TBM tunnel face lithology identification method based on self-supervised learning module

Abstract

The invention provides a TBM tunnel face lithology identification method based on a self-supervised learning module. The method comprises the following steps: constructing a self-supervised learning model, wherein the self-supervised learning model comprises an encoder, a projection head and a classifier; constructing an auxiliary data set by using the lithology-related public image data set, and training a self-supervised learning model by using the auxiliary data set to obtain a trained self-supervised learning model; and inputting to-be-identified rock slag image data into the trained self-supervised learning model, performing tunnel face lithology identification on the to-be-identified rock slag image by the trained self-supervised learning model by adopting a SimCLR algorithm, and judging the lithology of the to-be-identified rock slag image. According to the method, a self-supervised learning method is adopted, downstream generalization is carried out through a small part of images with lithologic labels, and lithologic classification can be carried out on TBM rock slag. Dependence of the model on label data volume can be slowed down, and lithology is identified and classified in the initial stage of tunnel construction.

Description

Technical Field

[0001] The present invention relates to the technical field of TBM (Tunnel Boring Machine) tunnel excavation, and in particular to a TBM tunnel face lithology identification method based on a self-supervised learning module. Background Art

[0002] TBMs, with their advantages of safe and efficient excavation, have become the preferred choice for long and large tunnel construction. Compared with drilling and blasting, TBMs have the automation and information technology to predict or identify rock mass information. Currently, long-distance hard rock tunnel construction often faces problems such as sudden changes in geological conditions and delayed rock type identification. Utilizing real-time TBM excavation data to accurately identify rock mass information at the tunnel face is of positive significance for safe and efficient TBM excavation. Lithology is a characteristic property of rock mass. Different lithologies have significant differences in mineral composition and structural composition. Studies have shown that lithology is correlated with indicators such as brittleness index and abrasiveness index. TBM rock breaking efficiency and tool wear vary significantly under different lithologies, directly affecting excavation efficiency and construction costs. Therefore, lithology is a key factor affecting the excavation and construction progress of TBMs in long-distance hard rock tunnels.

[0003] In practical applications, TBM tunnel faces are often obscured by shields, making it impossible to identify the face's lithology in real time. TBM rock debris image recognition generally uses supervised learning models, which rely on a large number of reliable labeled samples. However, the lack of labeled data in the early stages of new construction projects severely limits the model's generalization capabilities.

[0004] One existing method for identifying lithology at a TBM tunnel face involves training using data collected by mechanical sensors during tunneling (e.g., gear seal pressure, TBM inclination parameters, belt conveyor speed, and cutterhead torque and thrust). However, this data has limited relevance to the face lithology. Using supervised learning methods after tunnel penetration requires extensive labeled data for model training.

[0005] A TBM tunnel face lithology identification method in the above-mentioned prior art includes:

[0006] TBM mechanical parameters have no direct correlation with face lithology. Their values ​​are directly affected by TBM size (diameter, power, etc.). Different TBM sizes with the same lithology will have different mechanical parameters. Rock slag images are a direct product of face rock breaking. Their data labels are unique and reliable, unaffected by TBM size.

[0007] The supervised learning algorithm used in existing solutions requires sufficient data as a training basis. Often, after sufficient data has been collected, the tunnel project is close to completion, and its application value is limited. Summary of the Invention

[0008] The embodiment of the present invention provides a TBM tunnel face lithology identification method based on a self-supervised learning module, so as to effectively improve the excavation efficiency of the TBM tunnel.

[0009] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0010] A TBM tunnel face lithology identification method based on a self-supervised learning module includes:

[0011] Build a self-supervised learning model, which includes an encoder, a projection head, and a classifier;

[0012] constructing an auxiliary dataset using a public image dataset related to lithology, and training the self-supervised learning model using the auxiliary dataset to obtain a trained self-supervised learning model;

[0013] The rock slag image data to be identified is input into a trained self-supervised learning model. The trained self-supervised learning model uses the SimCLR algorithm to perform tunnel face lithology identification on the rock slag image to be identified, and determines the lithology of the rock slag image to be identified.

