Pipeline deformation visual measurement method and system based on a probe robot

Abstract

The application relates to the technical field of visual measurement, and discloses a pipeline deformation visual measurement method and system based on a detection robot, wherein the method comprises the following steps: collecting a pipeline inner wall image through an image sensor carried by the detection robot and pre-processing the pipeline inner wall image to obtain a pre-processed pipeline image; performing defect perception mask processing and periodic convolution coding to obtain a semantic latent vector; calculating a target signal-to-noise ratio according to a UWB transceiver in the detection robot to perform transmission; recovering a time domain pulse signal received by a ground receiving end to obtain the semantic latent vector and reconstruct a reconstructed pipeline image; extracting a pipeline contour point set of the reconstructed pipeline image, performing elliptical fitting on the pipeline contour point set, and constructing a continuous pipeline deformation curved surface; the method eliminates discontinuous artifacts at a pipeline image splicing position, guarantees the integrity of image semantics, realizes high-fidelity pipeline image reconstruction, and further realizes accurate measurement of pipeline deformation.

Description

Technical Field

[0001] This invention relates to the field of visual measurement technology, and in particular to a visual measurement method and system for pipe deformation based on a detection robot. Background Technology

[0002] As a crucial component of infrastructure, urban underground pipeline systems are prone to defects such as cracks, corrosion, and deposits over long-term operation, leading to pipeline deformation and even collapse. Traditional manual inspection methods are inefficient and pose safety hazards, making detection robots an important tool for pipeline inspection. However, multipath interference caused by metal walls and accumulated water in the sewer environment severely degrades wireless communication channels, often resulting in negative signal-to-noise ratios. Existing pipeline image transmission methods employ pixel-by-pixel compression encoding, which requires high bandwidth and suffers from high error rates in low signal-to-noise ratio environments, making it difficult to guarantee image quality. Furthermore, the 360-degree pipeline images acquired by image sensors exhibit discontinuities and artifacts at stitching points, affecting the accuracy of subsequent defect identification and deformation measurement. Summary of the Invention

[0003] This invention provides a visual measurement method and system for pipeline deformation based on a detection robot. This invention eliminates discontinuity artifacts at the stitching points of pipeline images, ensures the integrity of image semantics, achieves high-fidelity pipeline image reconstruction, and thus realizes accurate measurement of pipeline deformation.

[0004] In a first aspect, the present invention provides a visual measurement method for pipe deformation based on a detection robot, the method comprising:

[0005] The robot acquires images of the inner wall of the pipe using its image sensor, and then preprocesses these images to obtain a preprocessed pipe image.

[0006] The preprocessed pipeline image is subjected to defect-aware masking and periodic convolutional coding to obtain a semantic latent vector;

[0007] The target signal-to-noise ratio is calculated based on the UWB transceiver in the detection robot, and the semantic latent vector is mapped into a time-domain pulse signal for transmission based on the target signal-to-noise ratio.

[0008] The temporal pulse signal received by the ground receiver is recovered to obtain the semantic latent vector, and the image is reconstructed based on the semantic latent vector to obtain the reconstructed pipeline image.

[0009] Extract the pipe contour point set from the reconstructed pipe image, and perform ellipse fitting on the pipe contour point set to construct a continuous pipe deformation surface.

[0010] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, the step of acquiring images of the inner wall of the pipe using an image sensor mounted on a detection robot, and preprocessing the inner wall images of the pipe to obtain a preprocessed pipe image, includes:

[0011] The robot acquires images of the inner wall of the pipe using an image sensor, and projects these images onto a fisheye lens cylindrical surface to create a first inner wall image with circumferential angular coordinates and axial height coordinates.

[0012] The first inner wall image is subjected to adaptive histogram equalization and bilinear interpolation to obtain the second inner wall image.

[0013] The image region at the boundary of the second inner wall image is copied and filled cyclically to obtain a preprocessed pipe image.

[0014] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the step of performing defect-aware masking and periodic convolutional coding on the preprocessed pipeline image to obtain a semantic latent vector includes:

[0015] The preprocessed pipeline image is input into a lightweight feature extraction network, and semantic feature maps are extracted by concatenating depthwise convolution and pointwise convolution through depthwise separable convolution in the lightweight feature extraction network.

[0016] The semantic feature map is input into the defect-aware masking module, a defect masking matrix is ​​generated by the defect-aware masking module, and random masking is performed based on the defect masking matrix to obtain the masked sparse features.

[0017] The sparse features after masking are input into a circular convolutional encoder, and a circular convolution is performed in the circumferential direction to obtain a semantic latent vector.

[0018] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the step of inputting the semantic feature map into a defect-aware masking module, generating a defect masking matrix through the defect-aware masking module, and performing random masking based on the defect masking matrix to obtain masked sparse features includes:

[0019] The semantic feature map is input into the defect-aware masking module, and spatial convolution is performed on cracks, corrosion and sediment respectively to obtain the defect response map;

[0020] The defect response map is subjected to sigmoid activation function operation to obtain the defect mask matrix;

[0021] The mask value at each spatial location in the defect mask matrix is ​​compared with a set mask threshold. Locations with mask values ​​greater than the mask threshold are marked as defect pixels, preserving the original features of the semantic feature map. Locations with mask values ​​less than or equal to the mask threshold are generated with random mask labels.

[0022] The original features at the positions of the random mask markers are replaced with learnable embedding vectors, and then merged with the defective pixel positions that retain the original features to obtain the sparse features after masking.

[0023] In conjunction with the first aspect, in a fourth implementation of the first aspect of the present invention, the step of inputting the masked sparse features into a circular convolutional encoder and performing circular convolution in the circumferential direction to obtain a semantic latent vector includes:

[0024] The sparse features after the mask are input into the circular convolution encoder, and periodic boundary conditions are set at the boundaries of the circumferential angles of zero degrees and 360 degrees, so that the convolution kernel automatically takes values ​​cyclically at the boundary positions.

[0025] The first layer of the circular convolutional encoder with a stride of two is used to perform downsampling convolution processing on the sparse features after masking to obtain the first layer of encoded feature map.

[0026] The first layer of encoded feature map is sequentially input into the second layer of circular convolutional layer with a stride of 2, the third layer of circular convolutional layer with a stride of 2, and the fourth layer of circular convolutional layer with a stride of 1. Downsampling convolution is performed to progressively compress the spatial resolution and output the semantic latent vector.

