Fully mechanized caving face gangue content prediction method based on vibration signal information fusion

Abstract

The invention relates to the technical field of intelligent coal mining, in particular to a fully mechanized caving face gangue content prediction method based on vibration signal information fusion. On the basis, time domain segmentation, discrete Fourier transform and preprocessing of a graph modeling method based on cosine similarity joint k-NNG are carried out on data of various data sets; inputting the preprocessed data into a multi-channel fusion model based on a graph neural network, and generating fusion features based on a topological graph; inputting the features into a space-time diagram neural network prediction module, and finally outputting the gangue content; and meanwhile, an MSE loss function and a classification auxiliary loss function are adopted to carry out joint optimization on the model. The problems that an existing fully mechanized caving face gangue content detection / prediction method is low in precision and poor in robustness, and real-time continuous quantification cannot be achieved are solved, and end-to-end and high-precision fully mechanized caving face gangue content prediction is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent coal mining technology, specifically to a method for real-time prediction of gangue content during coal release in fully mechanized longwall mining faces by integrating information from multi-channel vibration signals, graph signal processing, and graph neural networks. Background Technology

[0002] Fully mechanized top coal caving mining is the mainstream high-yield and high-efficiency mining method for thick and extra-thick coal seams in my country. "Closing the gate when gangue is seen" is the operational principle of the coal caving process, which means that the coal caving gate should be closed in time when the top coal is exhausted and the gangue on the roof begins to collapse in large quantities, so as to ensure coal quality and improve resource recovery rate.

[0003] In existing technologies, the determination of gangue content mainly relies on manual observation or manual sampling and testing, which suffers from problems such as strong lag, high subjectivity, and inability to achieve real-time closed-loop control. Some automation attempts are mainly based on single-sensor vibration signals, sound signals, or images, but are affected by complex working conditions such as strong underground noise, continuous impact of multiple coal blocks, and random changes in coal and gangue block size, resulting in insufficient accuracy and robustness of single-source signal recognition.

[0004] In recent years, although some small-sample coal and gangue identification methods based on time-frequency analysis and machine learning have emerged, the following problems still exist: it is difficult to effectively characterize the strong non-stationary, nonlinear, and multi-scale characteristics of vibration signals; single-channel or simple multi-channel fusion ignores the spatial topological relationship between sensors and the complementarity between channels; there is a lack of end-to-end prediction models that directly derive gangue content (continuous numerical value) from vibration signals, with most remaining at the level of coal / gangue binary classification or discrete state classification, and currently lacking a method suitable for achieving high-precision continuous prediction of gangue content in fully mechanized longwall mining. Therefore, there is an urgent need for a method that can comprehensively utilize the spatial-temporal correlation characteristics of multi-channel vibration signals to achieve high-precision continuous prediction of gangue content, in order to support the intelligent "shutdown upon gangue sighting" decision-making in fully mechanized longwall mining faces. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention proposes a method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion, comprising the following steps:

[0007] S1. Obtain the vibration signal of the hydraulic support impacted by the collapse of coal gangue, establish the total dataset and divide it into training dataset and test dataset;

[0008] S2. Perform time-domain segmentation and Discrete Fourier Transform (DFT) on the sampled data to extract frequency domain amplitude features. Calculate the edge weights between nodes of each sampled data based on the frequency domain amplitude features using cosine similarity. Then, construct a topological graph by sparsification using the k-NNG method to establish a graph structure dataset.

[0009] S3. Build a gangue content prediction model based on the fusion of multi-channel vibration signal information using graph neural networks (GNNs). Input the graph structure data into the model to generate fusion features based on the topological graph. Then input the fusion features into the spatiotemporal graph neural network (ST-GNN) prediction module to output the predicted gangue content value.

[0010] S4. Optimize the model using a joint loss function;

[0011] S5. Model Training and Evaluation.

[0012] As a further improvement of the present invention, the time-domain segmentation in step S2 includes: using a fixed sampling frequency f s Vibration signals are collected according to a preset time window length T. w The original signal is segmented, with an overlap rate δ set between adjacent segments. Each segment is treated as a time-domain segment and represented as follows. ;

[0013] The DFT extraction of frequency domain amplitude features includes: performing a DFT on each time domain segment:

[0014] ;

[0015] Where j is the imaginary unit as described in this claim, and k is the frequency component index as described in this claim, representing the k-th frequency component. Let x be the complex transform result of the i-th time domain segment on the k-th frequency component, n be the sampling point index as described in this claim, and x be the sampling point index as described in this claim. n Let L be the sampled value of the input discrete-time signal, corresponding to the nth sample point of the i-th time segment, and L be the length of the signal segment, i.e., the total number of sample points participating in the DFT calculation. This formula uses a rotation factor. Extract the amplitude features of the signal at different frequency components; retain the amplitude of the positive frequency component as the frequency domain feature vector corresponding to that segment. The node feature matrix is ​​obtained. , where M is the number of time-domain segments.