[0014] Preferably, the self-supervised learning model is constructed, and the self-supervised learning model includes an encoder, a projection head, and a classifier, including:

[0015] A self-supervised learning model is constructed, comprising an encoder, a projection head, and a classifier. The encoder selects ResNet-50 as the basic encoding architecture and outputs feature data with a feature dimension of 2048. The projection head includes a lightweight feature projection head g(·). The structure of the feature projection head g(·) is implemented using a single-layer MLP with a ReLU activation function. The feature projection head g(·) consists of two fully connected networks. The first fully connected network uses the ReLU activation function to map the 2048-dimensional feature data output by the encoder to 512-dimensional feature data in the hidden layer. The second fully connected network further maps the 512-dimensional feature data in the hidden layer to 128-dimensional feature data in the contrastive learning space.

[0016] The classifier adopts a single-layer fully connected network and is optimized by a cross-entropy loss function. The classifier extracts common lithology representations from a public rock image library using the SimCLR algorithm, performs tunnel face lithology recognition on the rock slag image to be identified, and determines the lithology of the rock slag image to be identified.

[0017] Preferably, the construction of the auxiliary dataset using the public image dataset related to lithology includes:

[0018] A public image dataset related to lithology was selected as an auxiliary dataset. Abnormal data in the auxiliary dataset was cleaned, and the effective rock slag images in the auxiliary dataset were center-cropped. The cropped effective rock slag images were 2000×2000 pixels.

[0019] Preferably, the using the auxiliary data set to train the self-supervised learning model to obtain a trained self-supervised learning model includes:

[0020] (1) The self-supervised learning model is trained using the auxiliary dataset after data cleaning. The self-supervised learning model is trained on the image X in the input auxiliary dataset. i and X j Perform image enhancement respectively to generate the corresponding enhanced view x i -x' i and x j -x' j , construct positive sample pairs x i -x' i and x j -x' j And negative sample pairs x i -x j 、x i -x' j 、x' i -x j 、x' i -x' j ;

[0021] (2) The encoder in the self-supervised learning model extracts image features from the enhanced view and represents the image feature vector as h;

[0022] (3) After the image feature vector h is processed by the projection head, the image feature vector h is mapped to the contrastive learning space to obtain z;

[0023] (4) Calculate the similarity S between the positive and negative sample pairs respectively;

[0024]

[0025] Where: z i and z j is the sample pair feature extracted in the pre-training phase; τ is an adjustable temperature parameter;

[0026] (5) Calculate the loss function L to maximize the positive sample pairs and minimize the negative sample pairs;

[0027]

[0028] Where: exp is the exponential function; 1[k!=i] means that when k and i are not equal, they are 1; N is the batch size;

[0029] (6) Update the parameters of the encoder and projection head of the self-supervised learning model through backpropagation, freeze and save the trained encoder network, and complete the pre-training in the self-supervised learning model;

[0030] (7) Downstream tasks, for the input image X k , the encoder in the self-supervised learning model extracts image features to obtain the vector representation h k ;

[0031] (8) Use a small amount of labeled data to train the classifier in the self-supervised learning model, and represent the vector h k Perform lithology classification. After training is completed, a trained self-supervised learning model is obtained.

[0032] Preferably, the rock slag image data to be identified is input into a trained self-supervised learning model, and the trained self-supervised learning model uses the SimCLR algorithm to perform tunnel face lithology identification on the rock slag image to be identified, and determines the lithology of the rock slag image to be identified, including:

[0033] The trained self-supervised learning model uses the SimCLR algorithm to extract common lithologic representations from a public rock image library. The rock debris image data to be identified in the target dataset is input into the trained self-supervised learning model. The self-supervised learning model uses an encoder to extract 2048-dimensional feature data from the rock debris image to be identified. The extracted feature data is then input into a classifier, which uses the SimCLR algorithm to perform data augmentation on the feature data. The data augmentation includes random scaling and cropping, random horizontal flipping, random image distortion, and random grayscale conversion.