[0027] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present invention, the step of calculating the target signal-to-noise ratio based on the UWB transceiver in the detection robot, and mapping the semantic latent vector into a time-domain pulse signal for transmission based on the target signal-to-noise ratio, includes:

[0028] The target signal-to-noise ratio is calculated based on the received signal strength indication value measured by the UWB transceiver in the detection robot;

[0029] The semantic latent vector is input into the joint source-channel coding network. The dimension of the feature channel is adjusted according to the target signal-to-noise ratio. The coding redundancy rate is adaptively adjusted within a preset coding range. Error protection weights are applied to defective semantic channels to obtain the channel symbol sequence.

[0030] The channel symbol sequence is mapped into a time-domain pulse signal through UWB pulse position modulation, and the time-domain pulse signal is transmitted to the ground receiving end.

[0031] In conjunction with the first aspect, in the sixth implementation of the first aspect of the present invention, the step of recovering the time-domain pulse signal received by the ground receiver to obtain a semantic latent vector, and performing image reconstruction based on the semantic latent vector to obtain a reconstructed pipeline image, includes:

[0032] Viterbi decoding is performed on the time-domain pulse signal received by the ground receiver to recover the time-domain pulse signal as a semantic latent vector;

[0033] The spatial dimension of the semantic latent vector is flattened into a token sequence and then input into the Transformer decoder. The decoded semantic features are obtained by alternating processing of the multi-head self-attention mechanism and the feedforward network in the Transformer decoder.

[0034] The decoded semantic features are concatenated with the defect condition vectors of cracks, corrosion, and sediments, and then input into a gradient-penalized adversarial generative network. The spatial resolution is gradually restored through five layers of transposed convolutional upsampling in the gradient-penalized adversarial generative network to obtain the reconstructed pipeline image.

[0035] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, the step of performing Viterbi decoding on the time-domain pulse signal received by the ground receiver to recover the time-domain pulse signal as a semantic latent vector includes:

[0036] A pseudo-random pilot sequence is extracted from the time-domain pulse signal received from the ground receiver, and a multipath channel matrix is ​​calculated based on the pseudo-random pilot sequence.

[0037] A mesh graph structure is constructed based on the multipath channel matrix, and the squared Euclidean distance between the received signal and the desired signal is calculated as the branch metric for each path in the mesh graph structure.

[0038] The cumulative metric value at the current time is obtained by summing the path metric value at the previous time step, and the path with the minimum cumulative metric value for each state is retained as the surviving path based on the cumulative metric value.

[0039] The surviving path with the smallest cumulative metric value in the termination state is selected as the optimal decoding path. The optimal decoding path is then traced back to recover the time-domain pulse signal and decoded into a semantic latent vector.

[0040] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, the step of extracting the pipeline contour point set of the reconstructed pipeline image and performing ellipse fitting on the pipeline contour point set to construct a continuous pipeline deformation surface includes:

[0041] Perform an elliptic kernel morphological closing operation on the reconstructed pipeline image to form a pipeline contour point set;

[0042] Random sampling ellipse fitting is performed on the pipeline profile point set to obtain ellipse parameters, and the radial deviation between the measured radius of the profile point and the radius of the fitted ellipse is calculated based on the ellipse parameters.

[0043] An octree is constructed based on the radial deviation, and the deviation value within each leaf node is calculated based on the octree to generate a continuous pipeline deformation surface.

[0044] Secondly, the present invention provides a pipe deformation visual measurement system based on a detection robot, the pipe deformation visual measurement system based on a detection robot comprising:

[0045] The acquisition module is used to acquire images of the inner wall of the pipe through the image sensor mounted on the detection robot, and to preprocess the images of the inner wall of the pipe to obtain a preprocessed pipe image.

[0046] The encoding module is used to perform defect-aware masking and periodic convolutional encoding on the preprocessed pipeline image to obtain a semantic latent vector;

[0047] The transmission module is used to calculate the target signal-to-noise ratio based on the UWB transceiver in the detection robot, and to map the semantic latent vector into a time-domain pulse signal for transmission based on the target signal-to-noise ratio;

[0048] The reconstruction module is used to recover the time-domain pulse signal received by the ground receiver, obtain the semantic latent vector, and perform image reconstruction based on the semantic latent vector to obtain the reconstructed pipeline image.

[0049] The fitting module is used to extract the pipeline contour point set of the reconstructed pipeline image and perform ellipse fitting on the pipeline contour point set to construct a continuous pipeline deformation surface.

[0050] The technical solution provided by this invention achieves efficient image transmission and accurate deformation detection in low signal-to-noise ratio environments by constructing a pipeline deformation measurement method based on semantic communication and visual reconstruction using a detection robot. A defect-aware masking module generates adaptive masks for three types of defect areas: cracks, corrosion, and sediment. This preserves key semantic features while applying a high proportion of random masks to non-defect areas, significantly compressing data volume and reducing transmission bandwidth requirements. A circular convolutional encoder performs cyclic convolution with periodic boundary conditions in the circumferential angle direction, eliminating discontinuous artifacts at panoramic image stitching points and ensuring the integrity of image semantics. A joint source-channel coding network dynamically adjusts the feature channel dimension and coding redundancy rate based on real-time measured signal-to-noise ratio, applying error protection weights to the defect semantic channel, enabling the system to reliably transmit key defect information even under harsh channel conditions. At the ground receiver, after recovering the semantic vector through Viterbi decoding, a Transformer decoder and a gradient-penalized adversarial generative network are used to perform defect texture-guided adversarial generation, achieving high-fidelity pipeline image reconstruction. Based on the reconstructed image, a continuous deformable surface is constructed through random sampling ellipse fitting and octree spatial decomposition, enabling accurate measurement of pipeline deformation.

