Coal mine tunnel deformation detection basic model and detection method based on self-supervised learning

Abstract

A coal mine tunnel deformation detection basic model and method based on self-supervised learning, the model comprising a tunnel deformation point cloud data generation module, a data enhancement module and a tunnel deformation intelligent detection module connected in sequence, the tunnel deformation intelligent detection module comprising a self-supervised learning module and a deformation detection module; the method comprising: using the tunnel deformation point cloud data generation module to generate tunnel point cloud data with different deformation types through simulation; using the data enhancement module to perform deformation processing on the deformation tunnel point cloud data; constructing the self-supervised learning module to automatically extract features from unlabeled point cloud data; constructing the deformation detection module to realize accurate identification of the tunnel deformation area and evaluation of the deformation degree; and using the tunnel deformation intelligent detection module to automatically identify and evaluate the tunnel deformation area. The model and method can effectively improve the accuracy and efficiency of tunnel deformation detection and provide an efficient and intelligent solution for tunnel safety monitoring.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of point cloud target detection and intelligent monitoring of roadway deformation, and particularly relates to a coal mine roadway deformation detection basic model and method based on self-supervised learning. BACKGROUND

[0002] In recent years, with the rapid development of artificial intelligence, Internet of Things, 5G communication, edge computing and robot technology, the intelligent monitoring technology of mines has made significant progress. The application of artificial intelligence and deep learning technology in the field of mine monitoring has shown great potential. The rapid development of multi-modal data fusion, real-time environmental perception and geological structure prediction not only improves the safety and efficiency of mine production, but also makes automated and precise mine monitoring possible. In the field of mine safety monitoring, three-dimensional point cloud processing and sensor data analysis methods based on deep learning have achieved remarkable results in environmental monitoring, equipment health assessment and structure deformation detection. Through these intelligent monitoring technologies, potential risks such as geological structure changes, equipment failures and roadway deformations in mines can be detected and predicted in real time, and timely warnings can be issued to effectively prevent disasters and ensure the safety of miners and improve production efficiency.

[0003] In mine intelligent monitoring, common sensor types include LiDAR, high-definition camera, infrared thermal imaging sensor, ground penetrating radar (GPR), etc., each of which plays a unique role in different scenarios. LiDAR has obvious advantages over image sensors. LiDAR can provide accurate three-dimensional point cloud data, which can directly reflect spatial depth information, and is suitable for three-dimensional monitoring of roadway deformation and terrain changes. Image sensors can only provide two-dimensional images and lack depth perception. In addition, LiDAR is not affected by light conditions, dust and moisture in the mine environment, and can work stably in dark or harsh environments, while image sensors may have image blurring, detail loss and other problems in low light or disturbed environments. In terms of data processing, LiDAR can provide high-density point cloud data, which can accurately capture small deformations and perform three-dimensional modeling, while images need to be calculated to derive depth information. Therefore, LiDAR has more advantages in mine roadway deformation detection and spatial monitoring, and can provide more accurate and stable monitoring results.

[0004] Point cloud data for intelligent monitoring of mines is mainly collected by geological radar or laser radar equipment, which contains high-precision three-dimensional spatial information of the mine environment. However, the point cloud data of coal mine tunnels involves dense three-dimensional coordinate information in space, and the deformation at different time points is relatively subtle. It is very complex and labor-intensive to label the deformation information of each point, and the standardized labeling method is not mature, so using self-supervised learning method can effectively reduce the burden of manual labeling and avoid relying on complex and immature labeling process. Self-supervised learning can automatically learn the spatio-temporal features of point cloud data from unlabeled data, identify and model subtle deformation, and generate high-quality training data in an unsupervised manner, greatly improving the efficiency and accuracy of data processing. In addition, self-supervised learning can automatically adjust the model to adapt to different tunnel deformation patterns, thereby enhancing the generalization ability of the model and showing strong robustness in the face of data loss, noise interference and other practical problems. In addition, due to the slow occurrence of tunnel deformation and the limitation of mine safety and equipment, the number of data samples is seriously insufficient, so using three-dimensional point cloud deformation enhancement technology to simulate and generate tunnel point cloud deformation data has become an effective solution. By deforming existing three-dimensional point cloud data to different degrees, different geological changes and tunnel deformations can be simulated, thereby expanding the diversity of the data set. Such techniques include geometric transformations such as rotation, translation, scaling, and noise addition to point cloud data to simulate tunnel settlement, crack expansion, or deformation caused by geological disasters. In addition, by combining physical models or finite element analysis (FEA) methods with data collected by sensors such as laser radars, the deformation of point clouds can be further enhanced to simulate more realistic training data. These deformation-enhanced data not only help improve the generalization ability of deep learning models in real-world environments, but also improve the accuracy and reliability of tunnel deformation detection in data-scarce situations. These models can quickly and accurately identify abnormal changes in the mine environment through deep learning and point cloud processing techniques, thereby providing reliable risk assessment and early warning.

[0005] In summary, the three-dimensional point cloud deformation detection method for coal mine tunnels represents the forefront of the field of combining intelligent monitoring of mines with artificial intelligence technology, and has great potential to improve the accuracy of mine environment monitoring and risk prediction, providing more efficient and safe tools for mine managers to timely warn of potential geological deformation and equipment failure, thereby ensuring the safety of miners and optimizing production processes. Research and application in this field is at a critical stage of the development of intelligent mines, and further technological breakthroughs and application promotion are urgently needed. SUMMARY

[0006] To solve the above problems in the prior art, the present application provides a coal mine roadway deformation detection basic model based on self-supervised learning and a detection method, which has a simple structure and can effectively improve the intelligent degree of roadway deformation detection, effectively improve the accuracy of mine environment monitoring and risk prediction, and provide an efficient and intelligent solution for safety monitoring of coal mine roadways; the method can effectively solve the problems of scarce underground roadway deformation data and complex point cloud data annotation, reduce the dependence on manual annotation, effectively improve the accuracy and efficiency of roadway deformation detection, and significantly improve the ability of coal mine safety monitoring.

[0007] To achieve the above purpose, the present application provides a coal mine roadway deformation detection basic model based on self-supervised learning, which comprises a roadway deformation point cloud data generation module, a data enhancement module and a roadway deformation intelligent detection module.

[0008] The roadway deformation point cloud data generation module is used to generate deformed three-dimensional point cloud data based on the labeled coal mine roadway original point cloud data and according to the roadway geometric structure and different deformation types, and then send the three-dimensional point cloud data to the data enhancement module.

[0009] The data enhancement module is used to deform the three-dimensional point cloud data to different degrees, and at the same time, introduce noise and disturbance to realistically simulate actual data to expand the diversity of the data set and obtain an expanded data set, and then send the expanded data set to the self-supervised learning module.

[0010] The roadway deformation intelligent detection module comprises a self-supervised learning module and a deformation detection module.

[0011] The self-supervised learning module is used to automatically extract features from unlabeled point cloud data in the expanded data set through variational auto-encoding contrast learning and cross-entropy loss supervision, and send the obtained feature data to the deformation detection module.

[0012] The deformation detection module is used to accurately identify the roadway deformation area according to the feature data and accurately evaluate the severity of the deformation.

[0013] As a preferred, the different deformation types include subsidence deformation, extrusion deformation, expansion deformation and inclination deformation.

[0014] As a preferred, the data deformation processing includes rotation transformation, translation transformation, scaling transformation and noise addition.

