Roadway heading machine robust visual positioning method, device and system based on causal characterization learning

Abstract

The invention discloses a causal characterization learning-based robust visual positioning method, device and system for a tunnel boring machine, and the device comprises an industrial camera, an artificial target and a calculation system, the artificial target is arranged in a tunnel behind the tunnel boring machine, and the industrial camera is arranged on a body of the tunnel boring machine and is used for obtaining an image of the artificial target; the computing system estimates the pose of the aircraft body on line based on a camera pose regression deep learning model; the camera pose regression deep learning model comprises a multi-scale encoder, a feature fusion network and a pose regression network, the multi-scale encoder is used for extracting image multi-scale robust potential characterization, and pre-training is carried out based on a causal characterization learning method so as to realize image robust causal invariance feature extraction; the feature fusion network is used for processing potential representations of different scales and adding results or splicing the results in the depth direction, and the pose regression network is used for outputting six-degree-of-freedom pose parameters of the airframe. According to the invention, efficient and reliable on-line positioning of the tunnel boring machine under extreme working conditions is realized.

Description

Technical Field

[0001] This invention relates to a method for locating tunneling machines, and in particular to a robust visual positioning method, apparatus, and system for tunneling machines based on causal representation learning. Background Technology

[0002] Coal has long occupied a dominant position in the energy consumption structure. Underground coal mining requires extensive tunneling, which is a high-risk industry with frequent accidents. Achieving unmanned or minimally manned coal mine tunneling is an inevitable trend in the development of the coal mining industry. Autonomous positioning of tunnel boring machines (TBMs) is the primary condition for realizing intelligent tunneling and also a key issue limiting the level of intelligence of TBMs. Due to the harsh environment of underground tunnels and the extreme working conditions of the tunneling face, existing closed-space pose detection methods have certain shortcomings or difficulties: inertial positioning is affected by severe vibration of the machine body, resulting in serious cumulative errors; radio frequency positioning is greatly affected by multipath effects and signal blockage in narrow spaces with many devices; visual or lidar SLAM methods are prone to drift and cumulative errors in weak texture long straight tunnel environments due to feature scarcity and matching uncertainty, and lack loop closure conditions in tunneling scenarios, making it impossible to eliminate cumulative errors; visual positioning assisted by manual targets is severely affected by external interference such as occlusion, and feature extraction is unstable. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a robust visual positioning method for tunnel boring machines based on causal representation learning, so as to achieve accurate and robust visual positioning of tunnel boring machines under extreme working conditions.

[0004] Technical Solution: The positioning device of the present invention includes: an industrial camera, a manual target, and a computing system. The industrial camera is installed on the body of the tunneling machine to acquire images of the manual target. The manual target is installed in the tunnel behind the tunneling machine to provide rich information for full-state visual measurement. During the tunneling operation, the computing system estimates the body pose online based on a camera pose regression deep learning model. The camera pose regression deep learning model includes a multi-scale encoder, a feature fusion network, and a pose regression network. The multi-scale encoder is used to extract multi-scale robust latent representations of the image. It is pre-trained based on a causal representation learning method and can achieve robust causal invariant feature extraction of the image. The feature fusion network is used to process latent representations at different scales and add the results or stitch them together in the depth direction. The pose regression network is used to output the six-degree-of-freedom pose parameters of the body.

[0005] Optionally, the multi-scale encoder is based on a multi-scale variational autoencoder model and trained using a causal representation learning strategy. The multi-scale variational autoencoder model includes a multi-scale encoder and a decoder. The multi-scale latent representations of the original image and the intervention image are extracted by the multi-scale encoder, and the image is reconstructed by the decoder with the original image as the target. This allows the latent representation to contain all the effective information of the artificial target image. At the same time, a causal representation method based on decoupled contrastive learning is used to enable the latent representation to capture the robust causal invariance features of the artificial target image.

[0006] The positioning method of the present invention includes the following steps:

[0007] Artificial target image-camera pose data pairs were obtained by simulating a tunnel environment in a ground-based laboratory, and the obtained data pairs were augmented to obtain the original image-pose label dataset; among them, an intervention image dataset was generated based on the original image dataset.