[0051] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of an embodiment of the visual measurement method for pipeline deformation based on a detection robot according to the present invention;

[0054] Figure 2 This is a schematic diagram of one embodiment of the visual measurement system for pipeline deformation based on a detection robot according to the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0057] To facilitate understanding of this embodiment, a detailed description of a pipe deformation visual measurement method based on a detection robot, as disclosed in this embodiment of the invention, will be provided first. For example... Figure 1 As shown, this method includes the following steps:

[0058] 101. The image sensor on the robot is used to collect images of the inner wall of the pipe, and the images of the inner wall of the pipe are preprocessed to obtain a preprocessed pipe image;

[0059] Specifically, the robot's front-end image sensor continuously acquires panoramic images of the pipe's inner wall. The image sensor is a 360-degree panoramic camera with a fisheye lens structure, fixedly mounted so that its optical axis is perpendicular to the pipe's axis, ensuring complete coverage of the circumferential information during image projection. During image acquisition, the camera continuously acquires color RGB images at a set frame rate (e.g., 30 frames per second), and maps each acquired raw image frame to a first inner wall image using a cylindrical unfolding projection algorithm. This first inner wall image is defined in an unfolded plane coordinate system with circumferential angle θ and axial height h as coordinates, where θ is the angle range from 0° to 360°, and h corresponds to the axial length coordinate of the robot's current moving position. Due to the non-uniformity of the ambient lighting inside the pipe, inconsistent local exposure occurs in the image. Contrast-limited adaptive histogram equalization (HQE) is applied to the first inner wall image to enhance the local contrast of the luminance channel. The HQE algorithm divides the image into fixed-size (e.g., 8×8 pixels) grid regions and performs histogram equalization independently within each region. A contrast-limiting threshold is set to prevent excessive local enhancement. Simultaneously, bilinear interpolation is used to fuse adjacent grid boundaries, eliminating artifacts and boundary breaks caused by block processing, thus obtaining the second inner wall image. Fixed-width image blocks are copied to the edge regions at θ=0° and θ=360° of the second inner wall image and then circularly stitched together to form a cyclically filled image tensor with periodic boundary conditions. This cyclically filled image is the preprocessed pipe image.

[0060] 102. Perform defect-aware masking and periodic convolutional coding on the preprocessed pipeline image to obtain the semantic latent vector;

[0061] Specifically, the preprocessed pipeline image is input into a lightweight semantic feature extraction network. This network employs a lightweight structure based on depthwise separable convolutions, such as the MobileNetV3 architecture. In each feature extraction unit, a 3×3 depthwise convolution is first performed to extract local spatial features within a channel, followed by a 1×1 pointwise convolution to complete a linear combination between channels. This reduces computational complexity and parameter size while preserving multi-layered semantic expressive power. A channel attention mechanism is introduced in key layers to enhance the response to defect regions. After multi-layer stacking and non-linear activation, a semantic feature map is output. This semantic feature map is then input into a defect-aware masking module. This module constructs dedicated convolutional branches for three typical pipeline defects: cracks, corrosion, and sediment. Each branch activates the semantic features based on pre-trained parameters, and then calculates a continuous defect mask matrix using the Sigmoid function. Each value in the defect mask matrix represents the probability strength of the current location belonging to a certain type of defect. The mask matrix is ​​binarized based on a set defect threshold. Simultaneously, Bernoulli random sampling masking is performed with a set probability in non-defect regions. The masked feature locations are then replaced and filled with trainable embedding vectors, generating a sparse feature representation that preserves the semantic information of the defect region while compressing the invalid background. This masked sparse feature is input into a convolutional coding network with circular periodic boundary conditions. A specially designed circular convolution operator is used within the network to achieve intrinsic continuity between the circular angular directions θ=0° and θ=360° during convolution operations, ensuring no splicing artifacts in the edge regions and maintaining topological consistency. The latent vector tensor is output through multi-layer downsampling and channel compression.

[0062] 103. Calculate the target signal-to-noise ratio based on the UWB transceiver in the probe robot, and map the semantic latent vector into a time-domain pulse signal for transmission based on the target signal-to-noise ratio;

[0063] Specifically, based on the ultra-wideband (UWB) transceiver module installed on the probe robot, the received signal strength indication (RSSI) value is acquired in real time to characterize the physical quality of the current communication link. The RSSI value is calculated by converting the received signal power to the background noise power and then calculating the target signal-to-noise ratio in logarithmic form, reflecting the interference level of the UWB channel and the available data transmission bandwidth at the robot's location. The semantic latent vector is input into the deep joint source-channel coding network (JSCC network). In the network, the corresponding feature channel dimension adjustment factor is calculated according to the numerical range of the current target SNR. The dimensionality expansion or compression of the latent vector in the channel direction is adjusted through the channel attention mechanism. When the SNR is low, the number of channels is reduced to reduce the overall transmission code rate, and when the SNR is high, the channel dimension is expanded to enhance the reconstruction accuracy. At the same time, the coding redundancy rate is adaptively adjusted according to the target SNR within a preset range (e.g., 1 / 3 to 2 / 3) of the mapping curve, so that the overall coding scheme achieves a dynamic balance between redundancy and efficiency. An unequal error protection strategy is applied to the portion of the semantic latent vector corresponding to the strongly activated channels of the defect mask. This involves introducing specific error protection weighting factors to give higher error correction redundancy to channels with high defect response, ensuring that structural information such as cracks and corrosion is preferentially recovered under channel disturbances. After the above encoding process, the output is a complex-domain channel symbol sequence. This sequence is mapped to a time-domain pulse signal using a UWB modulation module with pulse position modulation. Each symbol is mapped to the position of a fixed-width pulse in multiple time slots, providing high bandwidth and anti-interference capabilities. The signal is then wirelessly broadcast to a ground receiver via an omnidirectional antenna mounted on a robot, using a center frequency of 4.5 GHz and a bandwidth of 500 MHz.

[0064] 104. The time-domain pulse signal received by the ground receiver is recovered to obtain the semantic latent vector, and the image is reconstructed based on the semantic latent vector to obtain the reconstructed pipeline image;

[0065] Specifically, at the receiving end, the time-domain pulse signal sent by the probe robot is received through an ultra-wideband receiving module. The receiving end extracts the pre-pilot pulse sequence for multipath channel estimation and constructs a channel matrix model to describe the propagation distortion caused by factors such as the metal wall inside the pipe and water vapor reflection. The estimated channel response is then input into a Viterbi decoder, which performs maximum likelihood path search based on a Trellis grid structure. A state-transition error-minimum matching operation is then performed on the received signal sequence to recover the original semantic latent vector. The original semantic latent vector contains multi-channel compressed feature representations and has a three-dimensional structure after recovery. The spatial dimension of the original semantic latent vector is flattened into a set of token sequences of fixed length, and these token sequences are input into a semantic decoder containing a multi-layer Transformer structure. Inside the Transformer decoder, the token sequences are linearly mapped to a unified-dimensional embedding representation, and then alternately processed through a multi-head self-attention mechanism and a feedforward network. The multi-head attention mechanism models long-distance dependencies in different semantic subspaces, while the feedforward network performs nonlinear feature transformations. After multi-layer iterative inference, the decoded semantic feature matrix is ​​output. The semantic feature matrix is ​​concatenated with conditional vectors representing defect categories—including one-hot encoded vectors for cracks, corrosion, and sediment—to form a composite input with semantic guidance. This composite input is then fed into the gradient-penalized adversarial generative network WGAN-GP to perform image generation. In the generative network, the generator employs a five-layer transposed convolutional module for spatial upsampling, followed by instance normalization and the LeakyReLU activation function after each layer. This progressively restores the high-resolution structure of the original image from the latent space, generating a reconstructed pipeline image with dimensions of 2048×512 pixels.

