An improved underground pipe circuit fault detection method based on YOLO26

Abstract

The application provides an underground pipe circuit fault detection method based on an improved YOLO26, and relates to the technical field of cable fault detection.The application constructs a thermal imaging simulation database through thermocouple coupling simulation and trains an improved YOLO26 model, which is deployed on a mobile robot inspection platform;the robot independently travels in the underground pipe, synchronously collects visible light and infrared thermal images, and performs dual-spectrum registration, the model performs end-to-end real-time inference to output a defect category, a position and a confidence;combined with pose information, the defect is positioned, high-risk defects trigger an alarm and return information, low-confidence results are reviewed and collected, and the automatic detection, positioning and early warning closed loop of the underground pipe cable defects are realized.

Description

Technical Field

[0001] This invention relates to the field of cable fault detection technology, and in particular to a fault detection method for underground pipe circuits based on an improved YOLO26. Background Technology

[0002] The safe and stable operation of the power system is a crucial guarantee for the national economy and people's livelihood. As an important carrier of urban power distribution networks, the health status of underground cables is directly related to the reliability of power supply. With the continuous expansion of urban power distribution networks, a large number of power cables are laid in underground ducts. However, underground ducts are enclosed spaces with complex environments, often characterized by dampness, dust, and poor lighting conditions. This makes it difficult for traditional manual inspections to achieve high-frequency, full-coverage monitoring, and also results in problems such as low efficiency, high cost, and operational safety risks.

[0003] Existing cable fault detection technologies primarily focus on post-fault diagnosis or offline testing, such as post-fault diagnostic methods like traveling wave methods and impedance methods, as well as offline detection methods like partial discharge. These technical approaches are generally better at locating and determining faults that have already occurred or are relatively significant, but they lack the ability to proactively identify early, minor anomalies that are mainly characterized by visible defects (such as minor damage to the insulation sheath, abnormal color and texture caused by localized overheating, and mechanical extrusion deformation). Furthermore, the traditional approach of first qualitatively identifying the fault, then coarsely locating it, then precisely locating it, and finally excavating for verification relies on instruments and human experience. This leads to long diagnostic cycles and a high probability of rework in long-distance or complex terrain conditions, and high-voltage operations and excavation work impose significant safety and cost burdens.

[0004] In recent years, the integration of robotics, computer vision, and artificial intelligence technologies has led to a significant trend in power facility operation and maintenance, with mobile robots equipped with sensors and capable of autonomous navigation conducting underground space inspections. Vision-based automatic defect detection typically employs deep learning-based target detection models for defect localization and classification, with typical models including two-stage detection frameworks and the YOLO series of single-stage detection frameworks. However, underground pipeline environments are characterized by extreme low light, reflections, oil contamination, cluttered backgrounds, and structural occlusion, resulting in significant image quality degradation. In such complex scenarios, existing visual inspection solutions generally face key challenges such as: easy missed detection of small targets and complex defects; insufficient environmental robustness leading to increased false alarms; difficulty in balancing real-time performance and accuracy; and a lack of training data specifically for underground cable defects that fails to cover diverse operating conditions. Furthermore, existing training data often lacks thermal imaging samples of underground pipeline cables, further hindering the improvement of digital and intelligent operation and maintenance levels.

[0005] To address the aforementioned issues, several similar solutions have been attempted to improve the situation: for example, improving the reliability of suspected fault identification through a human-machine interaction loop involving machine initial screening, remote manual review, and secondary machine fine inspection; or building a comprehensive perception platform using multi-sensor fusion and SLAM to assist fault analysis with multi-source information such as temperature, vision, and point clouds; or combining lightweight recognition with 3D reconstruction to map 2D defects to a 3D model for quantitative measurement; or using generative artificial intelligence to build a synthetic data augmentation training set to alleviate sample scarcity.

[0006] However, in general, the above solutions still have problems to varying degrees, such as insufficient adaptation to underground extreme lighting and shading environments, difficulty in simultaneously meeting the requirements of real-time inference and detection accuracy at the edge, and limited generalization due to differences between defect samples and real working conditions. There is still an urgent need for a cable defect online detection technology solution that can achieve high accuracy, strong robustness and meet the requirements of real-time deployment at the edge in the complex environment of underground pipe laying. Summary of the Invention

[0007] To overcome the shortcomings of the existing technology, the purpose of this invention is to provide an improved underground pipeline circuit fault detection method based on YOLO26, which can realize automatic detection, location, and early warning closed loop of underground pipeline cable defects, and improve the identification accuracy and real-time response.

[0008] To achieve the above objectives, the present invention provides the following solution: A fault detection method for underground pipeline circuits based on YOLO26, comprising: A thermal imaging simulation database was constructed. Based on the normal and fault conditions of the cable, a structural model was established and thermoelectric coupling simulation was performed to obtain thermal maps for each condition. The heat map is standardized and preprocessed, and the fault type and core area are labeled to form the thermal imaging simulation database; An improved YOLO26 model trained based on the thermal imaging simulation database is obtained, and the improved YOLO26 model is deployed in the computing unit of the mobile robot inspection platform. The mobile robot inspection platform is controlled to move autonomously within the underground pipework, enabling the dual-spectrum camera module to simultaneously acquire visible light images and infrared thermal images, and simultaneously obtain pose information. Perform dual-spectral temporal-spatial registration on the visible light image and the infrared thermal image to obtain a registered dual-spectral image; The registered bispectral image is input into the improved YOLO26 model for end-to-end real-time inference to obtain the detection results; the detection results include defect category, bounding box coordinates, and confidence level. When the confidence level is higher than the threshold, the defect location is mapped to the global pipeline map based on the pose information to achieve defect localization. When the defect category belongs to a preset high-risk defect type, an audible and visual alarm is triggered and the defect snapshot and location information are transmitted back via wireless communication. The detection log is also stored locally. When the confidence level is not higher than the threshold, the mobile robot inspection platform is controlled to trigger the gaze mode for verification and data collection.