[0034] The feature data after data enhancement processing are compared and analyzed with various common lithologic representations, and the probability p of mapping the feature data extracted by the classifier to the lithologic label is calculated by formula (2): i ;

[0035] p i =softmax(W c h i +b c ) (6)

[0036] Where: h i ∈R d is the feature data extracted by the classifier, W c ∈R k×d is the classification layer weight matrix, k is the number of categories; b c ∈R k is the bias term;

[0037] The classifier outputs the probability p corresponding to the set number of lithology labels i , the probability p of the highest finger i The corresponding lithology label is determined to be the lithology category corresponding to the slag image to be identified.

[0038] As can be seen from the technical solutions provided by the above-described embodiments of the present invention, the present invention employs a self-supervised learning approach, using publicly available lithology-related datasets to enhance the feature extraction capabilities of pre-trained models. By generalizing downstream data from a small number of images with lithology labels, the present invention can achieve lithology classification of TBM rock debris. This reduces the model's dependence on labeled data volume, enabling lithology identification and classification in the early stages of tunnel construction.

[0039] Additional aspects and advantages of the present invention will be set forth in part in the following description, will become apparent from the following description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 A schematic diagram illustrating a TBM tunnel face lithology identification method based on a self-supervised learning module according to an embodiment of the present invention;

[0042] Figure 2 A processing flow chart of a TBM tunnel face lithology identification method based on a self-supervised learning module is provided for an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of an image acquisition device built above a TBM conveyor belt for acquiring rock slag images of a TBM tunnel face is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0045] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the description of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or couplings. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0046] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention pertains. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such herein, will not be interpreted in an idealized or overly formal sense.

[0047] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0048] The implementation principle of the TBM tunnel face lithology identification method based on the self-supervised learning module provided by the embodiment of the present invention is as follows: Figure 1 As shown, the present invention first constructs a self-supervised learning model, which includes an encoder, a projection head and a classifier.

[0049] Encoder: ResNet-50 is selected as the basic encoding architecture, with an output feature dimension of 2048. During the pre-training phase, the encoder parameters are updated by comparing learning objectives; during downstream task fine-tuning, the encoder parameters are frozen and only used to extract high-level feature representations of the image.

[0050] Projection head: In the self-supervised learning module, a lightweight feature projection head g(·) is introduced. This feature projection head g(·) maps the high-dimensional features output by the encoder to an embedding space optimized for contrastive loss through nonlinear transformation. The structure of this feature projection head g(·) is implemented using a single-layer MLP with a ReLU activation function. The specific formal definition is shown in Equation (4). The projection head consists of two fully connected networks. The first fully connected network uses the ReLU activation function to map the high-dimensional features output by the encoder (2048 dimensions) to the hidden layer (512 dimensions). The second fully connected network further reduces the hidden layer (512 dimensions) features to the contrastive learning space (128 dimensions). This projection head structure enhances feature discriminability through nonlinear transformation and is jointly optimized with the encoder in the pre-training stage. However, in the downstream task, only the encoder output features are retained, and the projection head does not participate in the fine-tuning process.

[0051] z i =g(h i )=W (2) σ(W (1) h i )(1)

[0052] Where: W (1) ∈R d×d' and W (2) ∈R d'×d is the learnable weight matrix, and σ(·) represents the ReLU nonlinear activation function.

[0053] Classifier: The downstream classifier is a single-layer linear module (LogisticRegression) with an input dimension that matches the encoder's output feature dimension (2048) and an output dimension corresponding to the number of classes (n_classes = 6). The classifier is optimized using a cross-entropy loss function and an Adam optimizer (learning rate 3e-4). The classifier directly operates on the feature vectors extracted by the frozen encoder and predicts class probabilities through a fully connected layer. Through self-supervised pre-training, the classifier has learned visual representations with high generalization capabilities. Based on this, we adopt a representation freezing strategy for downstream task adaptation: all parameters of the pre-trained encoder are fixed, and only the top-level classifier is fine-tuned using a small number of labeled data. The classifier consists of a single-layer fully connected network, and its mathematical representation is shown in Equation (2).

[0054] p i =softmax(W c h i +b c )(2)

[0055] Where: h∈R d is the fixed feature representation extracted by the backbone network; W c ∈Rk×d is the classification layer weight matrix, k is the number of categories; b c ∈R k is the bias term; cross entropy loss function is used To optimize, p i It is the probability of mapping the feature data extracted by the classifier to the lithology label.