[0066] 105. Extract the pipeline contour point set from the reconstructed pipeline image, and perform ellipse fitting on the pipeline contour point set to construct a continuous pipeline deformation surface.

[0067] Specifically, morphological closing operations are performed on the reconstructed pipe image to suppress high-frequency artifacts and edge breaks introduced by semantic decoding. An elliptical structuring element is used as the morphological convolution kernel, and an appropriate radius parameter is set. Dilation operations fill small holes within the contour, and erosion operations restore the structural boundaries, resulting in a pipe inner wall image with continuous contour boundaries and a complete contour structure. The Canny edge detection algorithm is applied to the processed image to extract edge point sets, obtaining a pipe contour point set containing a large number of pixel coordinates. Based on this, the RANSAC random sampling consensus algorithm is applied for robust ellipse fitting. In each iteration, a minimum number of points are randomly selected from the contour point set to construct a candidate sample set, and the corresponding ellipse parameter vector is solved using the least squares method. During iteration, the optimal model is selected based on the number of interior points to maximize the ability to describe the true contour. After fitting, the geometric parameters of the fitted ellipse, including the center coordinates, major and minor axis lengths, and major axis angle, are used as a reference curve. The difference between the measured radius in polar coordinates and the radius of the fitted ellipse is calculated for all contour points sequentially to obtain the radial deviation function. The radial deviation function is embedded into an octree structure built in 3D space. The octree structure takes the entire pipeline space as the root node and recursively divides each spatial sub-block into eight child nodes until the spatial resolution reaches the set minimum deformation detection unit, such as a 0.5mm cube. During this process, the radial deviation values ​​contained in each leaf node are aggregated, such as by calculating the average value or weighted interpolation, and a local deformation function is constructed using a bicubic interpolation function. These are then stitched together on all the leaf nodes of the octree to form a continuous deformation surface map defined in 3D space.

[0068] In one specific embodiment, the process of performing step 101 may specifically include the following steps:

[0069] The robot collects images of the inner wall of the pipe using its image sensor, and then projects these images onto the cylindrical surface of a fisheye lens to create a first inner wall image with circumferential angular coordinates and axial height coordinates.

[0070] Adaptive histogram equalization and bilinear interpolation are performed on the first inner wall image to obtain the second inner wall image;

[0071] The image region at the boundary of the second inner wall image is copied and filled cyclically to obtain the preprocessed pipe image.

[0072] Specifically, the panoramic vision sensor mounted on the inspection robot is configured as a wide-angle lens with fisheye imaging characteristics. The lens can achieve high coverage imaging of the 360-degree circumferential range inside the pipe, and the sensor acquires a sequence of original images at set time intervals. Due to the obvious radial distortion and nonlinear projection relationship of the image formed by the fisheye lens, the original image is subjected to cylindrical unfolding transformation to restore the true geometric structure of the inner wall surface of the pipe. Based on the intrinsic parameter model of the fisheye camera, the spherical projection coordinates are mapped to the cylindrical unfolding coordinate system. The polar coordinate system of the fisheye image is aligned with the center of the pipe cross-section, and the angular coordinates θ are distributed horizontally from 0° to 360°, while the axial displacement h is arranged vertically along the robot's movement direction, thereby constructing a first inner wall image defined by the circumferential angle and axial height. The unprocessed first inner wall image suffers from uneven local contrast, unstable lighting, and blurred exposure areas. To address this, an adaptive histogram equalization mechanism with contrast limitations is employed. This mechanism divides the entire first inner wall image into several local regions (e.g., 8×8 pixel blocks). A separate grayscale histogram is calculated for each region, and its local maximum contrast threshold is limited. The histograms are then evenly distributed, and bilinear interpolation smooths the abrupt grayscale changes at the boundaries between regions. This avoids the over-enhancement and image noise amplification problems introduced by traditional histogram equalization, generating a second inner wall image with superior visual contrast. A cyclic filling operation is then performed on the boundary regions of the second inner wall image. Specifically, a fixed-width (e.g., 16 pixels) image region is copied at both ends of the θ direction. Pixels near θ=0° are copied and pasted to positions after θ=360°, and similarly, pixels near θ=360° are copied and pasted to positions before θ=0°. This creates a periodic boundary structure that maintains the continuity of image features during convolution calculations, resulting in a preprocessed pipeline image as the output.

[0073] In one specific embodiment, the process of performing step 102 may specifically include the following steps:

[0074] The preprocessed pipeline image is input into a lightweight feature extraction network, and semantic feature maps are extracted by concatenating depthwise separable convolution and pointwise convolution in the lightweight feature extraction network.

[0075] The semantic feature map is input into the defect-aware masking module, which generates a defect masking matrix and performs random masking based on the defect masking matrix to obtain the sparse features after masking.

[0076] The masked sparse features are input into a circular convolutional encoder, and circular convolution is performed in the circumferential direction to obtain the semantic latent vector.

[0077] Specifically, the preprocessed pipeline image is input into a lightweight semantic feature extraction network, built on the MobileNetV3 architecture, to maintain efficient feature extraction capabilities even under resource-constrained embedded devices. Within the feature extraction network, the core employs depthwise separable convolution operations to reduce the number of parameters and computational complexity. The core structure consists of two parts: a 3×3 depthwise convolutional layer for extracting local spatial structure information within each channel, and a 1×1 pointwise convolutional layer for linear combination and semantic aggregation along the channel dimension. This cascaded combination effectively captures defect textures and structural changes at different locations in the pipeline image, generating a semantic feature map with high-dimensional semantic expression capabilities. The semantic feature map is then input into a defect-aware masking module, which contains multi-branch convolutional structures to perceive three typical defect types: cracks, corrosion, and sediment. Based on pre-trained parameters, it outputs a three-channel defect response map. The defect response map is then activated with a sigmoid function to obtain a defect mask matrix with values ​​ranging from [0,1]. The value of each pixel in the matrix represents the probability that the current location belongs to a certain type of defect. To compress feature information in non-defective regions while preserving high semantic resolution in defective regions, a threshold is applied to the defect mask matrix for binarization. A Bernoulli-distributed random masking mechanism is introduced in low-response regions, replacing the original features with trainable embedding vectors at the mask locations, forming a sparse but semantically dense mask feature tensor. The masked sparse features are then input into a circular convolutional encoder with periodic boundary processing capabilities. This encoder is specifically designed for processing image structures with periodic topology along the θ direction. The convolutional kernel automatically cyclically samples at the boundaries along the circular angle direction, enabling continuous computation between θ=0° and θ=360°, effectively eliminating artifacts and information fragmentation problems in panoramic images at edge regions. The circular convolutional encoder progressively compresses the image spatial dimension through multiple downsampling convolution operations while enhancing semantic abstraction capabilities, outputting a latent vector tensor.

