Coal rock cutting state identification method based on semi-supervised adversarial discrimination network

Abstract

The invention relates to a coal rock cutting state identification method based on a semi-supervised adversarial discrimination network, and belongs to the technical field of coal rock cutting state identification. Comprising the following steps: acquiring historical vibration data of a cutting head when the heading machine historically cuts coal rock; preprocessing the historical vibration data, and taking the preprocessed historical vibration data as a training sample set; establishing a coal rock cutting state identification semi-supervised adversarial discrimination network, and training the coal rock cutting state identification semi-supervised adversarial discrimination network by using the labeled sample set and the unlabeled sample set to obtain a trained coal rock cutting state identification semi-supervised adversarial discrimination network; and acquiring current vibration data of a cutting head when the heading machine currently cuts the coal rock, inputting the current vibration data into the trained coal rock cutting state identification semi-supervised adversarial discrimination network, and determining the state of the current coal rock cut by the heading machine according to the output of the current vibration data. The coal rock cutting state recognition method is high in coal rock cutting state recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of coal-rock cutting state recognition, and in particular to a coal-rock cutting state recognition method based on a semi-supervised adversarial discriminant network. Background Art

[0002] During the coal mining process, coal-rock cutting status recognition technology can help workers accurately judge the current cutting status of various equipment, thereby effectively adjusting the working parameters of the equipment to protect the equipment and improve mining efficiency.

[0003] Coal-rock identification currently uses image processing technology to identify the visual differences between coal and rock for the identification of coal-rock cutting status. Although accurate identification can be achieved through continuous improvement of the algorithm, the method is only applicable to ideal environmental conditions. Lighting conditions, dust, and the dust-reducing water mist of the tunnel boring machine itself will cause great difficulties in image acquisition and identification, and even make it difficult to identify; sensor vibration signals, current signals and other time series signals are more suitable for industrial scenarios, with the advantages of high real-time performance, low cost and high reliability. With the rapid development of artificial intelligence (AI) technology, deep learning methods are becoming more and more popular due to their ability to adaptively learn features. Deep learning models have a multi-layer structure, which helps to represent hierarchical features, thereby achieving more accurate coal-rock identification effects. However, the practical application of these methods also faces a huge challenge, that is, a large number of labeled samples are required to train deep learning models for coal-rock identification. In terms of tunneling face mining equipment, obtaining a sufficient number of cutting feature samples is often a difficult task. Many coal-field enterprises have accumulated limited coal-rock cutting feature data. Even if sensors collect a large amount of data, manually labeling this data is a resource-intensive task. Therefore, effective use of unlabeled actual data is crucial to improving the performance of training models. Semi-supervised learning can use unannotated samples to better utilize the underlying structure of the data manifold. However, when dealing with recognition tasks with extremely limited annotated samples, it is difficult for deep learning models to achieve effective decision boundaries. At this time, the effectiveness of semi-supervised learning may deteriorate, resulting in low accuracy in coal and rock state identification using the network model obtained through semi-supervised learning. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network. The technical solution of the present invention is as follows:

[0005] A method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network, comprising:

[0006] S1, obtaining historical vibration data of the cutting head of the roadheader when cutting coal and rock;

[0007] S2, preprocessing the historical vibration data, and using the preprocessed historical vibration data as a training sample set, wherein the training sample set includes a labeled sample set consisting of a limited number of labeled samples and an unlabeled sample set consisting of a large number of unlabeled samples;

[0008] S3, establishing a semi-supervised adversarial discriminant network for coal-rock cutting state recognition, and using a labeled sample set and an unlabeled sample set to train the semi-supervised adversarial discriminant network for coal-rock cutting state recognition, to obtain a trained semi-supervised adversarial discriminant network for coal-rock cutting state recognition;

[0009] S4, obtain the current vibration data of the cutting head of the roadheader when it is currently cutting coal and rock, and input the current vibration data into the trained coal and rock cutting state recognition semi-supervised adversarial discriminant network, and determine the current state of the coal and rock cut by the roadheader according to the output of the trained coal and rock cutting state recognition semi-supervised adversarial discriminant network.