[0016] As a further improvement of the present invention, the step S2 of calculating the edge weights between nodes of each sampled data using cosine similarity includes: calculating the frequency domain feature vectors v corresponding to any two nodes. i With v j The edge weights between nodes with cosine similarity are represented as i1 and i2 are node indices, calculated as follows:

[0017] ;

[0018] The method of constructing a sparse topology graph using the k-NNG method includes: for each node, retaining only the edges between the k neighboring nodes with the highest cosine similarity to that node, setting the edge weights between that node and other nodes to zero, and obtaining a sparse adjacency matrix A; the method of establishing graph structure data includes: repeating the above calculation process for all samples of all channels to obtain the feature matrix set of each channel node. and the set of sparse adjacency matrices corresponding to each channel node. P is the number of channels, and a is the channel index as described in this claim.

[0019] As a further improvement of the present invention, the gangue content prediction model based on GNN multi-channel vibration signal information fusion in step S3 includes a GNN multi-channel information fusion module, an ST-GNN prediction module, and an MLP regression output module; the GNN multi-channel information fusion module includes an MLP dimensionality reduction module, a global adjacency matrix construction module, a multi-layer GCN message passing module, a channel attention module, and a multi-head attention fusion module.

[0020] As a further improvement to the present invention, based on the sparse adjacency matrix corresponding to each channel node... Construct a collection of diagonal block elements P is the number of channels, a is the channel index as described in this claim, and each diagonal block element satisfies , where I M Given an M-order identity matrix, channel connection weights are constructed based on the cosine similarity between the average eigenvectors of each channel. When the channel similarity... ( When the threshold is reached, connection weights are added at the corresponding positions to obtain the global adjacency matrix A. global Its structural form is as follows:

[0021] ;

[0022] Among them, W ab Let represent the weight matrix between the a-th channel and the b-th channel, where a and b are the channel indices as described in this claim, and the weight matrix satisfies ... , , and These are represented as the average feature vectors of the a-th and b-th channels, respectively. In the multi-layer GCN message passing module, each layer of GCN aggregates the feature information of adjacent nodes according to the global adjacency matrix. The aggregation range includes adjacent nodes within the same channel and adjacent nodes between different channels.

[0023] As a further improvement of the present invention, the channel attention module calculates the attention weight of each channel after each layer of GCN, and multiplies the attention weight by the channel features to achieve adaptive weighting of the channel features; the multi-head attention fusion module uses multiple attention heads to perform linear transformations of query, key, and value respectively, and the calculation method of each attention head is as follows:

[0024] ;

[0025] Where c is the attention head index as described in this claim, h is the total number of attention heads, and o c Let be the output vector of the c-th attention head, representing the weighted fusion feature extracted by that attention head, where 'a' is the channel index as described in claim 5, and 'softmax' is the output vector of the c-th attention head. a (·) is the normalization function for calculating the a-th channel in the c-th attention head. q c Let c be the query vector of the attention head, representing the global feature map. Let g be the query learnable weight matrix for the c-th attention head. avg The global average feature vector, obtained by performing global average pooling on the features of all channels, is used to guide attention allocation. k a,c Let be the key vector corresponding to the a-th channel in the c-th attention head. Let g be the key-learnable weight matrix for the c-th attention head. a Let be the node feature vector of the a-th channel after processing by the GCN layer. v a,c Let a be the value vector corresponding to the a-th channel in the c-th attention head. The weight matrix can be learned for the value of the c-th attention head, d c The scaling factor is the value of the vector k. a,c To minimize gradient vanishing due to excessively large dot products, the outputs of each attention head are concatenated to obtain the fused features of all channels based on the topological graph.

[0026] ;

[0027] Among them, F fusedFor the final calculated multi-channel fused feature matrix, Concat(·) represents the concatenation operation that joins the outputs of all attention heads into a single long vector, W. o This is a projection matrix that maps the stitched high-dimensional features back to the target dimension.

[0028] As a further improvement of the present invention, the ST-GNN prediction module includes a one-dimensional temporal convolutional layer (1D-Conv), a partitioned spatial convolutional layer, and a global pooling layer; the partitioned spatial convolutional layer divides the spatial neighborhood into three regions: a self-loop region, a near neighbor region, and a far neighbor region, and its calculation method is as follows:

[0029] ;