[0078] In one specific embodiment, the process of inputting the semantic feature map into the defect-aware masking module, generating a defect masking matrix through the defect-aware masking module, and performing random masking based on the defect masking matrix to obtain the masked sparse features can specifically include the following steps:

[0079] The semantic feature map is input into the defect-aware masking module, and spatial convolution is performed on cracks, corrosion and sediment respectively to obtain the defect response map.

[0080] The defect response map is activated by the sigmoid function to obtain the defect mask matrix;

[0081] The mask value at each spatial location in the defect mask matrix is ​​compared with a set mask threshold. Locations with mask values ​​greater than the mask threshold are marked as defect pixels to preserve the original features of the semantic feature map. Locations with mask values ​​less than or equal to the mask threshold are generated with random mask labels.

[0082] The original features at the positions of the random mask markers are replaced with learnable embedding vectors, and then merged with the defect pixel positions that retain the original features to obtain the sparse features after masking.

[0083] Specifically, the semantic feature map is input into the defect-aware masking module, which contains three independent convolutional branches corresponding to three typical pipe defects: cracks, corrosion, and sediment. Each branch uses a pre-trained 3×3 convolutional kernel to perform spatial convolution operations on the semantic feature map, extracting local activation patterns related to their respective target defects, and outputting a three-channel defect response map with a consistent shape. Each channel represents the activation intensity of a specific defect type at each location in the image space. To transform the response map into a probabilistic mask representation, a sigmoid activation function is applied to normalize its numerical range to [0,1], resulting in a defect mask matrix with the same size as the original image. In the defect mask matrix, the value of each pixel position represents the probability that its corresponding region belongs to a certain type of defect. The defect mask matrix is ​​compared point-by-point with a preset mask threshold. When the mask value at a certain position is greater than the preset threshold (e.g., τ=0.6), the current position is determined to be a significant defect region, and the original feature vector of the current position in the semantic feature map is directly retained. For positions with mask values ​​less than or equal to the threshold, they are considered to belong to non-critical regions or background regions, and random mask markers are generated based on a Bernoulli distribution to determine whether to cover the original features of the current position. For positions selected by random mask markers, their original features are uniformly replaced with a learnable embedding vector. The embedding vector is used as a parameter during training to co-optimize with other network parts, which has the function of compressing redundant information and maintaining the consistency of feature distribution. By fusing the defect regions with retained original features and the non-defect regions replaced with embedding vectors in the same feature map according to their positions, a masked sparse semantic feature map is constructed.

[0084] In one specific embodiment, the process of inputting the masked sparse features into a circular convolutional encoder and performing circular convolution in the circumferential direction to obtain the semantic latent vector can specifically include the following steps:

[0085] The sparse features after masking are input into the circular convolutional encoder, and periodic boundary conditions are set at the boundaries of the circumferential angles of zero degrees and 360 degrees, so that the convolutional kernel automatically takes values ​​cyclically at the boundary positions.

[0086] The first layer of the circular convolutional encoder with a stride of two is used to perform downsampling convolution on the sparse features after masking to obtain the first layer of encoded feature map.

[0087] The first layer of encoded feature map is sequentially input into the second layer of circular convolutional layer with a stride of 2, the third layer of circular convolutional layer with a stride of 2, and the fourth layer of circular convolutional layer with a stride of 1. Downsampling convolution is performed to progressively compress the spatial resolution and output the semantic latent vector.

[0088] Specifically, the sparse semantic feature map is input into a circular convolutional encoder with periodic boundary characteristics. A continuous mapping relationship between the circumferential angles θ=0° and θ=360° is explicitly set in the input tensor of the sparse semantic feature map, thus forming a logically topologically consistent circular structure at the tensor boundary. This circular structure, by configuring the padding mode of the convolution operation to "circular" at the implementation level, allows the receptive field located at the left edge of the image to directly cross to the right edge to obtain compensation values, and vice versa, thus avoiding boundary artifacts and structural fragmentation caused by zero padding in conventional convolutions, and ensuring continuous feature distribution in the θ direction when processing panoramic images. A first-layer circular convolutional layer with a stride of 2 performs a downsampling operation on the sparse feature map, reducing the spatial dimensions of the input feature map to half its original size in both the θ and axial directions while maintaining the kernel size and channel dimension. The output is the first-layer encoded feature map. The first layer's encoded feature map is used as an intermediate representation and input into a second ring convolutional layer with a stride of 2. This is further downsampled to compress the spatial dimension to one-quarter of its initial value, continuously extracting multi-scale semantic representations during the process. This multi-scale semantic representation is then input into a third ring convolutional layer with a stride of 2 for further spatial compression, extracting higher-level abstract features and forming a lower-resolution but higher-semantic-density representation. In the fourth layer, a ring convolutional layer with a stride of 1 performs a convolutional transformation on the output of the third layer. Without further downsampling, this integrates deep feature representations from the channel dimension and outputs a fixed-size, compact, high-semantic-density tensor as a semantic latent vector.

[0089] In one specific embodiment, the process of performing step 103 may specifically include the following steps:

[0090] The target signal-to-noise ratio is calculated based on the received signal strength indication value measured by the UWB transceiver in the detection robot;

[0091] The semantic latent vector is input into the joint source-channel coding network. The dimension of the feature channel is adjusted according to the target signal-to-noise ratio. The coding redundancy rate is adaptively adjusted within the preset coding range. Error protection weights are applied to the defective semantic channels to obtain the channel symbol sequence.

[0092] The channel symbol sequence is mapped into a time-domain pulse signal through UWB pulse position modulation and then transmitted to the ground receiver.