[0015] In the present application, by setting the roadway deformation point cloud data generation module, point cloud data of the roadway with different deformation types can be generated by simulation method, so that a large amount of simulated three-dimensional roadway deformation data can be obtained, and sufficient sample data is provided. Through the setting of the data enhancement module, the original point cloud data set is fully enhanced through rotation, translation, scaling, noise addition operation, providing more abundant and diversified samples for training model, which is beneficial to obtain detection model with higher detection precision. Through the setting of the self-supervised learning module, technical features related to roadway deformation can be automatically and quickly extracted from unlabeled point cloud data, which can provide reliable technical support for subsequent intelligent detection process. Through the setting of the deformation detection module, the roadway deformation area and deformation condition can be accurately identified according to the feature data, and the deformation program evaluation work can be carried out. The model has simple structure, which can effectively improve the intelligent degree of roadway deformation detection, can effectively improve the accuracy of mine environment monitoring and risk prediction, and provides an efficient and intelligent solution for safety monitoring of coal mine roadway.

[0016] The present application also provides a detection method of a coal mine roadway deformation detection basic model based on self-supervised learning, which adopts a coal mine roadway deformation detection basic model based on self-supervised learning, including the following steps:

[0017] Step one: using the roadway deformation point cloud data generation module to generate roadway point cloud data with different deformation types by simulation method;

[0018] For the labeled coal mine roadway original point cloud data P, according to the roadway geometric structure and different deformation types, combined with the physical model and numerical calculation, the deformation roadway point cloud data P is obtained by formula (1) f ; wherein P={(x i ,y i ,z i )|i=1,2,3,....,N}, (x i ,y i ,z i ) represents the i-th three-dimensional point coordinate in the roadway, and N is the total number of points in the point cloud;

[0019] P f =P·T+b (1);

[0020] In the formula, f represents the deformation type, T represents the transformation matrix related to the deformation type, and b represents the translation offset;

[0021] Step two: using the data enhancement module to deform the deformation roadway point cloud data;

[0022] The deformation roadway point cloud data P fThe point cloud data P after the enhancement processing is obtained by using formula (2) aug ;

[0023] P aug =(P f ·R)+t·(1+ε) (2);

[0024] In the formula, R is an enhancement transformation matrix; t is a translation vector; ε represents the noise added by randomization, and (1+ε) represents a scaling factor of the change by adding noise by randomization;

[0025] Step three: constructing a roadway deformation intelligent detection module and performing automatic identification and evaluation of the roadway deformation region;

[0026] S31: constructing a self-supervised learning module for automatically extracting features from unlabeled point cloud data;

[0027] The variational autoencoder and the contrastive learning are combined to form a variational autoencoding contrastive learning network structure, and the unlabeled point cloud data is subjected to self-supervised feature learning by using the variational autoencoding contrastive learning network structure; the latent feature representation of the point cloud data is extracted by using the variational autoencoder, and through contrastive learning, the distribution of the positive sample features extracted by the encoder converges to be similar, while distinguishing the distribution between the positive samples and the negative samples; at the same time, a cross-entropy loss function is introduced to supervise the training process, a generative adversarial mechanism is introduced, and the generated latent features are optimized through contrastive learning, so that the model can effectively capture the essential features in the point cloud data; at the same time, the cross-entropy loss is calculated by using the positive samples generated by two different data augmentations to retain more invariant features and promote the consistency of the distribution between the positive samples; the trained self-supervised learning network is used as the backbone network of the roadway deformation intelligent detection module to extract deep features related to roadway deformation from the input roadway point cloud data;

[0028] S32: constructing a deformation detection module to realize accurate identification of the roadway deformation region and evaluation of the deformation degree;

[0029] 3D candidate boxes are generated by using multiple full connection layers and convolutional layers, and the 3D candidate boxes are refined as the detection head of the roadway deformation intelligent detection module to realize accurate identification of the roadway deformation region and evaluation of the deformation degree according to the feature representation;

[0030] S33: using the roadway deformation intelligent detection module to perform automatic identification and evaluation of the roadway deformation region;

[0031] The roadway deformation intelligent detection module is composed of a self-supervised learning module and a deformation detection module; a three-dimensional sensor arranged in the roadway is used to automatically scan the roadway at a set sampling frequency and generate point cloud data, the point cloud data is input into the self-supervised learning module, the self-supervised learning module is used to automatically extract deep features related to roadway deformation from the point cloud data, and the obtained feature data is input into the deformation detection module, the deformation detection module is used for accurate identification of the roadway deformation area and evaluation of the deformation degree, and finally, the identified deformation area and the evaluation report are output.

[0032] As a preferred, in S31 of step three, the specific process of using a variational auto-encoding contrast learning network structure to perform self-supervised feature learning on unlabeled point cloud data is as follows:

[0033] S31-1: For the point cloud data obtained through data enhancement, two point cloud samples p and q are extracted each time for contrast learning, and the encoder of the contrast learning is G(·), and the features f and f are obtained after the samples pass through the encoder p and f q ;

[0034] S31-2: Use two single-layer fully connected networks F μ (·) and F σ (·) to project the features f p and f q to a high-dimensional invariant space respectively, so that each point cloud sample generates two vectors μ p ,σ p and μ q ,σ q , so that the latent features of the two point cloud samples follow a Gaussian distribution with mean μ and standard deviation σ:

[0035] S31-3: Use the KL divergence as a regularization term to constrain the feature distribution of all samples to a standard normal distribution with mean 0 and standard deviation 1: D KL (N(μ,σ),N(0,1)); the KL divergence is used according to formula (3) to calculate the distance between the Gaussian distributions of the two sample features, and the samples of the same category are constrained to similar Gaussian distributions, while the distributions of different categories are made as far apart as possible;

[0036]

[0037] In the formula, P(x) and Q(x) are the Gaussian distributions of the two sample features f p and f q ;

[0038] S31-4: Calculate the distance between two Gaussian distributions according to formula (4) using symmetric JS divergence to calculate the distribution similarity between the positive samples generated by two random data augmentations, and map the distance between the two samples to 0 and 1;

[0039]

[0040] where D KL (·) is the KL divergence between two distributions, the value of JS divergence is between 0 and 1, and the smaller the value represents the more similar the two distributions are;

[0041] S31-5: Make the feature distribution of the same class of samples similar, and make the feature distribution of different classes as far away as possible according to formula (5);

[0042]

[0043] where N represents the number of samples in the batch, i represents the i-th sample, k represents the k-th sample, P and Q are the Gaussian distributions of the feature of the two samples, and D JS (·) represents the JS divergence between two distributions;

[0044] S31-6: Obtain the feature distributions P(x) and Q(x) of the two samples p and q from the variational auto-encoding contrast learning network structure, and sample from P(x) and Q(x) to generate embeddings g p and g q ; input g p and g q to the decoder D(·) to generate new point clouds p' and q' to introduce randomness and make the generated point cloud data diverse;

[0045] S31-7: Normalize the point cloud p' and q' coordinates to the [0, 1] range and discretize them into a fixed size grid to convert the point cloud into a probability distribution; at the same time, also convert the input point clouds p and q into a probability distribution;

[0046] S31-8: Calculate the cross-entropy loss between the probability distribution of the generated point cloud and the probability distribution of the input point cloud according to formula (6) to evaluate the generation quality of the point cloud;

[0047]

[0048] where P(i) is the probability value of the i-th grid cell in the probability distribution of the input point cloud, and Q(i) is the probability value of the i-th grid cell in the probability distribution of the generated point cloud;

[0049] S31-9: According to formula (7), by minimizing the cross-entropy loss between the probability distribution of the generated point cloud and the probability distribution of the input point cloud, it is ensured that the generated red box cloud is highly consistent with the input point cloud in probability distribution;

[0050] min(CE(P,P') + CE(Q,Q')) (7);

[0051] In the formula, CE(P,P') is the cross-entropy loss between the input point cloud P and the generated point cloud P', and CE(Q,Q') is the cross-entropy loss between the input point cloud Q and the generated point cloud Q'.