[0030] Where H (l+1) and H (l) H represents the hidden layer matrices of the spatiotemporal convolutional layers at layers (l+1) and (l), respectively. (l) The terms constitute residual connections to prevent deep network degradation, where m is the regional spatial index as described in this claim, representing three regions (mapped to a self-loop region, a nearest neighbor region, and a far neighbor region, respectively), M m To obtain a learnable mask matrix, the Hadamard product is used. By performing element-wise weighting with the spatiotemporal convolution result, adaptive fine-tuning of the weights for different neighborhoods can be achieved. Z is the spatiotemporal adjacency matrix of the m-th region after decoupling. t The input spatiotemporal feature tensor represents the temporal sequence information after one-dimensional temporal convolution. Let be the learnable transformation matrix set for the m-th region in the l-th layer, and ReLU(·) be a linear activation function used to introduce nonlinear feature transformation; the spatiotemporal adjacency matrix of the ST-GNN prediction module is constructed based on the global adjacency matrix and the fully connected matrices of adjacent time steps, and is expressed as:

[0031] ;

[0032] Among them, A st This represents the spatiotemporal adjacency matrix, where u is the number of samples in adjacent time steps. Let be the global spatial adjacency matrix of the u-th sample, located on the diagonal, used to describe the static spatial topological relationships between different channels at the same sampling time. It is a P-order all-one matrix, where P is the number of channels as described in claim 5, i.e., a fully connected time step projection matrix, which establishes the temporal correlation between nodes between adjacent time steps, and simultaneously captures spatial and temporal dependencies.

[0033] As a further improvement of the present invention, the joint loss function in step S4 is the sum of the MSE loss and the classification auxiliary loss:

[0034] ;

[0035] in The regression loss is the MSE of the gangue content. The loss is categorized by coal discharge state, and λ is the balance coefficient.

[0036] As a further improvement of the present invention, the model training and evaluation in step S5 includes: initializing network parameters, inputting training samples into the model, calculating the joint loss and optimizing the model network parameters, and after each training round, evaluating the model based on MSE, MAE, and R... 2 The model's performance is evaluated using metrics, and the optimal model parameters are selected. The test set is then input into the trained model to assess its performance.

[0037] Secondly, the present invention also proposes a rock content prediction system for fully mechanized mining faces based on vibration signal information fusion, applied to the aforementioned method for predicting rock content in fully mechanized mining faces based on vibration signal information fusion, comprising:

[0038] The signal acquisition module is used to acquire the vibration signal of the hydraulic support tail beam impacted by the collapse of coal gangue;

[0039] The graph structure construction module is used to perform time-domain segmentation, DFT, cosine similarity calculation and k-NNG sparsification on the vibration signal to construct graph structure data.

[0040] The feature fusion module is used to input the graph structure data into a multi-channel fusion model based on GNN to generate fusion features based on the topological graph.

[0041] The prediction output module is used to input the fused features into the ST-GNN prediction module and output the predicted value of the gangue content.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. This invention introduces graph signal processing and graph neural network system into the prediction of gangue content in the working face of a fully mechanized amplification machine, realizing the end-to-end processing of vibration signals from non-Euclidean structured expression to spatiotemporal joint modeling, which can effectively characterize the strong non-stationary and nonlinear characteristics of vibration signals.

[0044] 2. This invention effectively captures the spatial topological relationships and complementary information between multiple sensors through k-NNG graph construction and dynamic attention mechanism between channels, overcoming the limitations of single sensor signals and simple multi-channel fusion methods, and improving robustness under complex working conditions.

[0045] 3. This invention realizes an end-to-end spatiotemporal map prediction framework, directly outputting continuous gangue content values, solving the problem that existing methods can only perform binary classification of coal and gangue or discrete state classification, and providing accurate and continuous quantitative basis for intelligent coal discharge.

[0046] 4. The computational complexity of this invention is controllable, making it suitable for deployment of downhole edge computing devices. It supports real-time prediction and closed-loop control, and the method has good scalability, allowing for further fusion of multimodal data such as acoustic signals and images. Attached Figure Description

[0047] Figure 1 This is a flowchart of the overall method of the present invention.

[0048] Figure 2 This is a schematic diagram illustrating the process of representing the single-channel vibration signal of the present invention.

[0049] Figure 3 This is a diagram illustrating the architecture of the gangue content prediction module based on multi-channel vibration signal information fusion according to the present invention.

[0050] Figure 4 This is a diagram showing the layout of the data acquisition device on the test bench.

[0051] Figure 5 This is a diagram to verify the effect of the model in the example.

[0052] Figure 6 The example shows the prediction results based on data collected on the experimental platform. Detailed Implementation

[0053] The technical solution of this application will be further described in detail below with reference to specific embodiments.

[0054] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0055] Reference Figure 4 As shown, a scaled-down test bench for hydraulic supports was constructed to simulate the actual top coal caving process. Specifically, a coal and gangue sliding test bench was built at a 1:5 scale, referencing the ZF8200 / 17 / 35 hydraulic support. The data acquisition devices on the test bench are arranged as follows. Figure 4 As shown, a 1A110E piezoelectric accelerometer was selected as the sensing input. The sensor was installed on the same component of the hydraulic support and the direction vectors were parallel to each other in space. The sensor's sensing input data was recorded using a MI-7008D portable data acquisition and analysis instrument with a sampling rate of 25600Hz.