[0093] Specifically, an ultra-wideband transceiver module installed on the exploration robot is used to evaluate the current communication channel status in real time, periodically collecting the received signal strength indicator (RSSI) value as a physical quantity reflecting link quality. The RSI value is obtained by detecting the signal power received by the receiving antenna through the underlying RF front-end and estimating the target signal-to-noise ratio (SNR) of the current channel based on the static background noise level in the environment. This SNR value represents the communication environment quality between the robot and the ground receiving base station. The semantic latent vector is input into a joint source-channel coding network. An adaptive control mechanism based on SNR adjustment is embedded in the network structure. The mechanism determines the channel state interval based on the currently estimated SNR level, thereby automatically adjusting the feature channel dimension of the latent vector during the coding stage. When the SNR is low, the number of channels is compressed to reduce transmission bit load and enhance redundancy error correction capability; conversely, when the SNR is high, the number of channels is increased to enhance reconstruction quality. During the encoding process, the defect response intensity output by the masking module is used as a guiding principle to apply error protection weights to the defect channels in the semantic latent vector. This ensures that channels related to defect structures such as cracks, corrosion, and sediments receive higher levels of protection resources in the coding redundancy allocation, improving the transmission stability of key semantic content under interference environments. The encoding result, after adjusting the feature channel dimensions and redundancy rate, is mapped to a set of channel symbol sequences. These sequences are represented as a set of modulation symbols in the complex domain and input to the ultra-wideband modulation module for modulation conversion. During modulation, pulse position modulation is used to map each channel symbol to one of multiple time slots within a fixed duration. Information is carried in the time domain using ultra-short pulses with high time resolution. Each symbol corresponds to a pulse located at a specific position. The resulting time-domain pulse signal has advantages such as high center frequency, wide bandwidth, strong penetration, and strong resistance to multipath interference. It is broadcast to the ground receiver via an omnidirectional antenna mounted on the robot.

[0094] In one specific embodiment, the process of performing step 104 may specifically include the following steps:

[0095] Viterbi decoding is performed on the time-domain pulse signal received by the ground receiver to recover the time-domain pulse signal as a semantic latent vector;

[0096] The spatial dimension of the semantic latent vector is flattened into a token sequence and then input into the Transformer decoder. The decoded semantic features are obtained by alternating processing through the multi-head self-attention mechanism and the feedforward network in the Transformer decoder.

[0097] The decoded semantic features are concatenated with the defect condition vectors of cracks, corrosion, and sediments and then input into a gradient-penalized adversarial generative network. The spatial resolution is gradually restored through five layers of transposed convolution upsampling in the gradient-penalized adversarial generative network to obtain the reconstructed pipeline image.

[0098] Specifically, at the ground receiver, the time-domain pulse signal received via the ultra-wideband communication link undergoes preliminary channel demodulation. After demodulation, the resulting symbol sequence is input into a Viterbi decoder for decoding. The Viterbi decoder employs a path metric search mechanism based on a state transition graph, progressively evaluating all possible coding paths and selecting the path with the minimum cumulative cost as the final decoding result. This achieves effective error correction and recovery of the modulated signal affected by noise interference and multipath distortion, outputting a structured semantic latent vector tensor. The spatial dimension of the semantic latent vector tensor is flattened into a one-dimensional token sequence in channel-priority order. Each token corresponds to a set of channel features in the latent vector, maintaining the sequential structure to preserve spatial semantic association. The token sequence is input into a multi-layered stacked Transformer decoder. Inside the Transformer decoder, the tokens are projected to a unified embedding dimension through a linear mapping. Then, they are processed alternately by a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism models global dependencies in the token sequence, allowing each feature to not only incorporate its own information but also perceive the global semantic context. The feedforward network independently performs nonlinear transformations on each position to enhance expressive power. After multi-layer stacking, the decoded high-dimensional semantic feature sequence is output. The decoded semantic features are concatenated with a defect type condition vector. The defect type condition vector represents three pipe defect types—cracks, corrosion, and sediment—in one-hot encoded form, giving the generator explicit semantic guidance during reconstruction. The concatenated joint features are input into the generator module of the gradient-penalized adversarial generative network. The generator employs a five-layer transposed convolution upsampling process, with each transposed convolution followed by normalization and a non-linear activation function. This process doubles the spatial resolution to progressively restore the size and structural details of the original image. During this process, the generator infers texture, edge, and shape information based on semantic features and combines conditional vectors to generate structural responses that match the target defect type. Simultaneously, the discriminator of the adversarial network undergoes stable training with a gradient-penalized structure. This constrains the model's learning process to maintain Lipschitz continuity, allowing the optimization objectives of structural realism and distribution consistency in the generated image to converge. The output reconstructed image spatially restores the original resolution, content-wise reproduces the realistic texture structure of the pipe's inner wall, and achieves high-fidelity reconstruction in key defect areas.

[0099] In one specific embodiment, the process of performing Viterbi decoding on the time-domain pulse signal received by the ground receiver to recover the time-domain pulse signal as a semantic latent vector can specifically include the following steps:

[0100] The pseudo-random pilot sequence is extracted from the time-domain pulse signal received from the ground receiver, and the multipath channel matrix is ​​calculated based on the pseudo-random pilot sequence.

[0101] A mesh graph structure is constructed based on the multipath channel matrix, and the squared Euclidean distance between the received signal and the desired signal is calculated as the branch metric for each path in the mesh graph structure.

[0102] The cumulative metric value at the current time is obtained by summing the path metric value from the previous time step, and the path with the minimum cumulative metric value for each state is retained as the surviving path based on the cumulative metric value.

[0103] The surviving path with the smallest cumulative metric value in the termination state is selected as the optimal decoding path. The optimal decoding path is then traced back to recover the time-domain impulse signal and decoded into a semantic latent vector.

[0104] Specifically, at the ground receiver, the structure of the time-domain pulse signal frames received from the UWB communication link is identified, and a pre-defined pseudo-random pilot sequence is extracted from the frame header. This pseudo-random pilot sequence, composed of combinations of known pulse positions, possesses good autocorrelation characteristics and anti-interference capabilities. After receiving the pilots, the pilot sequence is matched with the actual received signal. By estimating the changes in impulse response during transmission, a multipath channel matrix is ​​established for the current moment, describing the characteristics of the signal as it travels through multiple reflection paths, including time delay, gain, and phase shift. After the channel matrix calculation is completed, the decoding process transitions to a Viterbi lattice graph structure built based on the channel model. This Viterbi lattice graph is a two-dimensional path graph with the time step as the horizontal axis and the state set as the vertical axis. The evolution of the states is determined by the coding constraint length. Each path extends from the previous state to the current state and corresponds to a set of possible coded bit combinations. For each path, the squared Euclidean distance between the received signal samples and the expected transmitted signal under the corresponding path is evaluated as a branch metric, reflecting the matching quality of the path under the channel model. The current branch metric is summed with the cumulative metric from the previous time step to obtain the cumulative metric of the path in the current state. The path with the smallest cumulative metric is selected from all paths entering the same state and retained as the surviving path for the next time step. This path selection process is progressively advanced throughout the graph structure. After all time steps have been traversed, the decoder selects the path with the smallest cumulative metric from the set of terminated states as the globally optimal decoding path. It then backtracks along this optimal decoding path to the starting state, extracts the corresponding encoded bit sequence, and completes the recovery of the modulated and encoded original signal. The recovered bit sequence is then decoded and mapped into a semantic latent vector.