[0008] A multi-scale variational autoencoder (MAE) model and a camera pose regression deep learning model were constructed. The MAE model consists of a multi-scale encoder and a decoder. The MAE model was trained using the original image dataset and the corresponding intervention image dataset. This included: extracting multi-scale latent representations from the original and intervention images using the multi-scale encoder, and then reconstructing the images using the original images as targets using the decoder. This ensured that the latent representations contained all the effective information of the artificial target images. Simultaneously, a causal representation method based on decoupled contrastive learning was used to enable the latent representations to capture the robust causal invariance features of the artificial target images. The camera pose regression deep learning model consists of a trained multi-scale encoder, a feature fusion network, and a pose regression network. The camera pose regression deep learning model was trained using an intervention image-pose label dataset. This included: the trained multi-scale encoder extracting multi-scale latent representations from the intervention images; the feature fusion network processing the latent representations at different scales and summing or stitching the results along the depth direction; and the pose regression network outputting the six-DOF pose parameters of the organism.

[0009] In practical engineering deployments, the tunnel boring machine's pose is estimated online using a trained camera pose regression deep learning model based on the actual acquired images of artificial targets, thus achieving robust visual positioning of the tunnel boring machine.

[0010] Furthermore, the method for augmenting the acquired artificial target image-camera pose data pairs is as follows: using a controllable pose image generation method based on neural radiation field or 3D Gaussian sputtering, a large number of camera image-camera pose data pairs with arbitrary pose parameters are generated.

[0011] The process of generating an intervention image dataset based on the original image dataset involves adding various potential interference signals to the original images in the original image dataset to obtain the intervention image dataset.

[0012] Furthermore, the multi-scale variational autoencoder model is trained based on the designed lower bound of evidence. At the same time, a decoupled contrastive learning method for causal invariant representations is incorporated into the training process: for each batch of input data during model training, the latent representations of the original image and the intervention image are obtained through the multi-scale encoder. Based on the vector Pearson correlation coefficient, an inductive bias is designed to make the correlation matrix output by the multi-scale encoder approach the identity matrix. That is, it maximizes the correlation of the latent representations of the original image and the intervention image at the same index position in the feature dimension and minimizes the correlation at different index positions, so that the multi-scale latent representation is causally invariant to confounding factors, while making its dimensions independent of each other.

[0013] Furthermore, the training of the multi-scale variational autoencoder model is achieved by maximizing the following lower bound of evidence:

[0014] ;

[0015] in, Indicates the lower bound of evidence. Represent two distributions and KL divergence, In order to be in conform to Calculate the expectation under the distribution. For decoder models, For multi-scale encoder models, As a potential representation of the prior distribution, This represents the training parameters of the multi-scale encoder model. This represents the training parameters of the decoder model.

[0016] Furthermore, the inductive bias of the design is expressed as:

[0017] ;

[0018] ;

[0019] in, This is the inductive bias, i.e., the causal representation loss function. Represents the identity matrix. Correlation matrix Elements in, latent representation pass Obtained by sampling , This represents the total dimension of the multi-scale latent representation. , Representing the latent representations of the original and interventional images in a batch, respectively, in the ... and A vector composed of feature locations.

[0020] This invention also provides a robust visual positioning system for tunnel boring machines based on causal representation learning, comprising:

[0021] The dataset creation unit is used to acquire artificial target image-camera pose data pairs by simulating a tunnel environment in a ground laboratory, and to augment the acquired data pairs to obtain the original image-pose label dataset; based on the original image dataset, an intervention image dataset is generated.

[0022] The model building and training unit is used to build a multi-scale variational autoencoder model and a camera pose regression deep learning model. The multi-scale variational autoencoder model includes a multi-scale encoder and a decoder. It is trained using the original image dataset and the intervention image dataset. The process includes: extracting multi-scale latent representations of the original image and the intervention image respectively through the multi-scale encoder, and reconstructing the image with the original image as the target by the decoder. This ensures that the latent representations can contain all the effective information of the artificial target image. At the same time, a causal representation method based on decoupled contrastive learning is used to enable the latent representations to capture the robust causal invariance features of the artificial target image. The camera pose regression deep learning model includes a trained multi-scale encoder, a feature fusion network, and a pose regression network. It is trained using the intervention image-pose label dataset. The process includes: the trained multi-scale encoder is used to extract multi-scale robust latent representations of the intervention image; the feature fusion network is used to process the latent representations at different scales and add the results or stitch them together in the depth direction; and the pose regression network is used to output the six-degree-of-freedom pose parameters of the organism.

[0023] In practical engineering applications, the real-time positioning unit is used to estimate the pose of the tunnel boring machine online using a trained camera pose regression deep learning model based on the actual acquired images of artificial targets, thereby achieving robust visual positioning of the tunnel boring machine.

