CN-121980392-A - Intelligent recognition method for coal gangue vibration signals based on double branches Mamba and contrast learning

CN121980392ACN 121980392 ACN121980392 ACN 121980392ACN-121980392-A

Abstract

The invention provides an intelligent recognition method of coal gangue vibration signals based on double branches Mamba and contrast learning, which is characterized in that coal gangue impact vibration signals at a tail beam of a hydraulic support are collected and converted into time-frequency spectrograms, global features of the signals are extracted through a double-attention feature extraction model, decoupling of different branch features is achieved through JS divergence loss, the decoupled features are input into a double Mamba modeling model, fine-granularity local sequence features are extracted, the compactness of similar features and the separability of different types of features are constrained by a supervised NT-Xent contrast loss function, and the features are fused through a classifier module comprising a residual convolution and a multi-layer perceptron and recognition results of coal or gangue are output. The method solves the problems that the prior method ignores similarity and heterogeneity between classes, can realize feature decoupling and accurate classification without complex external data enhancement, has extremely high recognition precision and robustness in complex underground high-noise environments, and is suitable for intelligent control of fully mechanized caving face.

Inventors

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

Assignees

中国矿业大学

Dates

Publication Date: 20260505
Application Date: 20260401

Claims (10)

1. The intelligent recognition method for the coal gangue vibration signals based on double branches Mamba and contrast learning is characterized by comprising the following steps: step 1, arranging a vibration acceleration sensor on a tail beam of a hydraulic support of an experiment table, acquiring an original one-dimensional vibration signal generated in the process of sliding coal and gangue and impacting the tail beam by using the vibration acceleration sensor, performing characteristic interception on the vibration signal, converting the vibration signal into two-dimensional time-frequency spectrogram characteristics, performing data enhancement and normalization processing to obtain preprocessed input characteristics, and designing a label according to the category of the gangue; step 2, inputting the preprocessed features into a symmetrical double-attention feature extraction model to obtain a first hidden feature representation And a second hidden feature representation ; Step 3, establishing a branch difference loss function based on JS divergence, performing branch difference loss calculation on the two characteristic representations obtained in the step 2, and prompting the two branches to learn the characteristic differentiated on the global distribution by minimizing the branch difference loss function so as to realize decoupling of the coal-gangue sharing mixed characteristic; step 4, inputting the characteristics after decoupling in the step 3 into a double Mamba modeling model established based on Mamba networks, and outputting first branch characteristics And a second branching feature ; Step 5, a contrast loss function is established based on NT-Xent loss, in the double Mamba modeling model described in step 4, a positive sample pair and a negative sample pair are established by using supervised data with real labels, and contrast loss is calculated; step 6, utilizing a residual convolution module and a multi-layer perceptron mechanism to output two paths of characteristics of the double Mamba modeling model after constraint in the step 5 And (3) with Splicing and fusing are carried out on the channel dimension, interaction information between two branches is captured through a residual convolution module, nonlinear mapping dimension reduction is carried out through a multi-layer perceptron, and classification prediction probability is output Calculating cross entropy loss by using the predicted value and the true value to be used as classification loss; Step 7, taking the classification loss in the step 6, the branch difference loss in the step 3 and the comparison loss weighted summation in the step 5 as a total objective function, and performing iterative training on the whole model by using an Adam optimizer to obtain a trained coal gangue vibration signal intelligent recognition model; and 8, collecting real-time vibration signals in the actual fully-mechanized caving face, inputting the real-time vibration signals into the intelligent recognition model of the coal gangue vibration signals after the real-time vibration signals are processed in the step 1, and outputting the classification result that the current sliding substances are coal or gangue.
2. The intelligent recognition method for the coal gangue vibration signals based on the double branches Mamba and the contrast learning is characterized in that in the step 1, in the process of converting an original one-dimensional vibration signal into a two-dimensional time-frequency spectrogram, the number of short-time Fourier transform points and the window length are all set to 63, data enhancement operation on the time-frequency spectrogram comprises time masking, frequency masking and random translation, normalization operation is carried out on each data, each time-frequency spectrogram sample obtained through pretreatment is labeled according to the corresponding real category when the sample is collected, wherein the coal sample is endowed with a coal tag, the gangue sample is endowed with a gangue tag, and each sample corresponds to a unique supervision tag.