[0105] In one specific embodiment, the process of performing step 105 may specifically include the following steps:

[0106] Perform elliptic kernel morphological closing operations on the reconstructed pipeline image to form a pipeline contour point set;

[0107] Random sampling ellipse fitting is performed on the pipeline profile point set to obtain ellipse parameters, and the radial deviation between the measured radius of the profile points and the radius of the fitted ellipse is calculated based on the ellipse parameters.

[0108] An octree is constructed based on radial deviation, and the deviation value within each leaf node is calculated based on the octree to generate a continuous pipeline deformation surface.

[0109] Specifically, structural processing is performed on the pipe image to extract geometric boundary information. A morphological closing operation is employed to eliminate small holes and edge breaks caused by decoding noise or adversarial reconstruction flaws. The morphological closing operation uses an elliptical convolution kernel to perform a dilation-erosion operation on the image, enhancing the closure and continuity of the edge contours. After the closing operation, strong edge response regions are extracted from the image using edge detection operators, and a contour point set containing a large number of two-dimensional edge points is constructed in pixel coordinates to represent the actual boundary shape of the pipe's inner wall in the reconstructed image. To evaluate the deviation between the actual boundary and the ideal geometric model, an ellipse fitting operation is performed on the contour point set. To enhance the robustness of the fitting results, a random sampling consensus method is used, randomly selecting a minimum number of samples from the contour point set to solve the general ellipse equation. Error evaluation is performed on each solution, and the optimal solution is selected as the final fitting model through multiple iterations. After obtaining the geometric parameters of the elliptical model, including the center coordinates, major axis, minor axis, and principal axis direction angles, the polar radius of each contour point is compared with the radius of the fitted ellipse at the corresponding angular position. The difference is used to construct a radial deviation distribution, reflecting the degree of deviation of the actual deformation of the pipe's inner wall from the ideal elliptical contour. The radial deviation data is projected onto the pipe's spatial coordinate system, and a hierarchical spatial octree structure is constructed. The entire pipe model space is used as the root node, and the structure is divided into multiple sub-regions according to spatial dimensions. Each node can be recursively divided into eight sub-nodes until a set minimum resolution scale is reached. The radial deviation samples contained within each leaf node are statistically aggregated, for example, by taking the mean or interpolation fitting, thereby generating deformation values ​​on the smallest structural unit of the octree. Based on the spatial distribution of all leaf nodes, a continuously defined three-dimensional pipe deformation surface is reconstructed.

[0110] The above describes the pipe deformation visual measurement method based on a detection robot in the embodiments of the present invention. The following describes the pipe deformation visual measurement system based on a detection robot in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the pipeline deformation visual measurement system based on a detection robot in this invention includes:

[0111] The acquisition module is used to acquire images of the inner wall of the pipe through the image sensor mounted on the detection robot, and to preprocess the images of the inner wall of the pipe to obtain a preprocessed pipe image;

[0112] The encoding module is used to perform defect-aware masking and periodic convolutional encoding on the preprocessed pipeline image to obtain a semantic latent vector.

[0113] The transmission module is used to calculate the target signal-to-noise ratio based on the UWB transceiver in the probe robot, and to map the semantic latent vector into a time-domain pulse signal for transmission based on the target signal-to-noise ratio.

[0114] The reconstruction module is used to recover the time-domain pulse signal received by the ground receiver, obtain the semantic latent vector, and perform image reconstruction based on the semantic latent vector to obtain the reconstructed pipeline image.

[0115] The fitting module is used to extract the pipeline contour point set from the reconstructed pipeline image and perform ellipse fitting on the pipeline contour point set to construct a continuous pipeline deformation surface.

[0116] This invention constructs a pipeline deformation measurement method based on semantic communication and visual reconstruction using a detection robot, achieving efficient image transmission and accurate deformation detection in low signal-to-noise ratio (SNR) environments. A defect-aware masking module generates adaptive masks for three types of defect areas: cracks, corrosion, and sediment. This preserves key semantic features while applying a high proportion of random masks to non-defect areas, significantly compressing data volume and reducing transmission bandwidth requirements. A circular convolutional encoder performs cyclic convolution with periodic boundary conditions in the circumferential angle direction, eliminating discontinuities at panoramic image stitching points and ensuring the integrity of image semantics. A joint source-channel coding network dynamically adjusts the feature channel dimension and coding redundancy rate based on real-time measured SNR, applying error protection weights to the defect semantic channel, enabling reliable transmission of key defect information even under harsh channel conditions. At the ground receiver, after recovering the semantic vector through Viterbi decoding, a Transformer decoder and a gradient-penalized adversarial generative network are used to perform defect texture-guided adversarial generation, achieving high-fidelity pipeline image reconstruction. Based on the reconstructed image, a continuous deformable surface is constructed through random sampling ellipse fitting and octree spatial decomposition, enabling accurate measurement of pipeline deformation.