[0052] As a preferred, in step three S32, by segmenting the point cloud into foreground points and background points, the search space is reduced, the generation efficiency of the candidate box is improved, and a small number of high-quality 3D candidate boxes are directly generated from the original point cloud, providing a basis for subsequent bounding box refinement; The specific process is as follows:

[0053] S32-11: For the input point cloud X = {x1, x2, ···, x N}, where each point x i = (x, y, z, h, w, l, θ), (x, y, z) is the center point coordinate of the point cloud, (h, w, l) is the size length width height, and θ is the direction angle; Extract the feature F = E(X) through the trained self-supervised learning network, where F = {f1, f2, ···, f N} represents the feature corresponding to each point;

[0054] S32-12: A simple fully connected network is used to predict the probability of foreground points, where the input is the feature f i of each point, and the output is the probability p i that the point is a foreground point, as shown in formula (8);

[0055] p i = σ(W s f i +b i ) (8);

[0056] In the formula, σ(·) is a Sigmoid activation function, which is used to map the output value to the [0,1] range to represent the probability, W s and b s are the weight and bias parameters of the fully connected network, respectively;

[0057] S32-13: Estimate the parameters of the 3D candidate box using bin-based regression loss for each foreground point, divide the surrounding area into multiple discrete bins, estimate the center position using classification loss and residual regression loss, and since the distribution of target objects in the vertical direction is usually concentrated and the range of change is small, the loss in the y-axis direction is not calculated;

[0058] First, the offset of the target center (x p ,z p ) relative to the foreground point p is calculated according to formulas (9) and (10) respectively, and is discretized into bins to obtain the straight value bin assignment L bin (p) x , L bin (p) z along the x-axis and z-axis directions respectively.

[0059]

[0060] where (x(p), z(p)) is the coordinate of the foreground point p, (x p ,z p ) is the calculated center point coordinate, S is the search range set between 0 and 3, and δ is the size of the bin set between 100 and 255.

[0061] Second, the residual regression res(p) x , res(p) z of the target center is calculated according to formulas (11) and (12) respectively within the determined bin.

[0062]

[0063] where C is a normalization constant set to 50 to control the range of residuals.

[0064] As a preferred, in S32 of step three, by normalizing the coordinate system and global semantic features, more accurate bounding box parameters are learned to further refine the 3D candidate box generated in the previous step to improve detection accuracy.

[0065] S32-21: Determine each 3D candidate box b i , and slightly expand its range to include more context information, as shown in formula (13).

[0066]

[0067] where (x i ,y i ,z i ) is the point cloud center point coordinate, (h i ,wi , l i ) is the size length-width-height, θ is the direction angle, η is the constant of the enlarged range, which is set to 0.5;

[0068] S32-22: For each point p, it is judged whether it is in the enlarged candidate box, if so, the point and its features are retained, the pooled point features include the coordinates of the point, the reflection intensity, the segmentation mask and the global features, the pooled point coordinates are converted to the canonical coordinate system to eliminate rotation and position changes;

[0069] S32-23: Definition of canonical coordinate system, the origin is located at the center of the candidate box; the local x-axis and z-axis are parallel to the ground plane, the x-axis points to the head direction of the candidate box, and the z-axis is perpendicular to the x-axis; the y-axis is consistent with the y-axis of the LiDAR coordinate system;

[0070] S32-24: The converted local space features and global semantic features f(p) are combined, encoded into a unified feature vector through a fully connected layer, and these features are used for boundary box refinement and confidence prediction, finally, the boundary box refinement is performed using bin-based regression loss using formula (14);

[0071]

[0072] In the formula, L is the total loss of 3D boundary box refinement, B is the set of all 3D candidate boxes, |B| is the number of candidate boxes, Fcls(·) is a classification loss function, which is used to calculate the difference between the predicted confidence and the true label, prob i is the predicted confidence of the i-th candidate box, label i is the true label of the i-th candidate box, B pos is the set of positive sample candidate boxes, |B pos | is the number of positive sample candidate boxes, is the bin classification loss of the i-th positive sample candidate box, is the bin-based residual regression loss of the i-th positive sample candidate box.

[0073] As a preferred, in step one, the simulated generated roadway deformation point cloud data P f is obtained through sinking deformation, extrusion deformation, expansion deformation and tilt deformation operations.

[0074] Wherein, in the sinking deformation process, the sinking deformation is simulated by translation along the z-axis to generate the roadway deformation point cloud data P f , T1 is the transformation matrix of the lower layer deformation, s is the lower layer quantity, which is a random value set between 0-50;

[0075] During the extrusion deformation process, extrusion deformation is simulated by scaling in the x and y axes, and tunnel deformation point cloud data P is generated. f P f =P·T2, where T2 is the transformation matrix for extrusion deformation. α is the squeezing factor, which is a random value set between 0 and 1;

[0076] During the expansion deformation process, the expansion deformation is simulated by scaling up the entire point cloud proportionally, and the tunnel deformation point cloud data P is generated. f P f =P·T3, where T3 is the transformation matrix for the extended transformation. β is the expansion factor, which is a random value set between 0 and 1;

[0077] During the tilting deformation process, rotation is used to simulate the tilting deformation and generate tunnel deformation point cloud data P. f P f =P·T4, where T4 is the transformation matrix for tilt deformation. θ is the tilt angle, which is a random value set between 0 and 30 degrees.

[0078] As a preferred embodiment, in step two, the deformed roadway point cloud data Pf is enhanced through rotation transformation, translation transformation, scaling transformation, and noise addition. f ={(x i ′,y i ′,z i ′)|i=1,2,3,....,N0},(x i ′,y i ′,z i ′) represents the coordinates of the i-th 3D point in the generated tunnel point cloud data, and N0 is the total number of points in the generated tunnel point cloud;

[0079] In the rotation transformation process, the point cloud is rotated around the z-axis, and each point in the point cloud (x... i ′,y i ′,z i After rotation, the new point coordinates are obtained as follows: The rotation transformation formula is: Where θ1 is the rotation angle and R is the rotation matrix.

[0080] In the translation transformation process, translation is achieved by adding translation amounts along the x, y, and z axes. The translation transformation formula is as follows: Where t is the translation vector, t = (t x ,t y ,t z ), tx y z respectively represent the translation amount along the x, y, z axes;

[0081] In the scaling transformation process, multiply each coordinate of the point cloud by a scaling factor s x y z to realize scaling, and the scaling transformation formula is:

[0082] In the noise adding process, a random noise N(mu, sigma 2 ) of Gaussian distribution is added to each coordinate of the point cloud, and the noise adding formula is: Where mu is the mean and sigma is the standard deviation.

[0083] The present application proposes a kind of based on self-supervised learning coal mine roadway deformation detection basic model detection method, to solve the problem of data scarcity and annotation difficulty in coal mine roadway deformation monitoring. First, by simulating different types of roadway deformation to generate three-dimensional point cloud data with diversity, and further improve the richness of data and the robustness of model by combining data enhancement technology. Then, using self-supervised learning network, through variational auto-encoding contrast learning and cross-entropy loss supervision module, automatically extract features from unlabeled point cloud data, avoid complex manual annotation process. On this basis, a roadway deformation detection network is constructed, which realizes accurate identification of roadway deformation area and evaluation of deformation degree through detection head.

[0084] Compared with the prior art, the present application has the following advantages:

[0085] 1、The present application adopts self-supervised learning method, and features are learned from unlabeled three-dimensional point cloud data, which greatly reduces the demand for a large amount of labeled data. Compared with the traditional method of relying on a large amount of manually annotated data for training, the present application can accurately and efficiently detect roadway deformation based on a small amount of labeled data, reducing the cost and workload of data annotation.

[0086] 2、Through the deep feature extraction capability of self-supervised learning network, the present application can more accurately capture the details and structural changes of roadway deformation. Combined with the high-quality features obtained by self-supervised learning, the present application can effectively deal with various complex situations in the roadway environment, improve the accuracy and robustness of deformation detection, especially in different types and different sizes of roadway.