[0024] The present invention also provides an electronic device, the device comprising:

[0025] Memory containing executable program code;

[0026] A processor coupled to the memory;

[0027] The processor calls the executable program code stored in the memory to execute the method.

[0028] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method described.

[0029] Beneficial Effects: Compared with existing technologies, the significant technical effects of this invention are as follows: This invention introduces causal representation learning into the visual positioning task of tunnel boring machines for the first time. By constructing a multi-scale variational autoencoder and a causal invariance decoupled contrastive learning strategy, it effectively eliminates the influence of non-causal factors such as image noise, brightness changes, uncertain external light sources, visual occlusion, and motion blur on visual features, significantly improving the stability and generalization ability of visual feature representation and enhancing the positioning performance of tunnel boring machines under extreme working conditions. Due to the use of the causal representation learning strategy, the positioning model proposed in this invention can be directly deployed across scenarios without the need for retraining with a large amount of field data, significantly improving the engineering applicability of the proposed tunnel boring machine positioning method and enabling long-term continuous operation in complex and variable underground environments. This method can be used for tunnel boring machines, rock drilling rigs, continuous coal mining machines, and other tunnel boring equipment. Attached Figure Description

[0030] Figure 1 A schematic diagram showing the deployment of the visual positioning device for a tunnel boring machine;

[0031] Figure 2 This is a flowchart of the method of the present invention;

[0032] Figure 3 This is a schematic diagram of an image robust feature extraction model based on causal representation learning;

[0033] Figure 4 This is a schematic diagram of a deep learning model for camera pose regression. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0035] The robust visual positioning device for tunnel boring machines described in this invention includes an industrial camera, a human target, and a computing system. The industrial camera and the human target are deployed as follows: Figure 1As shown, an industrial camera is installed on the tunnel boring machine (TBM) body, and a structured artificial target is set in the tunnel behind the TBM. This artificial target comprises a rectangular LED array and multiple lasers at specific angles, providing rich information for full-state visual measurement and ensuring a large effective positioning range. The computing system estimates the machine's pose online based on a camera pose regression deep learning model. In practical engineering, to achieve long-distance continuous positioning of the TBM, when the TBM moves beyond the effective visual positioning range, the artificial target needs to be moved forward, and other precise positioning methods assisted by manual measurement (such as a total station) are used to measure the six-degree-of-freedom pose of the artificial target. The absolute pose state of the TBM body in the tunnel is obtained through coordinate transformation. For deep learning-based tunnel boring machine (TBM) positioning strategies, deep learning models require a large amount of training data. However, it is difficult to obtain sufficient image-pose label data pairs in actual underground tunnel excavation environments. Therefore, this invention obtains training data in a simulated tunnel environment in a laboratory setting: In the simulated tunnel environment, an industrial camera is mounted on a moving machine, and a human target is positioned behind the machine. A motion capture system is used to calibrate the pose parameters of the moving machine relative to the human target online, so as to efficiently obtain rich camera image-pose label data pairs as training data for the deep learning model.

[0036] This invention designs a robust visual localization method for tunnel boring machines based on causal representation learning. This method promotes the effective generalization of trained localization models in simulated tunnel environments, real-world engineering scenarios under different working conditions, and in situations with unknown and uncertain interference. This enables the direct deployment and use of the trained localization model in the field application of tunnel boring machines. For example... Figure 2 As shown, the positioning method of the present invention includes the following steps:

[0037] S1. Create the training dataset;

[0038] Training data is acquired by simulating a tunnel environment in a ground-based laboratory: In the simulated tunnel environment, an industrial camera is mounted on a mobile machine, and a human target is placed behind the machine. A motion capture system is used to calibrate the pose parameters of the mobile machine relative to the human target online, so as to efficiently acquire rich camera image-pose label data pairs as training data for deep learning models. The camera image-pose label data pair is a data pair formed by the image acquired at a certain moment and the pose of the mobile machine at that moment.

[0039] To further augment the training dataset and improve model performance, a controllable pose image generation method based on Neural Radiation Field (NeRF) or 3D Gaussian Sputtering (3DGS) is used to generate a large number of camera image-pose label data pairs with arbitrary pose parameters. Images acquired by industrial cameras and augmented images are used to form the original image of the artificial target.

[0040] Considering the potential interference of tunnel boring machine operating conditions on visual positioning (image noise, brightness variation, uncertain external light source, visual occlusion, motion blur, etc.), an artificial target image intervention strategy is designed to generate an intervention image; that is, potential external interference is added to the original image to generate an intervention image.