3. The intelligent recognition method for coal gangue vibration signals based on double branches Mamba and contrast learning according to claim 1 is characterized in that the specific method in the step 2 is as follows: The preprocessing features obtained in the step 1 are input into a branch 1, the branch 1 maps the input features through a first embedded mapping layer, a first group of query matrix, key matrix and value matrix are utilized to execute multi-head self-attention operation, and features in a time dimension and a frequency dimension are globally integrated to obtain a first hidden feature representation Inputting the preprocessed features obtained in the step 1 into a branch 2, mapping the input features by the branch 2 through a second embedded mapping layer, executing multi-head self-attention operation by utilizing a second group of query matrix, key matrix and value matrix, and performing global integration on the features in the time dimension and the frequency dimension to obtain a second hidden feature representation 。
4. The intelligent recognition method of coal gangue vibration signals based on double branches Mamba and contrast learning according to claim 3, wherein in step 2, the preprocessed input features are recorded as It is of dimension size Wherein, The number of frequency components is represented and, Representing a time step, in which the input features are calculated Expressed in bulk as a three-dimensional tensor , wherein, Representing a batch size; will input features After input to branch 1, first to Performing dimension transformation, and performing linear mapping, local convolution mapping and position coding superposition on the transformed features through an embedding mapping layer corresponding to the branch 1 to obtain the embedded features of the branch 1 Reusing the query projection matrix corresponding to branch 1 Key projection matrix Sum projection matrix Mapping the embedded features to the first Query matrix for individual attention heads Key matrix Sum matrix The calculation form is as follows: ; based on the query matrix, key matrix and value matrix, the first in branch 1 The output of the individual attention heads is calculated as: ; Wherein, the For single head attention feature dimension, all of branches 1 are taken The outputs of the individual attention heads are spliced and projected through the output projection matrix Performing linear transformation to obtain first hidden characteristic representation ; To the same input feature After input to branch 2, the same form of dimension transformation and embedding mapping operation as that of branch 1 is adopted to obtain the embedding characteristics of branch 2 Reusing the query projection matrix corresponding to branch 2 Key projection matrix Sum projection matrix Respectively generate the first Query matrix for individual attention heads Key matrix Sum matrix The calculation form is as follows: ; In branch 2 The output of the individual attention heads is calculated as: ; All of the branches 2 The outputs of the individual attention heads are spliced and projected through the output projection matrix Performing linear transformation to obtain second hidden characteristic representation ; Wherein, the branch 1 and the branch 2 respectively adopt mutually independent embedded mapping parameters, inquiry projection parameters, key projection parameters and value projection parameters, and are characterized in the same input Learning under conditions to obtain different and complementary representations of hidden features And 。
5. The intelligent recognition method for coal gangue vibration signals based on double branches Mamba and contrast learning according to claim 4, wherein in step 3, for a sample And Mapping to probability distribution via Softmax activation function And The branch difference loss is constructed based on the inverse form of Jensen-Shannon divergence: ; ; Wherein, the , Indicating Kullback-Leibler divergence, Representing Jensen-Shannon divergence, symbol' "Relative divergence marker" is used to indicate the degree of difference between the former probability distribution and the latter probability distribution; smoothing preset values to prevent zero removal errors by minimizing The characteristic distribution of the dual-attention branch output is caused to diverge from each other.
6. The intelligent recognition method for coal gangue vibration signals based on double branches Mamba and contrast learning according to claim 5, wherein the specific method in step 4 is as follows: establishing a double Mamba modeling model based on Mamba networks, wherein the double Mamba modeling model comprises a branch 1 and a branch 2, decoupling the first hidden feature in the step 3 The representation is input into the branch 1, the branch 1 utilizes the corresponding independent Mamba modeling unit to perform state space modeling on the first hidden feature representation, extracts a first fine-granularity local dynamic feature, and outputs a first branch feature And (3) decoupling the second hidden feature representation in step 3 Input into branch 2, branch 2 uses its corresponding independent Mamba modeling unit to model the state space of the second hidden feature representation, extract the second fine-grained local dynamic feature, and output the second branch feature 。