[0010] Optionally, the coal-rock cutting state recognition semi-supervised adversarial discriminant network includes a feature extractor, a recognition classifier and a discriminator; S3 includes:

[0011] S31, obtaining initial parameters configured for the feature extractor, recognition classifier and discriminator;

[0012] S32, training the feature extractor and the recognition classifier by configuring the feature extractor with initial parameters, the recognition classifier with initial parameters, and the labeled sample set to obtain updated parameters of the feature extractor and the recognition classifier;

[0013] S33, performing semi-supervised adversarial training of the feature extractor and the discriminator by configuring the updated parameters of the feature extractor, the labeled sample set, the unlabeled sample set, and the discriminator with the initial parameters, to obtain the updated parameters of the feature extractor and the updated parameters of the discriminator;

[0014] S34, continue to train the feature extractor, recognition classifier and discriminator through the updated parameters of the recognition classifier, the updated parameters of the feature extractor and the updated parameters of the discriminator, until the preset convergence conditions of the feature extractor, recognition classifier and discriminator are reached, and a trained coal rock cutting state recognition semi-supervised adversarial discriminant network is obtained.

[0015] Optionally, the S32 includes:

[0016] S321, extracting features of each labeled sample in the labeled sample set by using a feature extractor configured with initial parameters;

[0017] S322, inputting the features of each labeled sample into a recognition classifier configured with initial parameters, and after the recognition classifier configured with initial parameters outputs a classification result, calculating a classification loss value of the recognition classifier according to the classification result and the label of each labeled sample;

[0018] S323, constructing a triplet of a labeled sample set, wherein the triplet includes an anchor sample, a positive sample, and a negative sample;

[0019] S324, extracting features of the anchor sample, features of the positive sample, and features of the negative sample by using a feature extractor configured with initial parameters;

[0020] S325, calculating a triplet loss value according to the features of the anchor sample, the features of the positive sample, and the features of the negative sample;

[0021] S326, determining the overall loss function value of labeled supervised learning according to the classification loss value and the triplet loss value;

[0022] S327, back-propagating the overall loss function value of the labeled supervised learning to update the parameters of the feature extractor and the parameters of the recognition classifier, and obtaining the updated parameters of the feature extractor and the updated parameters of the recognition classifier.

[0023] Optionally, the S33 includes:

[0024] S331, extracting features of each labeled sample in the labeled sample set and features of each unlabeled sample in the unlabeled sample set by configuring a feature extractor with updated parameters;

[0025] S332, inputting the features of each labeled sample and each unlabeled sample into a discriminator configured with initial parameters, the discriminator configured with initial parameters discriminates the source of the features, and calculating the adversarial loss value of the discriminator and the adversarial loss value of the feature extractor according to the output of the discriminator configured with initial parameters;

[0026] S333, update the parameters of the discriminator according to the adversarial loss value of the discriminator, and update the parameters of the feature extractor according to the loss value of the feature extractor, and obtain the updated parameters of the discriminator and the updated parameters of the feature extractor respectively. The loss value of the feature extractor is determined according to the adversarial loss value of the feature extractor and the overall loss function value of labeled supervised learning.

[0027] Optionally, the preset convergence condition of the feature extractor is that the loss value of the feature extractor is maximized, the preset convergence condition of the recognition classifier is that the classification loss value of the classifier is minimized, and the preset convergence condition of the discriminator is that the adversarial loss value of the discriminator is minimized.

[0028] Optionally, when calculating the classification loss value of the recognition classifier according to the classification result and the label of each labeled sample, S322 is implemented by formula (1):

[0029]

[0030] In formula (1), Represents the classification loss value of the recognition classifier, Represents the labeled sample set X l The i-th labeled sample in , express The label, F represents the feature extractor, N c represents the number of defined cut states that need to be identified, express features, and C represents the recognition classifier.

[0031] Optionally, when calculating the triplet loss value according to the features of the anchor sample, the features of the positive sample and the features of the negative sample, S325 is implemented by formula (2):

[0032]

[0033] In formula (2), represents the triple loss value, Represents the labeled sample set X l The i-th labeled sample in , express The label, F represents the feature extractor, and N represents the labeled sample set X l The number of triplets in , represents the anchor sample, represents the characteristics of the anchor sample, represents a positive sample, represents the characteristics of the positive sample, represents negative samples, represents the characteristics of negative samples, and α represents the margin.