[0087] ​​​​3、The self-supervised learning network is fine-tuned by using a small amount of labeled data set, which can quickly adapt to different roadway environments and deformation types in practical applications. This method not only improves the adaptability of the model, but also ensures its real-time performance and accuracy in practical scenarios, so that the network can be efficiently deployed and updated online.

[0088] In summary, the technology not only reduces the dependence on a large amount of labeled data, but also improves the generalization ability and detection accuracy of the model in complex environments, providing an efficient and intelligent solution for coal mine roadway safety monitoring, which has important practical application value. This method can effectively solve the problems of lack of deformation data and complex point cloud data labeling in underground roadway, thereby improving the accuracy and efficiency of deformation detection, reducing the dependence on manual labeling, and significantly improving the ability of coal mine safety monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 is the principle diagram of the model part of the present application;

[0090] Figure 2 is the flow chart of the method part of the present application;

[0091] Figure 3 is the structure diagram of the self-supervised learning acting on the downstream detection task in the method part of the present application;

[0092] Figure 4 is the structure diagram of the variational auto-encoding contrast learning network in the method part of the present application;

[0093] Figure 5 is the structure diagram of the cross-entropy loss supervision network in the method part of the present application;

[0094] Figure 6 is the structure diagram of the deformation detection network in the method part of the present application. DETAILED DESCRIPTION

[0095] The application provides a coal mine roadway deformation detection basic model based on self-supervised learning, which comprises a roadway deformation point cloud data generation module, a data enhancement module, a self-supervised learning module and a deformation detection module.

[0096] The framework flowchart of the application is shown in Figure 1 and Figure 2 In order to more clearly illustrate the object, technical scheme and advantages of the application, the application will be further described in combination with the accompanying drawings and specific implementation steps. Figures 1 to 6

[0097] The application provides a coal mine roadway deformation detection basic model based on self-supervised learning, which comprises a roadway deformation point cloud data generation module, a data enhancement module and a roadway deformation intelligent detection module.

[0098] The roadway deformation point cloud data generation module is used for generating deformed three-dimensional point cloud data based on the labeled coal mine roadway original point cloud data and according to the roadway geometric structure and different deformation types, and then sending the three-dimensional point cloud data to the data enhancement module.

[0099] The data enhancement module is used for deforming the three-dimensional point cloud data in different degrees, introducing noise and disturbance to simulate the actual data, expanding the diversity of the data set, obtaining the expanded data set, and then sending the expanded data set to the self-supervised learning module.

[0100] The roadway deformation intelligent detection module comprises a self-supervised learning module and a deformation detection module.

[0101] The self-supervised learning module is used for automatically extracting features from the unlabeled point cloud data in the expanded data set by variational auto-encoding contrast learning and cross-entropy loss supervision, and sending the obtained feature data to the deformation detection module.

[0102] ​The deformation detection module is used for accurately identifying the deformation region of the roadway according to the feature data, and accurately evaluating the severity of deformation.

[0103] As a preferred, the different deformation types include subsidence deformation, extrusion deformation, expansion deformation and inclination deformation.

[0104] As a preferred, the data deformation processing includes rotation transformation, translation transformation, scaling transformation and noise addition.

[0105] In the present application, through the setting of the roadway deformation point cloud data generation module, roadway point cloud data with different deformation types can be generated by simulation method, so that a large amount of simulated three-dimensional roadway deformation data can be obtained, and sufficient sample data is provided; through the setting of the data enhancement module, the original point cloud data set can be fully enhanced by rotation, translation, scaling, noise addition operation, providing more abundant and diversified samples for training model, which is beneficial to obtain detection model with higher detection precision. Through the setting of the self-supervised learning module, technical features related to roadway deformation can be automatically and quickly extracted from unlabeled point cloud data, which can provide reliable technical support for subsequent intelligent detection process. Through the setting of the deformation detection module, the roadway deformation region and deformation condition can be accurately identified according to the feature data, and deformation program evaluation work can be carried out. The model has simple structure, which can effectively improve the intelligent degree of roadway deformation detection, can effectively improve the accuracy of mine environment monitoring and risk prediction, and provides an efficient and intelligent solution for safety monitoring of coal mine roadway.

[0106] The present application also provides a detection method of a coal mine roadway deformation detection basic model based on self-supervised learning.

[0107] Step one: using the roadway deformation point cloud data generation module to generate roadway point cloud data with different deformation types by simulation method;

[0108] The existing coal mine roadway point cloud data is used to generate simulated roadway deformation point cloud data, and the deformation roadway point cloud data generation module generates roadway point cloud data with different deformation types by simulation method; for the labeled coal mine roadway original point cloud data P, according to the roadway geometric structure and different deformation types such as subsidence, extrusion, inclination and expansion, and combining physical model and numerical calculation, the deformation roadway point cloud data P is obtained by using formula (1) f ; wherein P={(x i ,y i ,z i )|i=1,2,3,....,N}, (x i ,y iz i ) represents the i-th three-dimensional point coordinate in the roadway, and N is the total number of points in the point cloud;

[0109] P f = P · T + b (1);

[0110] In the formula, f represents the deformation type, T represents the transformation matrix related to the deformation type, and b represents the translation offset;

[0111] After the above deformation processing, the simulated roadway deformation three-dimensional point cloud data can be generated.

[0112] Step two: using a data enhancement module to perform deformation processing on the deformation roadway point cloud data;

[0113] The deformation roadway point cloud data P f is enhanced, and at the same time, noise and disturbance are introduced to realistically simulate actual data, and the point cloud data P aug after enhancement is obtained by using formula (2); the data enhancement operation can transform the point cloud through rotation, translation, scaling and the like, thereby increasing the diversity of the data and improving the robustness of the model. At the same time, noise information is added to simulate the error in the actual collection process, and the authenticity of the training data is increased. Through these enhancement techniques, the trained model can better adapt to complex actual environments and improve its generalization ability on real data.

[0114] P aug = (P f · R) + t · (1 + ε) (2);

[0115] In the formula, R is an enhanced transformation matrix, including rotation, translation, scaling and noise addition, etc.; t is a translation vector; ε represents the noise added by randomization, which can be Gaussian noise or discrete noise, and (1 + ε) represents the scaling factor of the noise change added by randomization;

[0116] Through the above operation, the generated enhanced data set P aug will have more deformation samples.

[0117] Step three: constructing a roadway deformation intelligent detection module and performing automatic identification and evaluation of the roadway deformation region;

[0118] S31: constructing a self-supervised learning module to automatically extract features from unlabeled point cloud data;

[0119] In view of the problem that three-dimensional point cloud data is difficult to label, a three-dimensional point cloud self-supervised learning network is constructed by using a method combining variational autoencoder and contrastive learning and simultaneously increasing cross-entropy loss supervision, the network mainly adopts a contrastive learning method to learn feature representation of point cloud data, and a variational autoencoder contrastive learning and cross-entropy loss supervision module are combined to design a network structure, so that the robustness and generalization ability of the model are enhanced. Figure 3 To achieve this object without labels, a specific pre-task needs to be designed to help the point cloud encoder extract enough information from the data to improve its encoding ability. Contrastive learning is a widely used pre-task. Contrastive learning is a special self-supervised learning. Unlike the encoding and decoding scheme of generative learning, it focuses on learning the common features between instances of the same class and distinguishing the differences between different classes.