7. The intelligent recognition method of coal gangue vibration signals based on double branches Mamba and contrast learning according to claim 6, wherein in step 4, the two branches of the double Mamba modeling model respectively receive the first hidden characteristic representation after decoupling in step 3 And a second hidden feature representation Wherein branch 1 represents the first hidden feature Modeling, branch 2 represents the second hidden feature Modeling is carried out; Representing the first hidden feature Input to branch 1, and branch 1 maps the input features through the first linear mapping layer to obtain principal components And gating signal The calculation form is as follows: ; Wherein, the And A learnable parameter matrix corresponding to the branch 1; Principal component Inputting the data into a one-dimensional convolution unit and obtaining updated characteristics through SiLU activation function processing And then to Performing linear projection to obtain input dependent parameters of the selective state space model And And combines the step length mapping function corresponding to the branch 1 And a learnable parameter Calculating a first discrete step size : ; The first discrete step size For discretizing continuous state space parameters and characterizing by means of a selective State Space Model (SSM) Dynamic modeling to obtain state space output Outputting the state space And gating signal after SiLU activation Performing element-by-element multiplication to obtain gating fusion characteristics, and performing linear transformation through an output mapping layer to obtain first branch output characteristics ; Representing the second hidden feature Input to branch 2, branch 2 adopts the same form of processing flow as branch 1, and obtains principal component through the second linear mapping layer And gating signal The calculation form is as follows: ; Wherein, the And A learnable parameter matrix corresponding to the branch 2; Principal component Inputting the data into a one-dimensional convolution unit and obtaining updated characteristics through SiLU activation function processing And then the updated characteristics are subjected to Performing linear projection to obtain input dependent parameters of the selective state space model And And combines the step length mapping function corresponding to the branch 2 And a learnable parameter Calculating a second discrete step size : ; The second discrete step size For discretizing continuous state space parameters and characterizing by selective state space models Dynamic modeling is carried out to obtain state space output Outputting the state space And gating signal after SiLU activation Performing element-by-element multiplication to obtain a gating fusion characteristic, and performing linear transformation through an output mapping layer corresponding to the branch 2 to obtain a second branch output characteristic ; Wherein, the branch 1 and the branch 2 are the same in network structure, and the linear mapping parameter, the convolution parameter, the gating parameter, the state space model parameter and the output mapping parameter are independent of each other, so that two paths of hidden features can be respectively and dynamically modeled, and different fine granularity features can be output And 。
8. The intelligent recognition method of coal gangue vibration signals based on double branches Mamba and contrast learning according to claim 7, wherein in step 5, the contrast loss function is based on a first branch characteristic output by a double Mamba modeling model And a second branching feature Build, set in one batch of batch size B, from the first branch feature The first to be taken out The individual sample feature vectors are noted as From a second branch feature The first to be taken out The individual sample feature vectors are noted as The true category labels are marked as Wherein, the method comprises the steps of, Representation and the first The samples have a sample index set of the same label, Indicating the current batch divided by the first All sample index sets except the individual samples themselves; based on the definition, the supervised NT-Xent contrast loss for branch 1 is calculated as: ; The supervised NT-Xent contrast loss for branch 2 was calculated as: ; Total contrast loss is defined as: ; Wherein, the Representation and the first Positive sample indexes of which the samples belong to the same class; representing candidate contrast sample indices; Representing positive sample sets The number of samples contained in the sample; Representing a similarity function for measuring the degree of similarity between two feature vectors; representing temperature parameters for adjusting the degree of smoothness of the similarity distribution by minimizing the total contrast loss The similar sample features are promoted to gather and the heterogeneous sample features are separated from each other in the feature space.
9. The intelligent recognition method of coal gangue vibration signals based on double branches Mamba and contrast learning according to claim 8, wherein in step 6, the specific processing flow of the classifier module is as follows: Constraining post-step 5 features And (3) with Channel-level stitching to obtain fusion features Will (i) be Sequential input contains 、 The convolution kernel and residual convolution module connected with the residual are subjected to batch normalization (Batchnorm) and SiLU activation function processing to obtain multi-scale fusion characteristics Will (i) be Input into a multi-layer perceptron (MLP) containing only a single hidden layer and output probabilities for classification via linear mapping When (when) Judging the sample class as coal; and judging the sample type as gangue.
10. The intelligent recognition method of coal gangue vibration signals based on double branches Mamba and contrast learning according to claim 9, wherein in step 7, the cross entropy loss calculation form is as follows: ; Wherein, the For the one-hot encoded component of the sample genuine label, For the prediction probability, K represents the total number of categories, ; The weighted total objective function is: ; Wherein, the 、、 Is a loss weight coefficient.