[0034] Optionally, when calculating the adversarial loss value of the discriminator and the adversarial loss value of the feature extractor according to the output of the discriminator configured with the initial parameters, S332 is implemented by formula (3) and formula (4) respectively:

[0035]

[0036]

[0037] In formula (3) and formula (4), represents the adversarial loss value of the discriminator, Represents the labeled sample set X lThe i-th labeled sample in , Represents the unlabeled sample set X u The i-th unlabeled sample in, E represents expectation, ~ represents belonging, D represents discriminator, F represents feature extractor, F(x i ) represents the sample x extracted by the feature extractor i Features, Represents the adversarial loss value of the feature extractor.

[0038] Optionally, in S333, the loss value of the feature extractor is determined by formula (5) according to the adversarial loss value of the feature extractor and the overall loss function value of labeled supervised learning:

[0039]

[0040]

[0041] In formula (5) and formula (6), represents the loss value of the feature extractor, represents the overall loss function value of labeled supervised learning, β is the parameter for weighing the classification loss and triple loss, Represents the classification loss value of the recognition classifier, Represents the triplet loss value.

[0042] All the above optional technical solutions can be combined arbitrarily, and the present invention does not provide detailed descriptions of the structures after the combinations.

[0043] By means of the above scheme, the beneficial effects of the present invention are as follows:

[0044] By using labeled sample sets and unlabeled sample sets to train the semi-supervised adversarial discriminant network for coal-rock cutting state recognition, the additional information in the unlabeled samples is utilized when training the semi-supervised adversarial discriminant network for coal-rock cutting state recognition, which can enhance the robustness of the network and improve its generalization ability for unseen data. Through adversarial training, the semi-supervised adversarial discriminant network for coal-rock cutting state recognition can effectively capture the underlying data distribution and obtain a more general feature representation, thereby enhancing the generalization ability of the network and improving the recognition performance, so that the accuracy of coal-rock cutting state recognition through the trained semi-supervised adversarial discriminant network is high.

[0045] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1It is a flow chart of a method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0048] The embodiment of the present invention provides a method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network, which can be implemented by any electronic device with computing function, such as a PC, a mobile terminal, a server, etc. Figure 1 As shown, the coal-rock cutting state recognition method based on the semi-supervised adversarial discriminant network provided by the embodiment of the present invention includes the following steps S1 to S4:

[0049] S1, obtaining historical vibration data of the cutting head of the roadheader when cutting coal and rock.

[0050] The historical vibration data includes the vibration data collected under various cutting conditions of coal and rock. The historical vibration data is collected by the vibration sensor installed on the cutting head of the tunnel boring machine and then sent to the electronic equipment.

[0051] S2, preprocessing the historical vibration data, and using the preprocessed historical vibration data as a training sample set, wherein the training sample set includes a labeled sample set consisting of a limited number of labeled samples and an unlabeled sample set consisting of a large number of unlabeled samples.

[0052] The preprocessing here may include filtering, outlier processing, etc. The specific content and implementation of the preprocessing are not elaborated in detail in the embodiment of the present invention. The training sample set is data for training a semi-supervised adversarial discriminant network for identifying coal-rock cutting states.

[0053] Specifically, suppose a small number of labeled samples from different coal-rock cutting states are recorded as labeled sample set X l and It is a monitoring sample of the tunnel boring machine. is the corresponding cut state, i.e., label. In addition to the limited labeled samples, a large amount of monitoring data in the preprocessed historical vibration data is unlabeled, recorded as the unlabeled sample set

[0054] S3, establish a semi-supervised adversarial discriminant network for coal-rock cutting state recognition, and use labeled sample sets and unlabeled sample sets to train the semi-supervised adversarial discriminant network for coal-rock cutting state recognition to obtain a trained semi-supervised adversarial discriminant network for coal-rock cutting state recognition.