[0120] The variational autoencoder and contrastive learning are combined to form a variational autoencoder contrastive learning network structure, and the characteristic of the variational autoencoder contrastive learning is to constrain the feature distribution instead of the feature value corresponding to each sample in the latent space. The unlabelled point cloud data is subjected to self-supervised feature learning by using the variational autoencoder contrastive learning network structure; the latent feature representation of the point cloud data is extracted by using the variational autoencoder, and through contrastive learning, the distribution of the positive sample features extracted by the encoder converges to be similar, while distinguishing the distribution between the positive sample and the negative sample; the purpose of the variational contrastive learning is to perform self-supervised feature learning on the point cloud data, while avoiding the hard constraint on the feature value in the traditional contrastive learning, so as to improve the generalization ability of the encoder, and the variational autoencoder contrastive learning network structure is as shown in the following Figure 4 Unlike the classic contrastive learning method which directly learns the feature value of the sample, the variational contrastive learning improves the robustness of the model by constraining the feature distribution in the latent space.

[0121] Contrastive learning may cause the model to pay excessive attention to specific key parts of the sample, affecting the generalization ability. Therefore, generation supervision is introduced to retain invariant features. The traditional self-reconstruction method focuses on global coordinate features, which reduces the generalization ability of the point cloud feature extractor. In order to solve this problem, a cross-entropy loss function is introduced to supervise the training process, a generative adversarial mechanism is introduced, the generated latent features are optimized through contrastive learning, so that the model can effectively capture the essential features in the point cloud data, thereby improving the performance of self-supervised learning. At the same time, the cross-entropy loss is calculated by using two different data augmentations to generate positive samples, so as to retain more invariant features and promote the consistency of the distribution between positive samples; the target of the cross-entropy loss supervision is to maximize the similarity of the positive sample pairs by the generator, and to maintain the distinguishability of different categories of samples in the latent space. The cross-entropy loss supervision network structure is as followsFigure 5 The trained self-supervised learning network is used as the backbone network of the tunnel deformation intelligent detection module to extract deep features related to tunnel deformation from the input tunnel point cloud data, and is specifically responsible for extracting deep features from the input tunnel point cloud data, which effectively capture the spatial layout, structural changes and deformation patterns of the tunnel.

[0122] S32: Construct a deformation detection module to accurately identify the deformation area of the tunnel and evaluate the deformation degree;

[0123] The self-supervised learning network model can learn robust feature representations from unlabeled point cloud data, which can capture the geometry and topology information of the point cloud, providing a high-quality feature basis for subsequent deformation detection tasks. After pre-training the self-supervised learning network, the weights of the encoder E(·) are saved, which will be used as the initialization weights for subsequent tasks. The pre-trained self-supervised learning encoder E(·) is used as the backbone network of the deformation detection network, which is responsible for extracting high-dimensional feature representations from the input point cloud. On the basis of the backbone network, a detection head is added, which is responsible for detecting the deformation area of the tunnel from the feature representation. The detection head mainly consists of two stages: 3D candidate box generation and 3D candidate box refinement.

[0124] After training the self-supervised learning network, a deep feature representation that can effectively extract features from point cloud data is obtained. Next, these feature representations obtained by self-supervised learning are applied to downstream tasks to construct a tunnel deformation detection network. The goal of this detection network is to accurately identify and detect the deformation of the tunnel based on the extracted point cloud features. The structure of the deformation detection network is as follows Figure 6 To improve the accuracy and robustness of the network and enhance the expression ability of the network, a multi-layer fully connected layer (MLP) and a convolution layer are used to generate 3D candidate boxes and refine 3D candidate boxes as the detection head of the tunnel deformation intelligent detection module, to accurately identify the deformation area of the tunnel and evaluate the deformation degree based on the feature representation. The detection head uses classification or regression methods to determine whether the tunnel has deformed and further identifies the type and degree of deformation to perform the tunnel deformation detection task. Finally, by optimizing the loss function of the overall network and combining the features obtained by self-supervised learning, the detection head can accurately identify the deformation of the tunnel and accurately evaluate it.

[0125] S33: Use the tunnel deformation intelligent detection module to automatically identify and evaluate the deformation area of the tunnel;

[0126] The roadway deformation intelligent detection module is composed of a self-supervised learning module and a deformation detection module; a three-dimensional sensor arranged in the roadway is used to automatically scan the roadway at a set sampling frequency and generate point cloud data, the point cloud data is input into the self-supervised learning module, the self-supervised learning module is used to automatically extract deep features related to roadway deformation from the point cloud data, and the obtained feature data is input into the deformation detection module, the deformation detection module is used for accurate identification of the roadway deformation area and evaluation of the deformation degree, and finally, the identified deformation area and the evaluation report are output.

[0127] As a preferred, in S31 of step three, the specific process of self-supervised feature learning of unlabeled point cloud data by using a variational auto-encoding contrast learning network structure is as follows:

[0128] S31-1: For the point cloud data obtained by data enhancement, two point cloud samples p and q are extracted for contrast learning each time, the encoder of contrast learning is G(·), and the features f p and f q are obtained after the samples pass through the encoder;

[0129] S31-2: Use two single-layer fully connected networks F μ (·) and F σ (·) to project the features f p and f q to a high-dimensional invariant space, respectively, so that each point cloud sample generates two vectors μ p ,σ p and μ q ,σ q , so that the latent features of the two point cloud samples follow a Gaussian distribution with mean μ and standard deviation σ:

[0130] S31-3: Use KL divergence as a regularization term to constrain the feature distribution of all samples to a standard normal distribution with mean 0 and standard deviation 1: D KL (N(μ,σ),N(0,1)); according to formula (3), the distance between the Gaussian distributions of the two sample features is calculated using KL divergence, and the samples of the same category are constrained to similar Gaussian distributions, while the distributions of different categories are as far apart as possible, which helps the model to generate samples similar to the training set;

[0131]

[0132] In the formula, P(x) and Q(x) are the Gaussian distributions of the two sample features f p and f q ;

[0133] S31-4: In order to maintain the distinguishing ability of the feature extractor, all samples are not expected to conform to the same distribution, while in contrast learning, the distribution similarity between positive samples generated by two random data augmentations needs to be calculated, and the similarity should be sequential independent. Therefore, the distance between two Gaussian distributions is calculated according to formula (4) using the symmetric JS divergence to calculate the distribution similarity between positive samples generated by two random data augmentations, and the distance between two samples is mapped to between 0 and 1;

[0134]

[0135] In the formula, D KL (·) is the KL divergence between two distributions, and the value of JS divergence is between 0 and 1, and the smaller the value represents the more similar the two distributions are;

[0136] S31-5: According to formula (5), make the feature distribution of samples in the same category similar, and make the feature distribution of samples in different categories as far away as possible;

[0137]

[0138] In the formula, N represents the number of samples in the batch, i represents the i-th sample, k represents the k-th sample, P and Q are the Gaussian distributions of the features of the two samples, and D JS (·) represents the JS divergence between two distributions;

[0139] S31-6: Obtain the feature distributions P(x) and Q(x) of the two samples p and q from the variational auto-encoding contrast learning network structure, and sample from P(x) and Q(x) to generate embeddings g p and g q ; input g p and g q to the decoder D(·) to generate new point clouds p' and q'. By sampling from the feature distribution, randomness can be introduced and the generated point cloud data has diversity, which helps the model to learn the latent distribution of the input point cloud, rather than just a single reconstruction. The decoder converts the feature embedding back to the point cloud space to generate an output similar to the input point cloud. This is a key step in evaluating the model's generation ability.

[0140] S31-7: In order to represent the point cloud as a comparable form, the coordinates of the point clouds p' and q' are normalized to the range [0, 1] and discretized to a fixed size grid to convert the point cloud to a probability distribution, which can intuitively represent the density and distribution of the point cloud, and facilitate the calculation of the difference between the generated point cloud and the input point cloud. At the same time, the input point clouds p and q are also converted to probability distributions;

[0141] S31-8: Calculate the cross-entropy loss between the probability distribution of the generated point cloud and the probability distribution of the input point cloud according to formula (6) to evaluate the generation quality of the point cloud.