[0117] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0118] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0119] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A pipe deformation visual measurement method based on a probe robot, characterized by, The method comprises the following steps: An image sensor carried by a detection robot is used to collect a pipe inner wall image, and the pipe inner wall image is preprocessed to obtain a preprocessed pipe image; The preprocessed pipe image is subjected to defect perception mask processing and periodic convolution coding to obtain a semantic latent vector; specifically, the preprocessed pipe image is input into a lightweight feature extraction network, semantic feature maps are extracted through deep convolution and point-by-point convolution in series in the lightweight feature extraction network by means of deep separable convolution; the semantic feature maps are input into a defect perception mask module, spatial convolution is performed on cracks, corrosion and deposits respectively to obtain defect response maps; sigmoid activation function operation is performed on the defect response maps to obtain a defect mask matrix; the mask values of each spatial position in the defect mask matrix are compared with a set mask threshold, the positions with mask values greater than the mask threshold are marked as defect pixels, the original features of the semantic feature maps are retained, for the positions with mask values less than or equal to the mask threshold, random mask marks are generated; the positions of the random mask marks are replaced with learnable embedding vectors to obtain sparse features after masking; the sparse features after masking are input into a ring convolution encoder, periodic boundary conditions are set at the boundary positions of the circumferential angle of zero degrees and 360 degrees, so that the convolution kernel automatically circulates at the boundary positions; the sparse features after masking are subjected to down-sampling convolution processing by a first layer of ring convolution layers with a step size of two in the ring convolution encoder to obtain a first layer of encoded feature maps; the first layer of encoded feature maps are sequentially input into a second layer of ring convolution layers with a step size of two, a third layer of ring convolution layers with a step size of two and a fourth layer of ring convolution layers with a step size of one, down-sampling convolution is performed to gradually compress the spatial resolution, and a semantic latent vector is output; A target signal-to-noise ratio is calculated by a UWB transceiver in the detection robot, and the semantic latent vector is mapped into a time-domain pulse signal based on the target signal-to-noise ratio for transmission; The time-domain pulse signal received by a ground receiving end is recovered to obtain a semantic latent vector, and an image is reconstructed based on the semantic latent vector to obtain a reconstructed pipe image; Pipe contour point sets of the reconstructed pipe image are extracted, and elliptical fitting is performed on the pipe contour point sets to construct a continuous pipe deformation surface.

2. The pipe deformation vision measurement method based on a probe robot according to claim 1, characterized by, The image sensor carried by the detection robot is used to collect a pipe inner wall image, and the pipe inner wall image is projected and mapped into a first inner wall image with circumferential angle coordinates and axial height coordinates through cylindrical expansion projection of a fish-eye lens; Adaptive histogram equalization and bilinear interpolation are performed on the first inner wall image to obtain a second inner wall image; Image regions are copied and cyclically filled at the boundaries of the second inner wall image to obtain a preprocessed pipe image. The target signal-to-noise ratio is calculated by the UWB transceiver in the detection robot, and the semantic latent vector is mapped into a time domain pulse signal based on the target signal-to-noise ratio for transmission.

3. The pipe deformation vision measurement method based on a probe robot according to claim 1, characterized by, ​ According to the received signal strength indication value measured by the UWB transceiver in the detection robot, the target signal-to-noise ratio is calculated; The semantic latent vector is input into a joint source channel coding network, the feature channel dimension is adjusted according to the target signal-to-noise ratio, the coding redundancy rate is adaptively adjusted within a preset coding range, an error protection weight is applied to a defective semantic channel, and a channel symbol sequence is obtained; The channel symbol sequence is mapped into a time domain pulse signal through UWB pulse position modulation, and the time domain pulse signal is transmitted to the ground receiving end.

4. The pipe deformation vision measurement method based on a probe robot according to claim 1, characterized by, The time domain pulse signal received by the ground receiving end is recovered to obtain a semantic latent vector, and an image is reconstructed based on the semantic latent vector to obtain a reconstructed pipeline image, including: The time domain pulse signal received by the ground receiving end is recovered to obtain a semantic latent vector, and an image is reconstructed based on the semantic latent vector to obtain a reconstructed pipeline image, including: The time domain pulse signal received by the ground receiving end is recovered to obtain a semantic latent vector, and an image is reconstructed based on the semantic latent vector to obtain a reconstructed pipeline image, including: The semantic latent vector is flattened into a token sequence and input into a Transformer decoder, and the decoding semantic features are obtained through the multi-head self-attention mechanism and the feedforward network in the Transformer decoder alternately processing.

5. The pipe deformation vision measurement method based on a probe robot according to claim 4, wherein The decoding semantic features and the defect condition vector of cracks, corrosion and deposits are spliced and input into a gradient penalty generative adversarial network, and the spatial resolution is gradually recovered through five-layer transposed convolution upsampling in the gradient penalty generative adversarial network to obtain a reconstructed pipeline image. The time domain pulse signal received by the ground receiving end is recovered to obtain a semantic latent vector, and an image is reconstructed based on the semantic latent vector to obtain a reconstructed pipeline image, including: The pseudo-random pilot sequence is extracted from the time domain pulse signal received by the ground receiving end, and a multipath channel matrix is calculated based on the pseudo-random pilot sequence; A grid graph structure is constructed based on the multipath channel matrix, and the squared Euclidean distance between the received signal and the expected signal is calculated as the branch metric value for each path in the grid graph structure; The path metric value at the previous time is accumulated to obtain the accumulated metric value at the current time, and the minimum accumulated metric path of each state is reserved as the survivor path based on the accumulated metric value; 6. The pipe deformation vision measurement method based on a probe robot according to claim 1, wherein The survivor path with the minimum accumulated metric value of the termination state is selected as the optimal decoding path, and the time domain pulse signal is recovered and decoded to obtain the semantic latent vector by backtracking the optimal decoding path. The pipeline contour point set of the reconstructed pipeline image is extracted, and the pipeline contour point set is elliptical fitting to construct a continuous pipeline deformation surface, including: The elliptical kernel morphological closing operation is performed on the reconstructed pipeline image to form a pipeline contour point set; Random sampling elliptical fitting is performed on the pipeline contour point set to obtain elliptical parameters, and the radial deviation between the measured radius of the contour point and the fitted elliptical radius is calculated based on the elliptical parameters; 7. A pipe deformation vision measurement system based on a probe robot, characterized by, An octree is constructed based on the radial deviation, and the deviation value in each leaf node is calculated based on the octree to generate a continuous pipeline deformation surface. A method for performing the pipeline deformation visual measurement based on a detection robot as claimed in any one of claims 1-6, comprising: An acquisition module is configured to acquire a pipeline inner wall image through an image sensor carried by a detection robot, and to obtain a preprocessed pipeline image by preprocessing the pipeline inner wall image. The coding module is used for defect-aware mask processing and periodic convolution coding on the pre-processed pipeline image to obtain a semantic latent vector; The transmission module is used for calculating a target signal-to-noise ratio of the UWB transceiver in the detection robot, mapping the semantic latent vector into a time-domain pulse signal for transmission based on the target signal-to-noise ratio; The reconstruction module is used for recovering the time-domain pulse signal received by the ground receiving end to obtain a semantic latent vector, and reconstructing an image based on the semantic latent vector to obtain a reconstructed pipeline image; The fitting module is used for extracting a pipeline contour point set of the reconstructed pipeline image, performing elliptical fitting on the pipeline contour point set, and constructing a continuous pipeline deformation surface.

Images