[0055] The method provided in the embodiment of the present invention aims to comprehensively utilize Xl and X u To train the semi-supervised adversarial discriminant network for coal-rock cutting state recognition. The semi-supervised adversarial discriminant network for coal-rock cutting state recognition consists of a multi-layer feature extractor F(·,θ f ) and the recognition classifier C(·,θ c ) and is trained to facilitate the recognition of coal-rock cutting states. f and θ c are the parameters of the feature extractor and recognition classifier, which need to be optimized during training. Through the trained coal-rock cutting state recognition semi-supervised adversarial discriminant network, the input sample x can be obtained i The final diagnostic decision, namely C(F(x i )).

[0056] In order to use the u In the embodiment of the present invention, a discriminator is introduced during training. l and X u To conduct adversarial learning between them, so as to improve the feature extractor F(·,θ f ) is regularized to capture generalized feature representations. At the same time, metric learning is applied to enhance the compactness of samples in the same cut state and the separability of samples in different categories of cut states in the deep feature space. l With the supervision of , more discriminative features can be learned in this step.

[0057] Specifically, in the method provided in the embodiment of the present invention, adversarial learning is used to capture the generalized feature representation by utilizing the coal-rock cutting state in the labeled sample set and the unlabeled sample set. Taking binary classification (coal and rock) as an example, F(X l ) and F(X u ) represent the distribution of labeled samples and unlabeled samples respectively. Before adversarial learning, due to the lack of sufficient cut-off state information, supervised training with extremely limited labeled samples may not obtain an ideal decision boundary. The decision boundary learned from limited labeled samples may not fully capture the feature distribution of a large number of unlabeled samples, that is, F(X u ), resulting in unsatisfactory recognition performance. To solve this problem, the present invention constructs a binary discriminator D(·,θ d ).θ d are the network parameters of the discriminator. By regularizing the training of F, adversarial learning is performed between D and F. Given a sample x i , the cutting state feature can be extracted by F, denoted as F(x i ). In the adversarial learning process, the purpose of D is to distinguish the labeled samples from the set X l Or unlabeled sample set X uThe source of the features extracted in is then updated to generate more general features to confuse D.

[0058] As shown in Table 1, it is a schematic diagram of the composition structure of the semi-supervised adversarial discrimination network for coal-rock cutting state recognition established in an embodiment of the present invention.

[0059] Table 1

[0060]

[0061] To ensure efficient feature extraction, the feature extractor contains five convolutional layers and pooling layers (from Conv_1&Pooling_1 to Conv_5&Pooling_5). In order to avoid over-labeling when there are limited labeled samples, the recognition classifier adopts a simple structure with fewer trainable parameters, that is, it includes two fully-connected layers, and the same is true for the discriminator.

[0062] In combination with the above content, in a specific embodiment, S3 includes:

[0063] S31, obtaining initial parameters configured for the feature extractor, recognition classifier and discriminator.

[0064] Among them, the initial parameters of the feature extractor, recognition classifier and discriminator configuration can be set manually based on experience.

[0065] S32, training the feature extractor and the recognition classifier by configuring the feature extractor with initial parameters, configuring the recognition classifier with initial parameters, and training the feature extractor and the recognition classifier with the labeled sample set to obtain updated parameters of the feature extractor and updated parameters of the recognition classifier.

[0066] Exemplarily, the S32 specifically includes:

[0067] S321, extracting features of each labeled sample in the labeled sample set by using a feature extractor configured with initial parameters.

[0068] S322, input the features of each labeled sample into the recognition classifier configured with initial parameters, and after the recognition classifier configured with initial parameters outputs the classification result, the classification loss value of the recognition classifier is calculated according to the classification result and the label of each labeled sample.

[0069] Specifically, when calculating the classification loss value of the recognition classifier according to the classification result and the label of each labeled sample, S322 can be implemented by formula (1):

[0070]

[0071] In formula (1), Represents the classification loss value of the recognition classifier, Represents the labeled sample set X l The i-th labeled sample in , express The label, F represents the feature extractor, N c represents the number of defined cut states that need to be identified, express features, and C represents the recognition classifier.

[0072] S323, constructing a triplet of labeled sample sets, wherein the triplet includes an anchor sample, a positive sample, and a negative sample.