[0041] S2. Build a robust visual localization model for tunnel boring machines based on causal representation learning, including a robust image feature extraction model and a camera pose regression deep learning model based on causal representation learning.

[0042] This invention establishes and trains a robust visual localization model for tunnel boring machines based on causal representation learning, which consists of two steps. The first step involves building and training an image robust feature extraction model based on causal representation learning, such as... Figure 3 As shown, on the one hand, a multi-scale variational autoencoder model (convolutional neural network architecture) is established, including a multi-scale encoder and a decoder. The multi-scale encoder extracts multi-scale latent representations of the original image and the intervention image respectively, and the decoder reconstructs the image with the original image as the target. This ensures that the latent representation can contain all the effective information of the artificial target image, and the multi-scale latent layer structure of the multi-scale encoder can improve the model's learning and expressive capabilities. On the other hand, a decoupled contrastive learning method for causal invariant representation is proposed. For each batch of data during the model training process... (Including original images and intervention images), the original images are obtained separately using a multi-scale encoder. and intervention images Potential representations and Based on the vector Pearson correlation coefficient, an effective inductive bias is designed to enable the multi-scale encoder to output a correlation matrix. Approaching the identity matrix That is, maximizing the latent representation and By minimizing the correlation at the same index position across feature dimensions and reducing the correlation at different index positions, the multi-scale latent representation is causally invariant to confounding factors while promoting the independence of its dimensions. Through a self-supervised learning strategy, using a large amount of raw and intervened image data, and based on a causal invariance-decoupled contrastive learning method, robust features of artificial target images with causal invariance are provided for the visual localization of tunnel boring machines.

[0043] How is the above-mentioned image robust feature extraction model based on causal representation learning trained (mathematical explanation): The multi-scale variational autoencoder aims to learn the joint distribution of an image sample and multi-scale latent representation variables. The expression is ,in Represents image samples, Representing multi-scale latent characterization variables, For the first Potential representation variables at each scale , The total number of scales for the latent characterization variables. Indicates the training parameters of the decoder model. For the prior distribution of multi-scale latent characterization variables, Let be the likelihood function (i.e., the decoder model). A multi-scale encoder is introduced. Approximating the unsolvable true posterior distribution ,in This represents the training parameters of the multi-scale encoder model. Based on variational inference, model training can be achieved by maximizing the lower bound of evidence (ELBO), as shown below:

[0044] ;

[0045] in, Describe a lower bound function for evidence. Represent two distributions and KL divergence, In order to be in conform to Calculate the expectation under the distribution. To enable the model to represent causality, the designed inductive bias can be expressed as:

[0046] ;

[0047] ;

[0048] in, This is the inductive bias (i.e., the causal representation loss function). Represents the identity matrix. Correlation matrix Elements in, latent representation pass Obtained by sampling , This represents the total dimension of the multi-scale potential representation (i.e., the total dimension of features). , Representing the latent representations of the original and interventional images in a batch, respectively, in the ... and The vector is composed of feature locations and then normalized using Z-score. The final model training can be achieved by simultaneously optimizing... and accomplish.

[0049] The second step is to use a deep learning model for camera pose regression, such as... Figure 4As shown, the multi-scale latent representations of the image (i.e., robust features in the figure) are extracted using the multi-scale encoder portion trained in the first step. Then, a feature fusion network and a pose regression network are used to estimate the six-DOF pose of the camera. The feature fusion network uses convolutional layers to process the latent representations at different scales and adds or concatenates the results along the depth direction. The pose regression network uses convolutional layers and fully connected layers to finally output the six-DOF pose parameters. During model training, the input image is the intervention image (with pose labels) to further improve the robustness of the localization model. Additionally, the parameters of the trained multi-scale encoder portion are frozen or only fine-tuned, with the main training focused on other parts of the model to maintain the robust feature extraction capability acquired through causal representation learning. In actual deployment, the input image is the currently acquired image to achieve online localization of the tunnel boring machine.

[0050] S3. In practical engineering deployments, based on the acquired images of artificial targets, a trained camera pose regression deep learning model is used to estimate the tunnel boring machine's pose online, achieving robust visual localization of the tunnel boring machine. This includes:

[0051] (1) Deploy the trained camera pose regression deep learning model on the computing system, and according to Figure 1 The diagram shows the initial setup of the positioning system in the tunnel to be excavated;

[0052] (2) Use other precise positioning methods for manual measurement (e.g., total station) to detect the pose parameters of the artificial target in the roadway;

[0053] (3) The tunnel boring machine performs tunneling work and uses a trained camera pose regression deep learning model to estimate the machine pose online;

[0054] (4) Continue tunneling forward until it exceeds the effective working range of visual positioning, then manually move the target forward;

[0055] (5) Repeat steps (2) to (4).