[0056] Six types of experimental data were collected: pure coal, 20% mixed gangue, 40% mixed gangue, 60% mixed gangue, 80% mixed gangue, and pure gangue (mass fraction). The data was divided into a training set (80% of the data) and a test set (20% of the data).

[0057] Reference Figure 2 The diagram shows the graphical representation process of a single-channel vibration signal, constructing graphical data for each channel's vibration signal. The specific calculation and processing flow is as follows:

[0058] With a fixed sampling frequency f s The vibration response generated by the collapse impact of coal and gangue on the test bench was continuously collected. The original vibration signal x collected can be regarded as a one-dimensional time series of length N, where n is the sampling point index of the original signal, and is expressed as:

[0059] ;

[0060] After resampling the sequence M times, the original vibration signal x can be regenerated as a signal of magnitude M. The time-domain signal dataset X can be represented as:

[0061] ;

[0062] Let i represent the index of the sample signal, then the content of the i-th data segment can be represented as:

[0063] ;

[0064] Time window length T w The choice of time window requires a trade-off between frequency resolution and frequency range. Longer segments can achieve higher frequency resolution, but may obscure the temporal location of individual impact events; shorter segments have the opposite effect. Considering the characteristics of coal gangue impact signals and to obtain a sufficient number of samples, a time window length of 50ms is chosen. Let the overlap rate of the resampled vibration signal samples be δ; in this embodiment, the sample overlap rate is set to 0.75.

[0065] After resampling the vibration signal of a single channel, each data segment x (M) This can be considered as a node. In this case, the size of the graph is determined by the number of samples (i.e., the number of resampling times) M. The discrete time-acceleration time-series signal is converted into a frequency-amplitude frequency domain signal using the Discrete Fourier Transform (DFT). For data segment x... (i) The method for calculating the discrete Fourier transform is as follows:

[0066] ;

[0067] In this formula, j is the imaginary unit, and k is the frequency component index, representing the k-th frequency component. Let x be the complex transform result of the i-th time-domain segment on the k-th frequency component, where n is the sampling point index and x is the frequency component. n Let L be the sampled value of the input discrete-time signal, corresponding to the nth sample point of the i-th time segment, and L be the length of the signal segment, i.e., the total number of sample points participating in the DFT calculation. This formula uses a rotation factor. The amplitude features of the signal at different frequency components are extracted. After the time-series signal is converted into frequency domain signal features through Discrete Fourier Transform, the positive and negative frequencies of different amplitudes are conjugate in the complex plane. Therefore, when extracting features for each sample, only the positive frequency components need to be retained to reduce computation. The frequency components are:

[0068] ;

[0069] Wherein, any frequency f in the frequency component f k It can be calculated using the following formula:

[0070] ;

[0071] The frequency domain characteristics corresponding to each sample can be represented by the amplitude of each frequency component f in that sample as follows:

[0072] ;

[0073] Obtain each data segment x (i) Frequency domain amplitude characteristics v (i) Then, the node feature matrix F is obtained, which is expressed as:

[0074] ;

[0075] The weight of the edge between any two nodes i1 and i2 is defined by cosine similarity. To avoid self-loops interfering with message passing, the weight of self-connections between nodes in the graph is set to 0, and the elements in the weight matrix W... The calculation method is as follows:

[0076] ;

[0077] The fully connected weight matrix W consumes a large amount of storage space and computational cost. Therefore, the k-NNG method is used to prune the edges of each node: for each node i, the edges with the k most similar neighboring nodes are retained, and the weights of the remaining edges are set to 0, resulting in a sparse adjacency matrix A, which is calculated as follows:

[0078] ;

[0079] Finally, by repeating the above calculation process for all samples in all channels, we can obtain the feature matrix set of P channel nodes. and the set of adjacency matrices corresponding to each channel. P is the number of channels, and a is the channel index;

[0080] Reference Figure 3 The diagram shows the architecture of a gangue content prediction module based on multi-channel vibration signal information fusion, and a gangue content prediction model based on GNN multi-channel vibration signal information fusion is built.

[0081] First, the training data feature matrix F obtained in step 2 is input into the initial convolutional layer for feature extraction, resulting in F', which is calculated as follows:

[0082] ;

[0083] Among them, W (1) b represents the weight matrix of the initial convolutional layer. (1) Z represents the bias term vector of the initial convolutional layer. (1) The intermediate feature matrix after preliminary feature extraction.