[0073] Specifically, although semi-supervised training can utilize feature information from a large number of unlabeled samples, it may have a negative impact when the feature distribution boundaries between different cut states are not clear. For example, the feature extractor may capture shared but useless features in labeled coal samples and unlabeled rock samples, and successfully confuse the discriminator in adversarial training. To solve this problem, the embodiment of the present invention introduces a discriminative feature enhancement method based on metric learning during training to improve the performance of semi-supervised training.

[0074] The basic idea of ​​discriminative feature enhancement based on metric learning is to minimize the difference between samples with the same cutting state while increasing the distance between samples of different categories. In this way, more distinguishing features of different cutting states can be captured, and the trained coal-rock cutting state recognition semi-supervised adversarial discriminant network can be made more discriminative. In specific implementation, the embodiment of the present invention constructs a triplet of labeled sample sets, which contains anchor samples from specific cutting states. Positive samples from the same cut state and negative samples from different cut states Among them, i, j and k represent different points, a is the abbreviation of anchor point, and p and n are the abbreviations of positive and negative samples respectively.

[0075] Specifically, when constructing the triplet of the labeled sample set, each labeled sample in the labeled sample set can be refined and labeled according to the cutting state.

[0076] S324, extracting features of anchor samples, features of positive samples, and features of negative samples through a feature extractor configured with initial parameters.

[0077] That is, when extracting the features of each labeled sample in S321 and extracting the features of the anchor sample, the features of the positive sample, and the features of the negative sample in step 324, the parameters of the feature extractor are the same.

[0078] S325, calculating the triplet loss value according to the features of the anchor sample, the features of the positive sample, and the features of the negative sample.

[0079] In a specific embodiment, when calculating the triplet loss value according to the features of the anchor sample, the features of the positive sample, and the features of the negative sample, S325 is implemented by formula (2):

[0080]

[0081] In formula (2), represents the triple loss value, Represents the labeled sample set X l The i-th labeled sample in , express The label, F represents the feature extractor, and N represents the labeled sample set X l The number of triplets in , represents the anchor sample, represents the characteristics of the anchor sample, represents a positive sample, represents the characteristics of the positive sample, represents negative samples, represents the characteristics of negative samples, and α represents the margin.

[0082] Optimizing the triplet loss enables learning more discriminative features, thereby ensuring clearer decision boundaries for further classification and diagnosis. Considering the manifold assumption, the embedded features of labeled and unlabeled samples with the same cut state are distributed in the same manifold. When the triplet loss is optimized for labeled samples, the learned features are more discriminative and the manifold distributions associated with different cut states are more separable.

[0083] Due to the insufficient information of coal-rock cutting state contained in the limited labeled samples, overfitting may occur when using triplet loss optimization alone. Moreover, when adversarial learning is used alone, the negative alignment of feature distributions between labeled samples and unlabeled samples may reduce the performance of semi-supervised learning. For semi-supervised adversarial learning, it is necessary to separate manifolds of different cutting states. Therefore, by comprehensively considering the two optimization objectives, the training method proposed in the embodiment of the present invention can effectively enhance the diagnostic performance of the semi-supervised adversarial discriminant network for coal-rock cutting state recognition.

[0084] S326, determining the overall loss function value of labeled supervised learning according to the classification loss value and the triplet loss value.

[0085] Specifically, the overall loss function value of labeled supervised learning is It can be expressed as formula (6):

[0086]

[0087] In formula (6), β is a parameter that balances classification loss and triplet loss. is the classification loss value, is the triplet loss value.

[0088] S327, back-propagating the overall loss function value of the labeled supervised learning to update the parameters of the feature extractor and the parameters of the recognition classifier, and obtaining the updated parameters of the feature extractor and the updated parameters of the recognition classifier.

[0089] S33, performing semi-supervised adversarial training on the feature extractor and the discriminator by configuring the updated parameters of the feature extractor, the labeled sample set, the unlabeled sample set and the discriminator with the initial parameters, to obtain the updated parameters of the feature extractor and the updated parameters of the discriminator.