[0056] In another embodiment of the present invention, a robust visual positioning system for tunnel boring machines based on causal representation learning includes:

[0057] The dataset creation unit is used to acquire artificial target image-pose label data by simulating a tunnel environment in a ground laboratory, and to augment the acquired data to obtain the original image dataset; based on the original image dataset, an intervention image dataset is generated.

[0058] The model building and training unit is used to build a multi-scale variational autoencoder model and a camera pose regression deep learning model. The multi-scale variational autoencoder model includes a multi-scale encoder and a decoder. It is trained using the original image dataset and the intervention image dataset. The process includes: extracting multi-scale latent representations of the original image and the intervention image through the multi-scale encoder, and reconstructing the image using the original image as the target image through the decoder. This ensures that the latent representations can contain all the effective information of the artificial target image. At the same time, a causal representation method based on decoupled contrastive learning is used to enable the latent representations to capture the robust causal invariance features of the artificial target image. The camera pose regression deep learning model includes a trained multi-scale encoder, a feature fusion network, and a pose regression network. It is trained using the intervention image dataset. The process includes: the trained multi-scale encoder is used to extract multi-scale latent representations of the intervention image; the feature fusion network is used to process the latent representations at different scales and add the results or stitch them together in the depth direction; and the pose regression network is used to output the six-degree-of-freedom pose parameters of the organism.

[0059] The real-time positioning unit is used to estimate the pose of the tunnel boring machine online using a trained camera pose regression deep learning model based on the actual acquired images of artificial targets, thereby achieving robust visual positioning of the tunnel boring machine.

[0060] In another embodiment of the present invention, an electronic device includes:

[0061] Memory containing executable program code;

[0062] A processor coupled to the memory;

[0063] The processor calls the executable program code stored in the memory to execute the method.

[0064] In another embodiment of the present invention, a computer program product includes a computer program / instructions that, when executed by a processor, implement the method described herein.

Claims

1. A robust visual positioning device for tunnel boring machines based on causal representation learning, characterized in that, include: The system comprises an industrial camera, a human target, and a computing system. The industrial camera is mounted on the tunnel boring machine (TBM) to acquire images of the human target, which is positioned in the tunnel behind the TBM to provide rich information for full-state visual measurement. During the TBM's tunneling operation, the computing system estimates the machine's pose online based on a camera pose regression deep learning model. This model includes a multi-scale encoder, a feature fusion network, and a pose regression network. The multi-scale encoder extracts robust latent representations of the image at multiple scales. It is pre-trained based on a causal representation learning method, enabling robust causal invariant feature extraction. The feature fusion network processes latent representations at different scales and sums or stitches the results along the depth direction. The pose regression network outputs the machine's six-degree-of-freedom pose parameters.

2. The robust visual positioning device for tunnel boring machines based on causal representation learning according to claim 1, characterized in that, The multi-scale encoder is based on a multi-scale variational autoencoder model and trained using a causal representation learning strategy. The multi-scale variational autoencoder model includes a multi-scale encoder and a decoder. The multi-scale latent representations of the original image and the intervention image are extracted by the multi-scale encoder, and the image is reconstructed by the decoder with the original image as the target. This allows the latent representation to contain all the effective information of the artificial target image. At the same time, a causal representation method based on decoupled contrastive learning is used to enable the latent representation to capture the robust causal invariance features of the artificial target image.

3. A robust visual localization method for tunnel boring machines based on causal representation learning, characterized in that, Includes the following steps: Artificial target image-camera pose data pairs were obtained by simulating a tunnel environment in a ground-based laboratory, and the obtained data pairs were augmented to obtain the original image-pose label dataset; among them, an intervention image dataset was generated based on the original image dataset. A multi-scale variational autoencoder (MAE) model and a camera pose regression deep learning model were constructed. The MAE model consists of a multi-scale encoder and a decoder. The MAE model was trained using the original image dataset and the corresponding intervention image dataset. This included: extracting multi-scale latent representations from the original and intervention images using the multi-scale encoder, and then reconstructing the images using the original images as targets using the decoder. This ensured that the latent representations contained all the effective information of the artificial target images. Simultaneously, a causal representation method based on decoupled contrastive learning was used to enable the latent representations to capture the robust causal invariance features of the artificial target images. The camera pose regression deep learning model consists of a trained multi-scale encoder, a feature fusion network, and a pose regression network. The camera pose regression deep learning model was trained using an intervention image-pose label dataset. This included: the trained multi-scale encoder extracting multi-scale latent representations from the intervention images; the feature fusion network processing the latent representations at different scales and summing or stitching the results along the depth direction; and the pose regression network outputting the six-DOF pose parameters of the organism. In practical engineering deployments, the tunnel boring machine's pose is estimated online using a trained camera pose regression deep learning model based on the actual acquired images of artificial targets, thus achieving robust visual positioning of the tunnel boring machine.