[0142]

[0143] In the formula, P(i) is the probability value of the i-th grid cell in the probability distribution of the input point cloud, and Q(i) is the probability value of the i-th grid cell in the probability distribution of the generated point cloud.

[0144] S31-9: Cross-entropy loss directly measures the similarity between the generated point cloud and the input point cloud. According to formula (7), by minimizing the cross-entropy loss, it can be ensured that the generated point cloud is highly consistent with the input point cloud in terms of probability distribution. Finally, the goal of the entire cross-entropy loss supervision module is to minimize the cross-entropy loss between the probability distribution of the generated point cloud and the probability distribution of the input point cloud.

[0145] min(CE(P,P′)+CE(Q,Q′))(7);

[0146] In the formula, CE(P,P′) is the cross-entropy loss between the input point cloud P and the generated point cloud P′, and CE(Q,Q′) is the cross-entropy loss between the input point cloud Q and the generated point cloud Q′.

[0147] As a preferred approach, in step S32 of step three, the point cloud is segmented into foreground and background points to reduce the search space and improve the efficiency of candidate box generation. A small number of high-quality 3D candidate boxes are directly generated from the original point cloud, providing a foundation for subsequent bounding box refinement. The specific process is as follows:

[0148] S32-11: For the input point cloud X = {x1, x2, ..., x...} N}, where each point x i = (x,y,z,h,w,l,θ), where (x,y,z) are the coordinates of the center point of the point cloud, (h,w,l) are the dimensions (length, width, height), and θ is the orientation angle; features F = E(X) are extracted through a trained self-supervised learning network, where F = {f1,f2,...,f...} N} represents the feature corresponding to each point;

[0149] S32-12: Predict the probability of foreground points using a simple fully connected network, where the input is the feature f of each point. i The output is the probability p that the point is a foreground point. i As shown in formula (8);

[0150] p i =σ(W s f i +bi ) (8);

[0151] where σ(·) is a Sigmoid activation function to map the output value to a probability in the range of [0, 1], W s and b s are the weight and bias parameters of the fully connected network, respectively.

[0152] S32-13: Estimate the parameters of the 3D candidate box using a bin-based regression loss for each foreground point, divide the surrounding area into multiple discrete bins, and use a classification loss and a residual regression loss to estimate the center position. Since the distribution of target objects in the vertical direction is usually concentrated and has a small range of variation, the loss in the y-axis direction is not calculated.

[0153] First, calculate the offset of the target center (x p ,z p ) relative to the foreground point p according to formulas (9) and (10), respectively, and discretize it into bins to obtain the straight value bin assignments L bin (p) x , L bin (p) z along the x-axis and z-axis directions, respectively.

[0154]

[0155] where (x(p), z(p)) is the coordinate of the foreground point p, (x p ,z p ) is the calculated center point coordinate, S is the search range set between 0 and 3, and δ is the size of the bin set between 100 and 255.

[0156] Second, calculate the residual regression res(p) x , res(p) z of the target center within the determined bin according to formulas (11) and (12), respectively.

[0157]

[0158] where C is a normalization constant set to 50 to control the range of residuals.

[0159] As a preferred, in S32 of step three, by normalizing the coordinate system and global semantic features, more accurate bounding box parameters are learned to further refine the 3D candidate box generated in the previous step to improve detection accuracy.

[0160] S32-21: Determine the 3D candidate box b iand slightly expand its range to contain more context information, as shown in formula (13);

[0161]

[0162] where (x i ,y i ,z i ) is the center point coordinate of the point cloud, (h i ,w i ,l i ) is the size length width height, θ is the direction angle, η is the constant of the expanded range, set to 0.5;

[0163] S32-22: For each point p, judge whether it is in the expanded candidate box, if so, keep the point and its features, the pooled point features include the point coordinates, reflection intensity, segmentation mask and global features, convert the pooled point coordinates to the canonical coordinate system to eliminate rotation and position changes;

[0164] S32-23: Definition of canonical coordinate system, the origin is located at the center of the candidate box; the local x-axis and z-axis are parallel to the ground plane, the x-axis points to the head direction of the candidate box, and the z-axis is perpendicular to the x-axis; the y-axis is consistent with the y-axis of the LiDAR coordinate system;

[0165] S32-24: Combine the converted local space features and global semantic features f(p) to encode into a unified feature vector through a fully connected layer, use these features to refine the bounding box and predict the confidence, finally, use formula (14) to use bin-based regression loss to refine the bounding box;

[0166]

[0167] where L is the total loss of 3D bounding box refinement, B is the set of all 3D candidate boxes, |B| is the number of candidate boxes, Fcls(·) is the classification loss function, which is used to calculate the difference between the predicted confidence and the true label, prob i is the predicted confidence of the i-th candidate box, label i is the true label of the i-th candidate box (1 for positive samples and 0 for negative samples), B pos is the set of positive sample candidate boxes (i.e. candidate boxes with large IoU with the true bounding box), |B pos | is the number of positive sample candidate boxes, is the bin classification loss of the i-th positive sample candidate box, is the bin-in residual regression loss of the i-th positive sample candidate box. This combination method makes full use of the feature extraction capability of the self-supervised learning network and the powerful detection capability of the deformation detection network, significantly improving the performance of the roadway deformation detection.

[0168] As a preferred, in step one, the simulated generated roadway deformation point cloud data P is obtained by sinking deformation, extrusion deformation, expansion deformation and tilt deformation operations f ;

[0169] In the sinking deformation process, the sinking deformation is simulated by translation along the z-axis, and the roadway deformation point cloud data P is generated f , T1 is the transformation matrix of the lower layer deformation, s is the lower layer quantity, which is a random value set between 0 and 50;

[0170] In the extrusion deformation process, the extrusion deformation is simulated by scaling in the x-axis and y-axis directions, and the roadway deformation point cloud data P is generated f , P f = P·T2, T2 is the transformation matrix of the extrusion deformation, α is the extrusion factor, which is a random value set between 0 and 1;

[0171] In the expansion deformation process, the expansion deformation is simulated by scaling the entire point cloud, and the roadway deformation point cloud data P is generated f , P f = P·T3, T3 is the transformation matrix of the expansion deformation, β is the expansion factor, which is a random value set between 0 and 1;

[0172] In the tilt deformation process, the tilt deformation is simulated by rotation, and the roadway deformation point cloud data P is generated f , P f = P·T4, T4 is the transformation matrix of the tilt deformation, θ is the tilt angle, which is a random value set between 0 and 30 degrees;

[0173] Through the above process, a large amount of simulated three-dimensional roadway deformation data is obtained based on the original point cloud data after batch transformation processing, thereby having sufficient sample data.

[0174] As a preferred, in step two, the deformation roadway point cloud data Pf is enhanced by rotation transformation, translation transformation, scaling transformation and noise addition, wherein P f = {(x i ′,y i ′,z i ′)|i=1,2,3,....,N0}, (xi ′,y i ′,z i ′) represents the i-th three-dimensional point coordinate in the generated roadway point cloud data, N0 is the total number of points in the generated roadway point cloud; the main purpose of the data enhancement module is to increase the diversity of data by performing various transformations on the generated roadway point cloud data, thereby improving the robustness and generalization ability of the model.

[0175] In the rotation transformation process, the point cloud is rotated around the z-axis, and the point (x i ′,y i ′,z i ′) in each point cloud is rotated to obtain new point coordinates The rotation transformation formula is: Where θ1 is the rotation angle, R is the rotation matrix, This rotation operation can rotate the point cloud in the z-axis plane, and different angle rotation data can be generated by randomly selecting θ1, θ1 is a random value set between 0 and 30 degrees

[0176] In the translation transformation process, translation is performed by adding translation amounts on the x, y, and z axes, and the translation transformation formula is: Where t is the translation vector, t = (t x ,t y ,t z ), t x ,t y ,t z represent the translation amounts along the x, y, and z axes, respectively; the translation amount is randomly generated in the range of 0 to 50.