[0090] Specifically, adversarial learning plays a vital role in achieving marginal distribution alignment between labeled and unlabeled samples. In scenarios where labeled and unlabeled samples are obtained from the same cut state, it is usually assumed that their distributions exhibit similarities. However, the limited size of labeled samples may lead to a deviation from the empirical distribution, resulting in a difference from the true distribution. This phenomenon is called empirical distribution mismatch, which highlights the difference between the distributions of labeled and unlabeled samples. In this case, it becomes meaningful to exploit adversarial learning of labeled and unlabeled samples. This approach aims to mitigate the difference between the empirical distribution and the true distribution by adjusting their respective distributions. By leveraging the additional information present in unlabeled samples, the robustness of the model can be enhanced and its generalization ability to unseen data can be improved. Through adversarial training, the semi-supervised adversarial discriminant network for coal rock cut state recognition can effectively capture the underlying data distribution, thereby enhancing its generalization ability.

[0091] Exemplarily, the S33 specifically includes:

[0092] S331 , extracting features of each labeled sample in the labeled sample set and features of each unlabeled sample in the unlabeled sample set through a feature extractor configured with updated parameters.

[0093] S332, input the features of each labeled sample and the features of each unlabeled sample into the discriminator configured with initial parameters, and the discriminator configured with initial parameters determines the source of the features (from labeled samples or unlabeled samples), and calculates the adversarial loss value of the discriminator and the adversarial loss value of the feature extractor based on the output of the discriminator configured with initial parameters.

[0094] In a specific embodiment, when calculating the adversarial loss value of the discriminator and the adversarial loss value of the feature extractor according to the output of the discriminator configured with initial parameters, S332 is implemented by formula (3) and formula (4) respectively:

[0095]

[0096]

[0097] In formula (3) and formula (4), represents the adversarial loss value of the discriminator, Represents the labeled sample set X l The i-th labeled sample in , Represents the unlabeled sample set X u The i-th unlabeled sample in, E represents expectation, ~ represents belonging, D represents discriminator, F represents feature extractor, F(x i ) represents the sample x extracted by the feature extractor i Features, Represents the adversarial loss value of the feature extractor.

[0098] S333, update the parameters of the discriminator according to the adversarial loss value of the discriminator, and update the parameters of the feature extractor according to the loss value of the feature extractor, and obtain the updated parameters of the discriminator and the updated parameters of the feature extractor respectively. The loss value of the feature extractor is determined according to the adversarial loss value of the feature extractor and the overall loss function value of labeled supervised learning.

[0099] In a specific embodiment, in S333, the loss value of the feature extractor is determined by formula (5) according to the adversarial loss value of the feature extractor and the overall loss function value of the labeled supervised learning:

[0100]

[0101] In formula (5) and formula (6), represents the loss value of the feature extractor, represents the overall loss function value of labeled supervised learning, β is the parameter for weighing the classification loss and triple loss, Represents the classification loss value of the recognition classifier, Represents the triplet loss value.

[0102] S34, continue to train the feature extractor, recognition classifier and discriminator through the updated parameters of the recognition classifier, the updated parameters of the feature extractor and the updated parameters of the discriminator, until the preset convergence conditions of the feature extractor, recognition classifier and discriminator are reached, and a trained coal rock cutting state recognition semi-supervised adversarial discriminant network is obtained.

[0103] In specific implementation, when the feature extractor, recognition classifier and discriminator are continued to be trained through the updated parameters of the recognition classifier, the updated parameters of the feature extractor and the updated parameters of the discriminator, the training method is the same as the principle in S32 and S33, and only the respective parameters are changed, which will not be elaborated in detail in the embodiment of the present invention.

[0104] In a specific embodiment, the preset convergence condition of the feature extractor is that the loss value of the feature extractor is maximized, the preset convergence condition of the recognition classifier is that the classification loss value of the classifier is minimized, and the preset convergence condition of the discriminator is that the adversarial loss value of the discriminator is minimized.