[0056] The input data of the self-supervised learning module is an image tensor, which is processed by image augmentation, including random scaling and cropping RandomResizedCrop (size=224), random horizontal flip RandomHorizontalFlip (p=0.5), random image distortion ColorJitter (p=0.8), and random grayscale RandomGrayscale (p=0.2).

[0057] The output data of the self-supervised learning module includes a probability distribution over six categories. The training phase records the loss, weighted F1 score, and accuracy, and the testing phase also generates a confusion matrix. The feature extraction phase maps the image into a 2048-dimensional feature vector, which is further converted into a 6-dimensional class probability vector during the classification phase.

[0058] The embodiment of the present invention provides a specific processing flow of a TBM tunnel face lithology identification method based on a self-supervised learning module. Figure 2 As shown, the processing steps include the following:

[0059] Step S10: construct a self-supervised learning model, where the self-supervised learning module includes an encoder, a projection head, and a classifier.

[0060] Step S20: constructing an auxiliary dataset using a public image dataset related to lithology.

[0061] Auxiliary datasets are public image datasets related to lithology, available from platforms such as Kaggle, such as the Kaggle rock classification database. This data, which does not require lithology labels, is used for comparative learning within the self-supervised learning model, enhancing the model's feature extraction capabilities.

[0062] The abnormal data in the auxiliary dataset are cleaned, and the effective rock slag images in the auxiliary dataset are center-cropped. The cropped effective rock slag images are 2000×2000 pixels to eliminate the redundant edge information of the image.

[0063] Step S30 , training the self-supervised learning model using the auxiliary data set after data cleaning to obtain a trained self-supervised learning model.

[0064] The self-supervised learning model is trained using the auxiliary dataset after data cleaning, and the auxiliary dataset is used for downstream tasks. The detailed process of the self-supervised learning model is as follows:

[0065] (1) The self-supervised learning model is trained on the image X in the auxiliary dataset. i and X j Perform image enhancement respectively to generate the corresponding enhanced view x i -x' i and x j -x' j , construct positive sample pairs x i -x' i and x j -x' j , construct negative sample pairs x i -x j 、x i -x' j 、x' i -x j 、x' i -x' j ;

[0066] (2) The encoder in the self-supervised learning model extracts image features from the enhanced view and represents the image feature vector as h;

[0067] (3) After the image feature vector h is processed by the projection head, the image feature vector h is mapped to the contrastive learning space and obtained as z

[0068] (4) Calculate the similarity S of the positive and negative sample pairs corresponding to z respectively;

[0069]

[0070] Where: z i and z j is the sample pair feature extracted in the pre-training phase; τ is an adjustable temperature parameter used to control the numerical range of cosine similarity and adjust the model's sensitivity to sample differences.

[0071] (5) Calculate the loss function L to maximize the positive sample pairs and minimize the negative sample pairs;

[0072]

[0073] Where exp is the exponential function; 1[k!=i] means that when k and i are not equal, they take 1; N is the batch size.

[0074] (6) Update the parameters of the encoder and projection head of the self-supervised learning model through backpropagation, freeze and save the trained encoder network, and complete the pre-training in the self-supervised learning model;

[0075] (7) Downstream tasks, for the input image X k , the encoder in the self-supervised learning model extracts image features to obtain the vector representation h k ;

[0076] (8) Use a small amount of labeled data to train the classifier in the self-supervised learning model, and represent the vector h k Perform lithology classification. After the training is completed, a trained self-supervised learning model is obtained.

[0077] The present invention uses public lithology-related datasets for pre-training, constructs a two-level architecture of "cross-domain pre-training-small sample fine-tuning" based on contrastive learning, enhances the feature extraction capability of the model through cross-domain feature migration, and uses only a small part of the target engineering dataset for downstream generalization to achieve face lithology classification of TBM rock debris.

[0078] In step S40 , the rock slag image data to be identified is input into the trained self-supervised learning model. The trained self-supervised learning model uses the SimCLR algorithm to perform tunnel face lithology identification on the rock slag image to be identified, and determines the lithology of the rock slag image to be identified.