[0177] In the scaling transformation process, each coordinate of the point cloud is multiplied by a scaling factor s x ,s y ,s z to realize scaling, and the scaling transformation formula is: This scaling operation can realize uniform scaling or non-uniform scaling of the point cloud using different scaling factors on the x, y, and z axes, where the scaling factor s x ,s y ,s z is a random value set between 0 and 1.

[0178] Point cloud data is usually affected by noise, so it is necessary to introduce noise in the data enhancement process. Noise can simulate errors in the actual acquisition process, such as scanner accuracy errors, environmental interference, etc. The present application mainly uses Gaussian noise to enhance the data set. In the noise addition process, a random noise N(μ,σ 2), N(0, σ 2 ) is a Gaussian noise with mean 0 and standard deviation σ, which can simulate small random fluctuations in the point cloud due to sensor errors or other factors. The noise addition formula is:

[0179] where μ is the mean and σ is the standard deviation.

[0180] The data augmentation module fully enhances the original point cloud dataset through the above rotation, translation, scaling, and noise addition operations, providing more diverse and rich samples for training the model.

[0181] In summary, the detection method of the coal mine roadway deformation detection basic model based on self-supervised learning proposed by the present application forms a complete intelligent monitoring solution through simulation of deformation data generation, data augmentation, self-supervised learning network construction, and optimization of the deformation detection network. This method not only effectively solves the problem of scarcity of coal mine roadway deformation data and complex labeling, but also significantly improves the generalization ability and detection accuracy of the model through self-supervised learning technology. In addition, through the effective combination of data augmentation and variational auto-encoding contrastive learning, it can adapt to different types of roadway deformation and show good robustness and adaptability. In practical applications, this method can significantly reduce the cost of manual labeling and significantly improve the monitoring efficiency, providing an efficient and reliable technical means for coal mine safety monitoring. With the development of mine intelligence, the present application is expected to be widely applied in more mine scenes and provide strong technical support for the safety production of mines.

Claims

1. A method for detecting deformation in coal mine roadways based on self-supervised learning, comprising the following steps: Step 1: Use the tunnel deformation point cloud data generation module to generate tunnel point cloud data with different deformation types through simulation methods; For the labeled raw point cloud data of coal mine roadways Based on the roadway geometry and different deformation types, and combined with physical models and numerical calculations, point cloud data of deformed roadways are obtained using formula (1). ;in, , Indicates the first in the alleyway Three-dimensional point coordinates, N The total number of points in the point cloud; （1）； In the formula, Indicates the deformation type. This represents the transformation matrix related to the deformation type. Indicates the translation offset; Step 2: Use the data augmentation module to deform the point cloud data of the deformed roadway; Point cloud data of deformed tunnels Enhancement processing is performed, and noise and disturbances are introduced to realistically simulate the actual data. The enhanced point cloud data is obtained using formula (2). ; （2）； In the formula, It is an enhancement transformation matrix; It is a translation vector; This represents the noise added during randomization. This represents the scaling factor used to introduce noise variations through randomization. Step 3: Construct an intelligent detection module for tunnel deformation and perform automated identification and assessment of tunnel deformation areas; S31: Construct a self-supervised learning module for automatically extracting features from unlabeled point cloud data; By combining variational autoencoders and contrastive learning, a variational autoencoder-contrast learning network structure is formed, which is then used to perform self-supervised feature learning on unlabeled point cloud data. A variational autoencoder is used to extract latent feature representations from point cloud data. Through contrastive learning, the distribution of positive sample features extracted by the encoder converges to similarity, while distinguishing the distribution between positive and negative samples. Simultaneously, a cross-entropy loss function is introduced to supervise the training process, and a generative adversarial mechanism is introduced to optimize the generated latent features through contrastive learning. This enables the model to effectively capture the essential features in the point cloud data. Furthermore, cross-entropy loss is calculated using positive samples augmented with two different datasets to retain more invariant features and promote consistency in the distribution of positive samples. The trained self-supervised learning network is used as the backbone network of the intelligent tunnel deformation detection module to extract depth features related to tunnel deformation from the input tunnel point cloud data. S32: Construct a deformation detection module to achieve accurate identification of roadway deformation areas and assessment of the degree of deformation; 3D candidate boxes are generated using multi-layer fully connected layers and convolutional layers, and the 3D candidate boxes are refined as the detection head of the intelligent detection module for roadway deformation, so as to achieve accurate identification of roadway deformation areas and assessment of the degree of deformation based on latent feature representation. S33: Automated identification and assessment of roadway deformation areas using a roadway deformation intelligent detection module; A roadway deformation intelligent detection module is composed of a self-supervised learning module and a deformation detection module. Using 3D sensors deployed in the roadway, the module automatically scans the roadway at a set sampling frequency and generates point cloud data. This point cloud data is input into the self-supervised learning module, which automatically extracts depth features related to roadway deformation from the point cloud data. The obtained feature data is then input into the deformation detection module, which accurately identifies the roadway deformation area and assesses the degree of deformation. Finally, the module outputs the identified deformation area and an assessment report.

2. The method for detecting deformation of coal mine roadways based on self-supervised learning according to claim 1, characterized in that, In step S31 of step three, the specific process of performing self-supervised feature learning on the unlabeled point cloud data using the variational autoencoder contrastive learning network structure is as follows: S31-1: For point cloud data obtained through data augmentation, extract two point cloud samples each time. and Comparative learning is performed, and the encoder for comparative learning is... The sample obtains features after passing through the encoder. and ; S31-2: Using two single-layer fully connected networks and Features and Projecting onto a high-dimensional invariant space, so that each point cloud sample generates two vectors. and Let the latent feature representations of the two point cloud samples follow the following Gaussian distribution: , ; S31-3: Using KL divergence as a regularization term, constrain the feature distribution of all samples to a standard normal distribution with a mean of 0 and a standard deviation of 1: According to formula (3), KL divergence is used to calculate the distance between the Gaussian distributions of two sample features, and samples of the same class are constrained to be similar Gaussian distributions, while the distributions of different classes are made as far apart as possible. （3）； In the formula, and These are two sample features. and Gaussian distribution; S31-4: The distance between two Gaussian distributions is calculated using the symmetric JS divergence according to formula (4) to calculate the distribution similarity between positive samples generated by two random data augmentations, and the distance between positive samples generated by two random data augmentations is mapped to between 0 and 1; （4）； In the formula, It calculates the KL divergence between two distributions. The value of the JS divergence is between 0 and 1. The smaller the value, the more similar the two distributions are. S31-5: According to formula (5), make the feature distributions of samples of the same category similar, and make the feature distributions of samples of different categories as far apart as possible; （5）； In the formula, Indicates the number of samples in the batch. Indicates the first sample, Indicates the first One sample, and These are Gaussian distributions of the features of the two samples, respectively. This indicates the calculation of the JS divergence between two distributions; S31-6: Obtain two samples from the variational autoencoder contrastive learning network structure. and The corresponding latent feature Gaussian distribution and and from and Mid-sampling generates embeddings and ;Will and Input to decoder Generate new point clouds and This introduces randomness and makes the generated point cloud data more diverse; S31-7: Point Cloud and The coordinates are normalized to the range [0, 1] and discretized into a fixed-size grid to convert the point cloud into a probability distribution; simultaneously, the input point cloud is... and It is also converted into a probability distribution; S31-8: Calculate the cross-entropy loss between the probability distribution of the generated point cloud and the probability distribution of the input point cloud according to formula (6) to evaluate the generation quality of the point cloud. （6）； In the formula, It is the th in the probability distribution of the input point cloud The probability value of each grid cell. It is the th in the probability distribution of generating point clouds The probability value of each grid cell; S31-9: According to formula (7), the generated point cloud is highly consistent with the input point cloud in terms of probability distribution by minimizing the cross-entropy loss between the probability distribution of the generated point cloud and the probability distribution of the input point cloud; （7）； In the formula, It is an input point cloud probability distribution and generating point clouds probability distribution Cross-entropy loss between Input point cloud samples probability distribution and generating point clouds probability distribution Cross-entropy loss between them.