[0105] In other words, the optimization goal of the recognition classifier C is: The optimization goal of the discriminator D is: The optimization goal of the feature extractor F is: max F

[0106] In summary, the semi-supervised adversarial discriminant network for coal-rock cutting state recognition provided by the embodiment of the present invention allows full use of the coal-rock cutting state information contained in unlabeled samples during the training process, and learns more general feature representations and improves recognition performance through adversarial training. By combining metric learning based on triplet loss, the proposed semi-supervised adversarial learning framework enhances feature clustering within the same category and improves feature separation between different categories, effectively assisting the learning process, thereby producing excellent recognition performance, especially when the availability of labeled samples is limited, the recognition performance is even better.

[0107] S4, obtain the current vibration data of the cutting head of the roadheader when it is currently cutting coal and rock, and input the current vibration data into the trained coal and rock cutting state recognition semi-supervised adversarial discriminant network, and determine the current state of the coal and rock cut by the roadheader according to the output of the trained coal and rock cutting state recognition semi-supervised adversarial discriminant network.

[0108] Specifically, the trained coal-rock cutting state recognition semi-supervised adversarial discriminant network outputs the probabilities of various cutting states, and the embodiment of the present invention selects the cutting state with the highest probability as the current coal-rock state cut by the roadheader.

[0109] In summary, the embodiments of the present invention provide a new method for identifying coal and rock cutting states when labeled samples are limited. The proposed semi-supervised adversarial discriminant network for identifying coal and rock cutting states combines the coal and rock cutting state information contained in unlabeled samples, and learns more general feature representations and improves recognition performance through adversarial training. In addition, the semi-supervised adversarial discriminant network for identifying coal and rock cutting states uses triplet loss optimization to improve the discrimination of features of different cutting categories, strengthens feature clustering within the same category, and improves feature separation between different categories, effectively assisting the learning process, thereby producing excellent recognition performance, especially when the availability of labeled samples is limited. By analyzing the current vibration signal of the cutting head of the tunnel boring machine collected on site, the proposed method can realize coal and rock recognition with only limited labeled samples, and has strong robustness and generalization capabilities.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network, characterized in that: include: S1, obtaining historical vibration data of the cutting head of the roadheader when cutting coal and rock; S2, preprocessing the historical vibration data, and using the preprocessed historical vibration data as a training sample set, wherein the training sample set includes a labeled sample set consisting of a limited number of labeled samples and an unlabeled sample set consisting of a large number of unlabeled samples; S3, establishing a semi-supervised adversarial discriminant network for coal-rock cutting state recognition, and using a labeled sample set and an unlabeled sample set to train the semi-supervised adversarial discriminant network for coal-rock cutting state recognition, to obtain a trained semi-supervised adversarial discriminant network for coal-rock cutting state recognition; S4, obtain the current vibration data of the cutting head of the roadheader when it is currently cutting coal and rock, and input the current vibration data into the trained coal and rock cutting state recognition semi-supervised adversarial discriminant network, and determine the current state of the coal and rock cut by the roadheader according to the output of the trained coal and rock cutting state recognition semi-supervised adversarial discriminant network.

2. The method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network according to claim 1 is characterized in that: The coal-rock cutting state recognition semi-supervised adversarial discriminant network includes a feature extractor, a recognition classifier and a discriminator; S3 includes: S31, obtaining initial parameters configured for the feature extractor, recognition classifier and discriminator; S32, training the feature extractor and the recognition classifier by configuring the feature extractor with initial parameters, the recognition classifier with initial parameters, and the labeled sample set to obtain updated parameters of the feature extractor and the recognition classifier; S33, performing semi-supervised adversarial training of the feature extractor and the discriminator by configuring the updated parameters of the feature extractor, the labeled sample set, the unlabeled sample set, and the discriminator with the initial parameters, to obtain the updated parameters of the feature extractor and the updated parameters of the discriminator; S34, continue to train the feature extractor, recognition classifier and discriminator through the updated parameters of the recognition classifier, the updated parameters of the feature extractor and the updated parameters of the discriminator, until the preset convergence conditions of the feature extractor, recognition classifier and discriminator are reached, and a trained coal rock cutting state recognition semi-supervised adversarial discriminant network is obtained.