[0079] Field application. Figure 3 As shown, in an embodiment of the present invention, an image acquisition device is built above the conveyor belt of the TBM. The image acquisition device acquires rock slag images of the TBM tunnel face in real time, and uses the acquired rock slag images to construct a target data set.

[0080] The rock slag image data to be identified in the target data set is input into the trained self-supervised learning model. The trained self-supervised learning model uses the SimCLR algorithm to perform tunnel face lithology recognition on the rock slag image to determine the lithology of the rock slag image to be identified.

[0081] The SimCLR algorithm extracts common lithologic representations from a public rock image library and then uses a small number of rock debris samples from current projects to train a downstream classifier for target generalization. This allows for lithologic classification of TBM rock debris at tunnel faces, even in the early stages of tunnel construction, when training samples are scarce. The data augmentation operations in the SimCLR algorithm are as follows, along with the corresponding PyTorch APIs: RandomResizedCrop (size = 224), RandomHorizontalFlip (p = 0.5), ColorJitter (p = 0.8), and RandomGrayscale (p = 0.2).

[0082] The self-supervised learning model extracts feature data (2048 dimensions) from the rock debris image to be identified through the encoder, and inputs the extracted feature data into the classifier. The downstream classifier uses the SimCLR algorithm to perform data enhancement processing on the feature data, and then compares and analyzes the feature data with various common lithologic representations. The similarity probability between the feature data and various common lithologic representations is calculated using Equation (6).

[0083] p i =softmax(W c h i +b c ) (6)

[0084] Where: h i ∈R d is the feature data extracted by the classifier, W c ∈R k×d is the classification layer weight matrix, k is the number of categories; b c ∈R k is the bias term;

[0085] The classifier outputs the probability p corresponding to the lithology label of the set number of categories (for example, n_classes = 6) i , the probability p of the highest finger i The corresponding lithology label is determined to be the lithology category corresponding to the slag image to be identified.

[0086] In summary, the data source of the embodiment of the present invention is more reliable: it discards mechanical parameters that are affected by TBM equipment specifications and selects rock slag images as input to identify tunnel face lithology. As a direct product of TBM cutterhead rock breaking, the visual characteristics of rock slag, such as morphology and texture, are intrinsically related to lithology. This gives the data labels objective uniqueness and scale invariance, avoids the problem of systematic deviation of mechanical parameters caused by equipment differences, and enables the model to generalize across projects and equipment.

[0087] The model has greater application prospects: This invention overcomes the strong reliance of supervised learning on massive amounts of labeled data by constructing a two-stage architecture: cross-domain pre-training and small-sample fine-tuning. Using self-supervised comparative learning (SimCLR), SimCLR extracts common lithologic representations from a public rock image library. It then uses a small number of engineering rock slag samples for feature space alignment. Ultimately, this model achieves an agile deployment model with zero-labeling cold start in the early stages of construction and progressive, precise iteration in the mid-term. This allows for face lithologic classification of TBM rock slag in the early stages of tunnel construction, even when training samples are scarce.

[0088] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0089] From the above description of the embodiments, it can be seen that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus the necessary general-purpose hardware platform. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0090] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without making any creative efforts.

[0091] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A TBM tunnel face lithology identification method based on a self-supervised learning module, characterized in that: include: Build a self-supervised learning model, which includes an encoder, a projection head, and a classifier; constructing an auxiliary dataset using a public image dataset related to lithology, and training the self-supervised learning model using the auxiliary dataset to obtain a trained self-supervised learning model; The rock slag image data to be identified is input into a trained self-supervised learning model. The trained self-supervised learning model uses the SimCLR algorithm to perform tunnel face lithology identification on the rock slag image to be identified, and determines the lithology of the rock slag image to be identified.