3. The method for detecting deformation in coal mine roadways based on self-supervised learning according to claim 2, characterized in that, In step S32 of step three, the point cloud is segmented into foreground and background points to reduce the search space and improve the efficiency of candidate box generation. A small number of high-quality 3D candidate boxes are directly generated from the original point cloud, providing a foundation for subsequent bounding box refinement. The specific process is as follows: S32-11: For input point cloud , where each point , These are the coordinates of the center point of the point cloud. It refers to the dimensions: length, width, and height. It refers to direction and angle; features are extracted through a well-trained self-supervised learning network. ,in, This represents the feature corresponding to each point; S32-12: Predicting the probability of foreground points using a fully connected network, where the input is the features of each point. The output is the probability that the point is a foreground point. As shown in formula (8); （8）； In the formula, It is the Sigmoid activation function, used to map output values ​​to a probability representation within the range [0,1]. and These are the weights and bias parameters of a fully connected network; S32-13: For each foreground point, bin-based regression loss is used to estimate the parameters of the 3D candidate bounding box. The surrounding region is divided into multiple discrete bins, and classification loss and residual regression loss are used to estimate the center position. Since the target object is relatively concentrated in the vertical direction and has a small range of variation, no further classification loss is used. Losses in the axial direction are calculated; First, calculate the target center according to formulas (9) and (10) respectively. Compared to the foreground The offset is then discretized into bins to obtain the values ​​along the [path]. axis, bin assignment values ​​in the axial direction , ; （9）； （10）； In the formula, It is the front attraction. coordinates This calculates the coordinates of the center point, and S is the search range, set between 0 and 3. This refers to the size of the bin, set to between 100 and 255. Secondly, the residual regression of the target center is calculated within the defined bin according to formulas (11) and (12), respectively. , ; （11）； （12）； In the formula, C is the normalization constant, set to 50, which is used to control the range of the residuals.

4. The method for detecting deformation of coal mine roadways based on self-supervised learning according to claim 3, characterized in that, In step S32 of step three, more accurate bounding box parameters are learned by standardizing the coordinate system and global semantic features, and the generated 3D candidate boxes are further refined to improve detection accuracy. S32-21: Determine each 3D candidate bounding box And expand its scope to include more contextual information, as shown in formula (13); （13）； In the formula, These are the coordinates of the center point of the point cloud. It refers to the dimensions: length, width, and height. It refers to direction and angle. This is a constant that expands the range, set to 0.5; S32-22: For each 3D point in the input point cloud It determines whether the point is within the expanded candidate box. If so, it retains the point and its features. The pooled point features include the point's coordinates, reflection intensity, segmentation mask, and global features. The pooled point coordinates are transformed to the normal coordinate system to eliminate rotation and position changes. S32-23: Definition of the standard coordinate system, with the origin located at the center of the candidate box; shaft and The axis is parallel to the ground plane. The axis points towards the head of the candidate box. Axis perpendicular to axis; S32-24: Transform the local spatial features and global semantic features Combined, the bounding boxes are encoded into a unified feature vector through a fully connected layer. The unified feature vector is used to refine the bounding boxes and predict confidence. Finally, the bounding boxes are refined using the bin-based regression loss according to formula (14). （14）； In the formula, is the total loss for refining the 3D bounding boxes, and B is the set of all 3D candidate boxes. It is the number of candidate boxes. It is a classification loss function used to calculate the difference between the predicted confidence level and the true label. It is the first The prediction confidence of each candidate box. No. The true labels of each candidate box It is the set of positive candidate boxes. It represents the number of positive candidate boxes. It is the first bin classification loss for each positive candidate box It is the first The bin-in-bin residual regression loss for each positive candidate box.

5. The method for detecting deformation of coal mine roadways based on self-supervised learning according to claim 4, characterized in that, In step one, simulated deformation tunnel point cloud data is obtained through subsidence deformation, compression deformation, expansion deformation, and tilting deformation operations. ; Among them, during the subsidence deformation process, by along The translation of the axis is used to simulate subsidence deformation and generate point cloud data of the deformed tunnel. , , The transformation matrix is ​​for the subsidence deformation. , This is the subsidence amount, which is a random value set between 0 and 50; During the extrusion deformation process, through shaft and Scaling along the axis to simulate extrusion deformation and generate point cloud data of deformed tunnels. , , The transformation matrix for extrusion deformation. , This is the compression factor, which is a random value set between 0 and 1; During the expansion deformation process, the entire point cloud is scaled up proportionally to simulate the expansion deformation and generate deformed tunnel point cloud data. , , To extend the transformation matrix, , This is the expansion factor, which is a random value set between 0 and 1; During the tilting deformation process, rotation is used to simulate the tilting deformation and generate point cloud data of the deformed tunnel. , , The transformation matrix is ​​for tilt deformation. , The rotation angle is a random value set between 0 and 30 degrees.

6. The method for detecting deformation of coal mine roadways based on self-supervised learning according to claim 5, characterized in that, In step two, the deformed tunnel point cloud data is processed through rotation transformation, translation transformation, scaling transformation, and noise addition. Enhancement processing is performed, among which, , This indicates the first element in the generated tunnel point cloud data. Three-dimensional point coordinates, The total number of points in the generated tunnel point cloud; During the rotation transformation, the point cloud is surrounded Rotate the axis, each point in the point cloud After rotation, the new point coordinates are obtained as follows: The rotation transformation formula is: ,in, For rotation matrix, , The rotation angle; During the translation transformation, by , , Translation is achieved by adding a translation amount to the axis. The translation transformation formula is: ,in, It is a translation vector. , , , They represent along , , Translation of the axis; During the scaling transformation, each coordinate of the point cloud is multiplied by a scaling factor. , , Scaling is achieved using the following formula: ; During the noise addition process, a Gaussian-distributed random noise is added to each coordinate of the point cloud. The formula for adding noise is: , , ,in, It is the mean. That is the standard deviation.

7. A coal mine roadway deformation detection device based on self-supervised learning, used to implement the coal mine roadway deformation detection method based on self-supervised learning as described in any one of claims 1 to 6, characterized in that, This includes a tunnel deformation point cloud data generation module, a data enhancement module, and a tunnel deformation intelligent detection module; The tunnel deformation point cloud data generation module is used to generate deformed three-dimensional point cloud data based on the labeled original point cloud data of coal mine tunnels, according to the tunnel geometry and different deformation types, and then send the three-dimensional point cloud data to the data enhancement module. The data augmentation module is used to perform deformation processing on 3D point cloud data to different degrees. At the same time, noise and disturbance are introduced to realistically simulate the actual data, so as to expand the diversity of the dataset and obtain an extended dataset. The extended dataset is then sent to the self-supervised learning module. The intelligent tunnel deformation detection module includes a self-supervised learning module and a deformation detection module: The self-supervised learning module is used to automatically extract features from unlabeled point cloud data in the extended dataset through variational autoencoder contrastive learning and cross-entropy loss supervision, and send the obtained feature data to the deformation detection module. The deformation detection module is used to accurately identify the deformation area of ​​the roadway based on feature data and to accurately assess the severity of the deformation.

8. The coal mine roadway deformation detection device based on self-supervised learning according to claim 7, characterized in that, The different deformation types include subsidence deformation, compression deformation, expansion deformation, and tilting deformation.

9. A coal mine roadway deformation detection device based on self-supervised learning according to claim 8, characterized in that, Data transformation includes rotation, translation, scaling, and noise addition.

Images