3. The method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network according to claim 2 is characterized in that: The S32 includes: S321, extracting features of each labeled sample in the labeled sample set by using a feature extractor configured with initial parameters; S322, inputting the features of each labeled sample into a recognition classifier configured with initial parameters, and after the recognition classifier configured with initial parameters outputs a classification result, calculating a classification loss value of the recognition classifier according to the classification result and the label of each labeled sample; S323, constructing a triplet of a labeled sample set, wherein the triplet includes an anchor sample, a positive sample, and a negative sample; S324, extracting features of the anchor sample, features of the positive sample, and features of the negative sample by using a feature extractor configured with initial parameters; S325, calculating a triplet loss value according to the features of the anchor sample, the features of the positive sample, and the features of the negative sample; S326, determining the overall loss function value of labeled supervised learning according to the classification loss value and the triplet loss value; S327, back-propagating the overall loss function value of the labeled supervised learning to update the parameters of the feature extractor and the parameters of the recognition classifier, and obtaining the updated parameters of the feature extractor and the updated parameters of the recognition classifier.

4. The method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network according to claim 2 or 3, characterized in that: The S33 includes: S331, extracting features of each labeled sample in the labeled sample set and features of each unlabeled sample in the unlabeled sample set by configuring a feature extractor with updated parameters; S332, inputting the features of each labeled sample and each unlabeled sample into a discriminator configured with initial parameters, the discriminator configured with initial parameters discriminates the source of the features, and calculating the adversarial loss value of the discriminator and the adversarial loss value of the feature extractor according to the output of the discriminator configured with initial parameters; S333, update the parameters of the discriminator according to the adversarial loss value of the discriminator, and update the parameters of the feature extractor according to the loss value of the feature extractor, and obtain the updated parameters of the discriminator and the updated parameters of the feature extractor respectively. The loss value of the feature extractor is determined according to the adversarial loss value of the feature extractor and the overall loss function value of labeled supervised learning.

5. The method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network according to claim 4 is characterized in that: The preset convergence condition of the feature extractor is that the loss value of the feature extractor is the maximum, the preset convergence condition of the recognition classifier is that the classification loss value of the classifier is the minimum, and the preset convergence condition of the discriminator is that the adversarial loss value of the discriminator is the minimum.

6. The method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network according to claim 3 is characterized in that: When calculating the classification loss value of the recognition classifier according to the classification result and the label of each labeled sample, S322 is implemented by formula (1): In formula (1), Represents the classification loss value of the recognition classifier, Represents the labeled sample set X l The i-th labeled sample in , express The label, F represents the feature extractor, N c represents the number of defined cut states that need to be identified, express features, and C represents the recognition classifier.

7. The method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network according to claim 3 or 6, characterized in that: When calculating the triplet loss value according to the features of the anchor sample, the features of the positive sample and the features of the negative sample, S325 is implemented by formula (2): In formula (2), represents the triple loss value, Represents the labeled sample set X l The i-th labeled sample in , express The label, F represents the feature extractor, and N represents the labeled sample set X l The number of triplets in , represents the anchor sample, represents the characteristics of the anchor sample, represents a positive sample, represents the characteristics of the positive sample, represents negative samples, represents the characteristics of negative samples, and α represents the margin.

8. The method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network according to claim 4 is characterized in that: When calculating the adversarial loss value of the discriminator and the adversarial loss value of the feature extractor according to the output of the discriminator configured with the initial parameters, S332 is implemented by formula (3) and formula (4) respectively: In formula (3) and formula (4), represents the adversarial loss value of the discriminator, Represents the labeled sample set X l The i-th labeled sample in , Represents the unlabeled sample set X u The i-th unlabeled sample in, E represents expectation, ~ represents belonging, D represents discriminator, F represents feature extractor, F(x i ) represents the sample x extracted by the feature extractor i Features, Represents the adversarial loss value of the feature extractor.

9. The method for identifying coal-rock cutting state based on a semi-supervised adversarial discriminant network according to claim 8, characterized in that: In S333, the loss value of the feature extractor is determined by formula (5) according to the adversarial loss value of the feature extractor and the overall loss function value of the labeled supervised learning: In formula (5) and formula (6), represents the loss value of the feature extractor, represents the overall loss function value of labeled supervised learning, β is the parameter for weighing the classification loss and triple loss, Represents the classification loss value of the recognition classifier, Represents the triplet loss value.

Images