Description

Intelligent recognition method for coal gangue vibration signals based on double branches Mamba and contrast learning Technical Field The invention belongs to the field of intelligent mining of coal mines, and particularly relates to an intelligent recognition method for coal gangue vibration signals based on double branches Mamba and contrast learning. Background Coal is used as one of core energy sources and is widely applied to the fields of industry, power generation, metallurgy, civil use and the like for a long time. Along with the acceleration of the carbon reduction process and the improvement of the environmental protection requirement, the coal industry is urgently required to realize energy conservation, consumption reduction and green exploitation while guaranteeing the production energy. Fully-mechanized caving mining (LTCC) is widely used at home and abroad as a main technology for mining thick coal seams because of its capability of effectively improving coal recovery rate. However, in the top coal caving link of fully mechanized mining, currently, workers are mainly relied on to operate a coal caving port through experience, and gangue is extremely easy to mix. The gangue can not be directly used as fuel as solid waste, and can also obviously reduce coal quality, thereby greatly increasing washing and selecting cost and energy consumption in later stage. Therefore, the accurate, efficient and automatic identification of coal and gangue is realized in the whole top coal caving process, and the method has become a key technical problem to be solved in intelligent mine construction. Aiming at the challenge, domestic and foreign scholars propose a plurality of coal gangue identification methods based on different sensing modes, and the method mainly comprises image identification, ray detection, vibration signal detection and the like. The image recognition method relies on the differences of colors, textures and shapes of coal and gangue, but the underground environment of the deep coal mine is usually filled with high-concentration dust, and serious problems of insufficient illumination and shielding exist, so that the image data contains a large amount of noise, and the reliability of visual detection is greatly reduced. Although the ray detection technology can effectively overcome the interference of environmental dust, the practical application of the ray detection technology in the pit is limited by high equipment cost and potential radiation hazard to operators, and is difficult to popularize on a large scale. In contrast, the identification method based on the vibration signal has been one of the most potential technical directions in recent years due to the advantages of high reliability, low cost and environmental friendliness. Because of different physical and mechanical properties of coal and gangue, the coal and gangue can generate different vibration modes when striking the tail beam of the hydraulic support. The traditional vibration signal processing method mostly depends on fixed mathematical transformation and artificial feature design, generally only global features are extracted, and local or specific frequency band information is ignored, so that the differentiation degree of different categories in a feature space is weaker. With the development of deep learning technology, neural networks exhibit superior automatic feature extraction capability in vibration signal processing. However, most existing deep learning recognition models are optimized mainly using cross entropy loss, and this paradigm, while maximizing classification accuracy, lacks explicit constraints on feature spatial distribution. In actual working conditions, the vibration signals of the coal and the gangue show the dual characteristics of extremely similar signals of the same kind and weak signal differences of different kinds. Neglecting the complex local time-frequency structural characteristics can cause the problem that the characteristics are easily overlapped and the boundary is fuzzy when the model faces to underground environmental noise and operation condition changes, and seriously weakens the generalization capability of the model. In order to solve the problem of uneven feature distribution, part of researches introduce a contrast learning strategy, and feature space is optimized by pulling similar samples and pushing different samples. However, the application of the conventional contrast learning method in coal gangue vibration signal identification has obvious limitations that on one hand, the conventional contrast learning is excessively dependent on an external data enhancement technology to construct positive and negative sample pairs, so that extra environmental noise and uncertainty are easily introduced in vibration signal processing to destroy the original physical semantics of signals, and on the other hand, most of the conventional models adopt a single feature extraction path, so