[0084] ;

[0085] Among them, W (2) b represents the weight matrix of the second convolutional layer. (2) This represents the bias term vector of the second convolutional layer, ReLU(·) is the linear activation function, μ is the feature mean of the current batch, and σ 2 Let ε be the feature variance, γ be the numerical stability term to prevent the denominator from being zero, γ and β be the learnable scaling and translation factors, F' be the output feature matrix after normalization and activation, and the node feature dimension after output is L'.

[0086] Construct the adjacency matrix A of the global graph global First, we need to calculate the diagonal block elements in each channel. The calculation method is as follows:

[0087] ;

[0088] W represents the weight matrix between the a-th channel and the b-th channel. ab It can be represented as:

[0089] ;

[0090] Where a and b are channel indices, s ab The average feature cosine similarity is calculated as follows:

[0091] ;

[0092] The average feature vector of the channels is calculated as follows:

[0093] ;

[0094] like ( If the threshold is used, then connection weights are added at the corresponding positions in the global adjacency matrix. In summary, the global adjacency matrix can be constructed as follows:

[0095] ;

[0096] Secondly, multiple layers of GCN need to be set up to transmit information. First, establish the initial information. And concatenate the node features of all channels:

[0097] ;

[0098] A multi-channel GCN layer, where l is the layer number and is set to 2, with message passing between each layer, can be represented as:

[0099] ;

[0100] in, This is a degree matrix, representing the degree of each node. To construct a normalized Laplace operator, The aggregation of learnable parameter weight matrices through any node v of a certain channel P:

[0101] ;

[0102] Where, m v Let v be the message vector obtained by aggregating the neighborhood of node v, where v is the target node whose feature needs to be updated. u Let v be a neighboring node of node v, and v u The neighborhood set of node v d v d u They are node v and node v respectively. u The degree. For the neighbor node v in layer l u The feature vector of node v, and the feature update method of layer l+1 is as follows:

[0103] ;

[0104] in, Let v be the updated feature vector of node v at layer l+1. Let be the feature vector of node v at layer l. To avoid information overload caused by simple concatenation, channel attention can be introduced after each layer. The pooling operation at each layer is as follows:

[0105] ;

[0106] in, This represents the time-domain average feature representation of channel a at layer l. Channel attention coefficients are also introduced. The calculation method is as follows:

[0107] ;

[0108] Where q is a learnable query vector used to calculate the attention weights for channel a. Let be the learnable weight matrix for the attention module of channel a, tanh(·) be the hyperbolic tangent activation function, and softmax(·) be the normalization function. Then, update the weighted feature matrix of node v based on this.

[0109] ;

[0110] Read the sample topology feature representation for each channel The expression using global average pooling is as follows:

[0111] ;

[0112] in, The features representing a specific level of node v are denoted here as the final layer. The final fusion employs multi-head attention cross-channel fusion, where c is the attention head index and h is the total number of attention heads. Each attention branch completes a QKV linear transformation through learnable Query weights, Key weights, and Value weights, satisfying the following calculation method:

[0113] ;

[0114] ;

[0115] ;

[0116] Where, q c Let c be the query vector of the attention head, representing the global feature map. Let g be the query learnable weight matrix for the c-th attention head. avg k is the global average eigenvector. a,c Let be the key vector corresponding to the a-th channel in the c-th attention head. Let g be the key-learnable weight matrix for the c-th attention head. av is the node feature vector of the a-th channel after processing by the GCN layer; a,c Let a be the value vector corresponding to the a-th channel in the c-th attention head. The weight matrix can be learned for the c-th attention head. The calculation method for each attention head is as follows:

[0117] ;

[0118] Among them, o c Let d be the output vector of the c-th attention head, representing the weighted fusion features extracted by that attention head. c The scaling factor is the value of the vector k. a,c The feature dimensions are adjusted to prevent excessively large dot products from causing gradient vanishing. Finally, all attention heads are concatenated to obtain the fused features of all channels based on the topological graph.

[0119] ;

[0120] Among them, F fused For the final calculated multi-channel fused feature matrix, Concat(·) represents the concatenation operation that joins the outputs of all attention heads into a single long vector, W. o This is a projection matrix that maps the stitched high-dimensional features back to the target dimension.

[0121] The spatiotemporal convolutional layer model treats u sampled neighboring nodes as temporal sequences of length, fuses these nodes, and models them as a spatiotemporal graph. :

[0122] ;

[0123] in, yes A set of nodes, a spacetime graph The initial characteristics are:

[0124] ;

[0125] To ensure temporal smoothness, the temporal convolutional layer is calculated as follows:

[0126] ;

[0127] in, The output temporal convolution feature vector represents the temporal sequence information after one-dimensional temporal convolution processing. Conv1D(·,kernel=3) represents a one-dimensional temporal convolution operation with a kernel size of 3.

