency component index as defined in the present claim, represents a kth frequency component, For complex conversion result of the ith time domain segme