2. The method according to claim 1, characterized in that The self-supervised learning model is constructed, and the self-supervised learning model includes an encoder, a projection head, and a classifier, including: A self-supervised learning model is constructed, comprising an encoder, a projection head, and a classifier. The encoder selects ResNet-50 as the basic encoding architecture and outputs feature data with a feature dimension of 2048. The projection head includes a lightweight feature projection head g(·). The structure of the feature projection head g(·) is implemented using a single-layer MLP with a ReLU activation function. The feature projection head g(·) consists of two fully connected networks. The first fully connected network uses the ReLU activation function to map the 2048-dimensional feature data output by the encoder to 512-dimensional feature data in the hidden layer. The second fully connected network further maps the 512-dimensional feature data in the hidden layer to 128-dimensional feature data in the contrastive learning space. The classifier adopts a single-layer fully connected network and is optimized by a cross-entropy loss function. The classifier extracts common lithology representations from a public rock image library using the SimCLR algorithm, performs tunnel face lithology recognition on the rock slag image to be identified, and determines the lithology of the rock slag image to be identified.

3. The method according to claim 2, characterized in that The construction of an auxiliary dataset using a public image dataset related to lithology includes: A public image dataset related to lithology was selected as an auxiliary dataset. Abnormal data in the auxiliary dataset was cleaned, and the effective rock slag images in the auxiliary dataset were center-cropped. The cropped effective rock slag images were 2000×2000 pixels.

4. The method according to claim 3, characterized in that The method of using the auxiliary data set to train the self-supervised learning model to obtain a trained self-supervised learning model includes: (1) The self-supervised learning model is trained using the auxiliary dataset after data cleaning. The self-supervised learning model is trained on the image X in the input auxiliary dataset. i and X j Perform image enhancement respectively to generate the corresponding enhanced view x i -x' i and x j -x' j , construct positive sample pairs x i -x' i and x j -x' j And negative sample pairs x i -x j 、x i -x' j 、x' i -x j 、x' i -x' j ; (2) The encoder in the self-supervised learning model extracts image features from the enhanced view and represents the image feature vector as h; (3) After the image feature vector h is processed by the projection head, the image feature vector h is mapped to the contrastive learning space to obtain z; (4) Calculate the similarity S between the positive and negative sample pairs respectively; Where: z i and z j is the sample pair feature extracted in the pre-training phase; τ is an adjustable temperature parameter; (5) Calculate the loss function L to maximize the positive sample pairs and minimize the negative sample pairs; Where: exp is the exponential function; 1[k!=i] means that when k and i are not equal, they are 1; N is the batch size; (6) Update the parameters of the encoder and projection head of the self-supervised learning model through backpropagation, freeze and save the trained encoder network, and complete the pre-training in the self-supervised learning model; (7) Downstream tasks, for the input image X k , the encoder in the self-supervised learning model extracts image features to obtain the vector representation h k ; (8) Use a small amount of labeled data to train the classifier in the self-supervised learning model, and represent the vector h k Perform lithology classification. After training is completed, a trained self-supervised learning model is obtained.

5. The method according to claim 4, characterized in that The aforementioned step of inputting the rock slag image data to be identified into a trained self-supervised learning model, wherein the trained self-supervised learning model uses the SimCLR algorithm to perform tunnel face lithology identification on the rock slag image to be identified, and determines the lithology of the rock slag image to be identified, including: The trained self-supervised learning model uses the SimCLR algorithm to extract common lithologic representations from a public rock image library. The rock debris image data to be identified in the target dataset is input into the trained self-supervised learning model. The self-supervised learning model uses an encoder to extract 2048-dimensional feature data from the rock debris image to be identified. The extracted feature data is then input into a classifier, which uses the SimCLR algorithm to perform data augmentation on the feature data. The data augmentation includes random scaling and cropping, random horizontal flipping, random image distortion, and random grayscale conversion. The feature data after data enhancement processing are compared and analyzed with various common lithologic representations, and the probability p of mapping the feature data extracted by the classifier to the lithologic label is calculated by formula (2): i ; p i =softmax(W c h i +b c ) (6) Where: h i ∈R d is the feature data extracted by the classifier, W c ∈R k×d is the classification layer weight matrix, k is the number of categories; b c ∈R k is the bias term; The classifier outputs the probability p corresponding to the set number of lithology labels i , the probability p of the highest finger i The corresponding lithology label is determined to be the lithology category corresponding to the slag image to be identified.

Images