[0128] Each convolutional domain of a spatial convolutional layer needs to capture three different types of spatial dependencies: from the neighborhood, near neighbors, and far neighbors. It is a set of edges, and the spatial edges will reuse A. global The time side uses a matrix of all 1s. Connecting adjacent nodes, the spatiotemporal adjacency matrix A is obtained. st Represented as:

[0129] ;

[0130] Among them, A st This represents the spatiotemporal adjacency matrix, where u is the number of samples in adjacent time steps. Let be the global spatial adjacency matrix of the u-th sample, located on the diagonal, used to describe the static spatial topological relationships between different channels at the same sampling time. It is a P-order all-one matrix, where P is the number of channels, i.e., the fully connected time step projection matrix, which establishes the temporal correlation between nodes between adjacent time steps, and captures both spatial and temporal dependencies.

[0131] The spatial convolutional layer is divided into three regions: the self-loop region, the nearest neighbor region, and the far neighbor region. Its calculation method is as follows:

[0132] ;

[0133] Among them, H (l+1) and H (l) H represents the hidden layer matrices of the spatiotemporal convolutional layers at layers (l+1) and (l), respectively. (l) The terms form residual connections to prevent deep network degradation, where m is the regional spatial index, representing three regions (mapped to the self-loop region, the nearest neighbor region, and the far neighbor region, respectively), M m To obtain a learnable mask matrix, the Hadamard product is used. By performing element-wise weighting with the spatiotemporal convolution result, adaptive fine-tuning of the weights for different neighborhoods can be achieved. Z is the spatiotemporal adjacency matrix of the m-th region after decoupling. t The input spatiotemporal feature tensor represents the temporal sequence information after one-dimensional temporal convolution. Let be the learnable transformation matrix set for the m-th region in the l-th layer. ReLU(·) is a linear activation function used to introduce non-linear feature transformation. The computational results, after temporal smoothing and spatial aggregation, are then used for the final regression module calculation after pooling. The global pooling layer is calculated as follows:

[0134] ;

[0135] Where P is the number of channels and u is the number of adjacent samples. The final output feature matrix is ​​obtained by performing MLP regression on the previous global pooling result g to obtain the final predicted value of the gangue content. The calculation method is as follows:

[0136] ;

[0137] in, This is the weight matrix of the regression layer. Let σ be the bias vector of the regression layer, and σ(·) be the activation function, which compresses the output value to the (0,1) interval, conforming to the physical definition of gangue content (percentage). Finally, the model's predicted gangue content is obtained. .

[0138] The model training parameters are shown in Table 1.

[0139] Table 1 Model training parameters

[0140]

[0141] The model is optimized using a joint loss function:

[0142] ;

[0143] in The regression loss is the amount of gangue. The loss is categorized by coal discharge state, and λ is the balance coefficient.

[0144] Model training and experimental results (refer to) Figure 5 and Figure 6 Model training process: Initialize network parameters, input training samples into the model, calculate joint loss, and optimize model network parameters. After each training round, based on MSE, MAE, and R... 2 The effectiveness of the model is evaluated using various metrics, and the model parameters with the best performance are compared and selected.

[0145] Ablation experiments were conducted on each module of the method model proposed in this embodiment, and the results are shown in Table 2.

[0146] Table 2 Ablation Experiment Results

[0147]

[0148] As shown in Table 2, the results demonstrate the effectiveness of each module of the proposed method model in extracting and fusing information features of data from each channel, and show good results.

[0149] Reference Figure 5 , Figure 5 The experimental results for the classification-aided task are shown in the figure. Figure 5 In the figures (a) and (b), the trends of accuracy and loss function for binary classification (using only pure coal and pure gangue samples) of the model are shown, respectively. It is evident that the model performs well in binary classification tasks. Figure 5 (c) and (d) are confusion matrices of the six-class classification performance of the model. With a gangue content of less than 25% classified as positive and a gangue content of greater than 25% classified as negative, the model achieves a gangue identification accuracy of 0.9453 in the full-class classification task. This further verifies that the method of this invention can capture the vibration signal topology data features of the samples and the features related to the gangue content, thereby improving the performance of gangue identification and gangue content prediction.

[0150] Reference Figure 6 , Figure 6 This is a scatter plot of the prediction results obtained from data collected on the experimental platform according to an embodiment of the present invention. Fifteen prediction results were randomly selected from each type of sample, clearly demonstrating the close intra-class relationships and significant inter-class distinctions among the samples.

[0151] Finally, the model performance was evaluated based on the test set data, and the evaluation results are shown in Table 3.

[0152] Table 3 Model Evaluation Indicators

[0153]

[0154] Based on the comprehensive evaluation of the model obtained from Table 3, the model performs well in the prediction task with MSE=0.033, MAE=0.025<0.03, and R0. 2 =0.89 > 0.88, which meets the requirements for real-time control in engineering. Based on the above results, this invention demonstrates good performance in the field of predicting gangue content.