4. The robust visual localization method for tunnel boring machines based on causal representation learning according to claim 3, characterized in that, The method for augmenting the acquired artificial target image-camera pose data pairs is as follows: a large number of camera image-camera pose data pairs with arbitrary pose parameters are generated using a controllable pose image generation method based on neural radiation field or 3D Gaussian sputtering. The process of generating an intervention image dataset based on the original image dataset involves adding various potential interference signals to the original images in the original image dataset to obtain the intervention image dataset.

5. The robust visual localization method for tunnel boring machines based on causal representation learning according to claim 3, characterized in that, The multi-scale variational autoencoder model is trained based on the designed lower bound of evidence. At the same time, a decoupled contrastive learning method for causal invariant representation is incorporated into the training process: for each batch of input data during the model training process, the latent representations of the original image and the intervention image are obtained by the multi-scale encoder respectively. Based on the vector Pearson correlation coefficient, an inductive bias is designed to make the correlation matrix output by the multi-scale encoder approach the identity matrix. That is, it maximizes the correlation of the latent representations of the original image and the intervention image at the same index position in the feature dimension and minimizes the correlation at different index positions, so that the multi-scale latent representation is causally invariant to confounding factors, while making its dimensions independent of each other.

6. The robust visual localization method for tunnel boring machines based on causal representation learning according to claim 5, characterized in that, Training a multiscale variational autoencoder model is achieved by maximizing the following lower bound of evidence: ； in, Indicates the lower bound of evidence. Represent two distributions and KL divergence, In order to be in conform to Calculate the expectation under the distribution. For decoder models, For multi-scale encoder models, As a potential representation of the prior distribution, This represents the training parameters of the multi-scale encoder model. This represents the training parameters of the decoder model.

7. The robust visual localization method for tunnel boring machines based on causal representation learning according to claim 5, characterized in that, The inductive bias of the design is expressed as: ； ； in, This is the inductive bias, i.e., the causal representation loss function. Represents the identity matrix. Correlation matrix Elements in, latent representation pass Obtained by sampling , This represents the total dimension of the multi-scale latent representation. , Representing the latent characterization of a batch of raw and intervened images respectively in the 1st... and A vector composed of feature locations.

8. A robust visual positioning system for tunnel boring machines based on causal representation learning, characterized in that, include: The dataset creation unit is used to acquire artificial target image-camera pose data pairs by simulating a tunnel environment in a ground laboratory, and to augment the acquired data pairs to obtain the original image-pose label dataset; based on the original image dataset, an intervention image dataset is generated. The model building and training unit is used to build a multi-scale variational autoencoder model and a camera pose regression deep learning model. The multi-scale variational autoencoder model includes a multi-scale encoder and a decoder. It is trained using the original image dataset and the intervention image dataset. The process includes: extracting multi-scale latent representations of the original image and the intervention image respectively through the multi-scale encoder, and reconstructing the image with the original image as the target by the decoder. This ensures that the latent representations can contain all the effective information of the artificial target image. At the same time, a causal representation method based on decoupled contrastive learning is used to enable the latent representations to capture the robust causal invariance features of the artificial target image. The camera pose regression deep learning model includes a trained multi-scale encoder, a feature fusion network, and a pose regression network. It is trained using the intervention image-pose label dataset. The process includes: the trained multi-scale encoder is used to extract multi-scale robust latent representations of the intervention image; the feature fusion network is used to process the latent representations at different scales and add the results or stitch them together in the depth direction; and the pose regression network is used to output the six-degree-of-freedom pose parameters of the organism. In practical engineering applications, the real-time positioning unit is used to estimate the pose of the tunnel boring machine online using a trained camera pose regression deep learning model based on the actual acquired images of artificial targets, thereby achieving robust visual positioning of the tunnel boring machine.

9. An electronic device, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor invokes the executable program code stored in the memory to execute the method according to any one of claims 3 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method described in any one of claims 3 to 7.

Images