[0155] This invention provides a method for predicting the gangue content in fully mechanized longwall mining faces based on vibration signal information fusion. This method can be applied to real-time monitoring of gangue content and intelligent coal release control during fully mechanized longwall mining, demonstrating clear industrial applicability. It achieves high-precision continuous prediction end-to-end, with controllable computational complexity, making it suitable for deployment with underground edge computing equipment and providing reliable technical support for intelligent coal mining.

[0156] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for predicting the gangue content in a fully mechanized mining face based on vibration signal information fusion, characterized in that, Includes the following steps: S1. Obtain the vibration signal of the hydraulic support impacted by the collapse of coal gangue, establish the total dataset and divide it into training dataset and test dataset; S2. Perform time-domain segmentation and Discrete Fourier Transform (DFT) on the sampled data to extract frequency domain amplitude features. Based on the frequency domain amplitude features, calculate the edge weights between nodes of each sampled data using cosine similarity. Then, construct a topological graph by sparsification using the k-NNG method to establish a graph structure dataset. S3. Construct a multi-channel vibration signal information fusion model based on graph neural network (GNN) to predict the gangue content. Input the graph structure data into the model to generate fusion features based on the topological graph. Then input the fusion features into the spatiotemporal graph neural network (ST-GNN) prediction module to output the predicted gangue content value. S4. Optimize the model using a joint loss function; S5. Model Training and Evaluation.

2. The method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion according to claim 1, characterized in that, The time-domain segmentation in step S2 includes: using a fixed sampling frequency f s Vibration signals are collected according to a preset time window length T. w The original signal is segmented, with an overlap rate δ set between adjacent segments. Each segment is treated as a time-domain segment and represented as follows. ; The DFT extraction of frequency domain amplitude features includes: performing a DFT on each time domain segment: ； Where j is the imaginary unit as described in this claim, and k is the frequency component index as described in this claim, representing the k-th frequency component. Let x be the complex transform result of the i-th time domain segment on the k-th frequency component, n be the sampling point index as described in this claim, and x be the sampling point index as described in this claim. n Let L be the sampled value of the input discrete-time signal, corresponding to the nth sample point of the i-th time segment, and L be the length of the signal segment, i.e., the total number of sample points participating in the DFT calculation. This formula uses a rotation factor. Extract the amplitude features of the signal at different frequency components; retain the amplitude of the positive frequency component as the frequency domain feature vector corresponding to that segment. The node feature matrix is ​​obtained. , where M is the number of time-domain segments.

3. The method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion according to claim 1, characterized in that, Step S2, which involves calculating the edge weights between nodes of each sampled data using cosine similarity, includes: calculating the frequency domain feature vectors v corresponding to any two nodes. i With v j The edge weights between nodes with cosine similarity are represented as i1 and i2 are node indices, calculated as follows: ； The method of constructing a sparse topology graph using the k-NNG method includes: for each node, retaining only the edges between the k neighboring nodes with the highest cosine similarity to that node, setting the edge weights between that node and other nodes to zero, and obtaining a sparse adjacency matrix A; the method of establishing graph structure data includes: repeating the above calculation process for all samples of all channels to obtain the feature matrix set of each channel node. and the set of sparse adjacency matrices corresponding to each channel node. P is the number of channels, and a is the channel index as described in this claim.

4. The method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion according to claim 1, characterized in that, The gangue content prediction model based on GNN multi-channel vibration signal information fusion in step S3 includes a GNN multi-channel information fusion module, an ST-GNN prediction module, and an MLP regression output module; the GNN multi-channel information fusion module includes an MLP dimensionality reduction module, a global adjacency matrix construction module, a multi-layer GCN message passing module, a channel attention module, and a multi-head attention fusion module.

5. The method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion according to claim 4, characterized in that, Based on the sparse adjacency matrix corresponding to each channel node Construct a collection of diagonal block elements P is the number of channels, a is the channel index as described in this claim, and each diagonal block element satisfies , where I M Given an M-order identity matrix, channel connection weights are constructed based on the cosine similarity between the average eigenvectors of each channel. When the channel similarity... , When the threshold is reached, connection weights are added at the corresponding positions to obtain the global adjacency matrix A. global Its structural form is as follows: ； Among them, W ab Let represent the weight matrix between the a-th channel and the b-th channel, where a and b are the channel indices as described in this claim, and the weight matrix satisfies ... , , and These are represented as the average feature vectors of the a-th and b-th channels, respectively. In the multi-layer GCN message passing module, each layer of GCN aggregates the feature information of adjacent nodes according to the global adjacency matrix. The aggregation range includes adjacent nodes within the same channel and adjacent nodes between different channels.

6. The method for predicting the gangue content in a fully mechanized mining face based on vibration signal information fusion according to claim 4, characterized in that, The channel attention module calculates the attention weights for each channel after each layer of GCN, and multiplies the attention weights by the channel features to achieve adaptive weighting of the channel features; the multi-head attention fusion module uses multiple attention heads to perform linear transformations on the query, key, and value respectively, and the calculation method for each attention head is as follows: ； Where c is the attention head index as described in this claim, h is the total number of attention heads, and o c Let be the output vector of the c-th attention head, representing the weighted fusion feature extracted by that attention head, where 'a' is the channel index as described in claim 5, and 'softmax' is the output vector of the c-th attention head. a (·) is the normalization function for calculating the a-th channel in the c-th attention head. q c Let c be the query vector of the attention head, representing the global feature map. Let g be the query learnable weight matrix for the c-th attention head. avg The global average feature vector, obtained by performing global average pooling on the features of all channels, is used to guide attention allocation. k a,c Let be the key vector corresponding to the a-th channel in the c-th attention head. Let g be the key-learnable weight matrix for the c-th attention head. a Let be the node feature vector of the a-th channel after processing by the GCN layer. v a,c Let a be the value vector corresponding to the a-th channel in the c-th attention head. The weight matrix can be learned for the value of the c-th attention head, d c The scaling factor is the value of the vector k. a,c To minimize gradient vanishing due to excessively large dot products, the outputs of each attention head are concatenated to obtain the fused features of all channels based on the topological graph. ； Among them, F fused For the final calculated multi-channel fused feature matrix, Concat(·) represents the concatenation operation that joins the outputs of all attention heads into a single long vector, W. o This is a projection matrix that maps the stitched high-dimensional features back to the target dimension.

7. The method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion according to claim 4, characterized in that, The ST-GNN prediction module includes a one-dimensional temporal convolutional layer (1D-Conv), a spatial partitioning convolutional layer, and a global pooling layer. The spatial partitioning convolutional layer divides the spatial neighborhood into three regions: a self-looping region, a near neighboring region, and a far neighboring region. Its calculation method is as follows: ； Where H (l+1) and H (l) H represents the hidden layer matrices of the spatiotemporal convolutional layers at layers (l+1) and (l), respectively. (l) The terms constitute residual connections to prevent deep network degradation, where m is the regional spatial index as described in this claim, representing three regions respectively, mapped to a self-loop region, a nearest neighbor region, and a far neighbor region, M. m To obtain a learnable mask matrix, the Hadamard product is used. By performing element-wise weighting with the spatiotemporal convolution result, adaptive fine-tuning of the weights for different neighborhoods can be achieved. Z is the spatiotemporal adjacency matrix of the m-th region after decoupling. t The input spatiotemporal feature tensor represents the temporal sequence information after one-dimensional temporal convolution. Let be the learnable transformation matrix set for the m-th region in the l-th layer, and ReLU(·) be the linear activation function used to introduce non-linear feature transformation; The spatiotemporal adjacency matrix of the ST-GNN prediction module is constructed based on the global adjacency matrix and the fully connected matrix of adjacent time steps, and is expressed as follows: ； Among them, A st This represents the spatiotemporal adjacency matrix, where u is the number of samples in adjacent time steps. Let be the global spatial adjacency matrix of the u-th sample, located on the diagonal, used to describe the static spatial topological relationships between different channels at the same sampling time. It is a P-order all-one matrix, where P is the number of channels as described in claim 5, i.e., a fully connected time step projection matrix, which establishes the temporal correlation between nodes between adjacent time steps, and simultaneously captures spatial and temporal dependencies.

8. The method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion according to claim 1, characterized in that, The joint loss function mentioned in step S4 is the sum of the MSE loss and the classification auxiliary loss: ； in The regression loss is the MSE of the gangue content. The loss is categorized by coal discharge state, and λ is the balance coefficient.

9. The method for predicting the gangue content of a fully mechanized mining face based on vibration signal information fusion according to claim 1, characterized in that, The model training and evaluation described in step S5 includes: initializing network parameters, inputting training samples into the model, calculating joint loss and optimizing model network parameters, evaluating model performance based on MSE, MAE, and R² metrics after each training round, selecting the optimal model parameters, and inputting the test set into the trained model to evaluate performance.

10. A system for predicting the gangue content of a fully mechanized longwall mining face based on vibration signal information fusion, applied to the method for predicting the gangue content of a fully mechanized longwall mining face based on vibration signal information fusion as described in any one of claims 1-9, characterized in that, include: The signal acquisition module is used to acquire the vibration signal of the hydraulic support tail beam impacted by the collapse of coal gangue; The graph structure construction module is used to perform time-domain segmentation, DFT, cosine similarity calculation and k-NNG sparsification on the vibration signal to construct graph structure data. The feature fusion module is used to input the graph structure data into a multi-channel fusion model based on GNN to generate fusion features based on the topological graph. The prediction output module is used to input the fused features into the ST-GNN prediction module and output the predicted value of the gangue content.

Images