[0009] Preferably, a thermal imaging simulation database is constructed, and a structural model is established based on the normal and fault conditions of the cable, and thermoelectric coupling simulation is performed to obtain thermal maps for each condition, including: Establish the structural model that includes the cable structure and the duct structure; Set fault area parameters for the structural model corresponding to the normal operating condition and the fault operating condition, respectively. Define material properties and set electric and thermal boundary conditions for the structural model; The structural model is divided into meshes, and the mesh of the fault region is refined. Perform the thermoelectric coupling simulation and extract the thermal map.

[0010] Preferably, the fault condition includes one or more of the following: skin damage, regular burn-out, hollowing-out damage, and severe damage.

[0011] Preferably, the thermal image is standardized and preprocessed, and the fault type and core area are labeled to form the thermal imaging simulation database, including: The heatmap is subjected to standardized preprocessing with a unified format and temperature value representation range; The heat map is labeled based on the fault type; Based on the core region, core region annotation information is generated on the heat map; The heat map, the label annotations, and the core area annotation information are stored together to form the thermal imaging simulation database.

[0012] Preferably, obtaining the improved YOLO26 model trained based on the thermal imaging simulation database includes: Training samples are obtained from the thermal imaging simulation database, and geometric transformations of translation and rotation are performed on the training samples; The training samples are subjected to color-changing processing and noise injection to simulate imaging differences and environmental interference; Perform a trapezoidal transformation on the training samples to simulate changes in shooting angle; The improved YOLO26 model is trained based on the training samples after the geometric transformation, color change processing and noise injection, and trapezoidal transformation processing.

[0013] Preferably, obtaining the improved YOLO26 model trained based on the thermal imaging simulation database includes: When training the improved YOLO26 model, a label allocation strategy prioritizing small targets and occluded targets is adopted; A dynamic loss weight adjustment strategy is used to weight the training loss; A hybrid optimizer is used to update the parameters of the improved YOLO26 model to enhance its ability to distinguish defect categories and improve convergence stability.

[0014] Preferably, the backbone network of the improved YOLO26 model includes a C2f_DWS_Add module, which divides the input features into a first part and a second part along the channel dimension; the second part of the features is transformed by depthwise separable convolution; the transformed features and the first part of the features are fused by element-wise addition (Add) to obtain fused features; the fused features and the second part of the features are concatenated along the channel dimension (concat) to form the output features.

[0015] Preferably, the improved YOLO26 model is a deployment-optimized model, and the deployment optimization includes: Introduce quantization-aware training to obtain quantization model parameters; Batch normalization layers are integrated before quantization to reduce inter-layer redundancy; A weight magnitude-based pruning method is used to remove redundant weights or channels, and a channel retention threshold is set to ensure feature extraction capability.

[0016] Preferably, the detection head of the improved YOLO26 model is configured to output the detection results based on multi-scale feature maps, and to suppress or crop the detection output used for small-scale feature maps, retaining only the detection output of medium-scale and large-scale feature maps.

[0017] Preferably, the gaze pattern includes: Control the mobile robot inspection platform to pause briefly or move at a low speed at the current position; The dual-spectrum camera module is controlled to acquire multiple frames of visible light images and infrared thermal images from multiple observation angles; For each frame of the visible light image and the corresponding frame of the infrared thermal image, perform the dual-spectral temporal-spatial registration to obtain multiple sets of registered dual-spectral images; The registered bispectral images of each group are input into the improved YOLO26 model for end-to-end real-time inference to obtain multiple sets of detection results; Consistency determination is performed based on the defect categories and confidence levels of multiple sets of test results to generate verification test results; When the confidence level of the verification test result is higher than the threshold, the defect location is mapped to the global pipeline map based on the pose information to achieve defect localization. When the defect category of the verification test result belongs to a preset high-risk defect type, an audible and visual alarm is triggered and the defect snapshot and location information are transmitted back via wireless communication, and the test log is stored locally.

[0018] The present invention discloses the following beneficial effects: (1) This invention constructs a thermal imaging simulation database and obtains thermal maps by performing thermoelectric coupling simulation on the normal and fault conditions of cables. Then, it performs standardized preprocessing on the thermal maps and marks the fault types and core areas for training and improving the YOLO26 model. This alleviates the problem of scarce training data caused by the difficulty in obtaining real defect samples and insufficient coverage of working conditions in underground pipeline scenarios from the source. This enables the model training to have scalable sample supply and more complete fault morphology coverage, thereby improving the stability and generalization ability of different fault types.

[0019] (2) The present invention uses a mobile robot inspection platform to move autonomously in underground pipes, and uses a dual-spectrum camera module to simultaneously acquire visible light images and infrared thermal images. The two types of images are then registered in time and space to form a registered dual-spectrum image, which is then input into the model for inference. Thus, under typical pipe conditions such as low light, reflection, occlusion, and cluttered background, complementary detection input can still be obtained, reducing the missed detection and false detection caused by environmental interference in the single imaging mode, and improving the separability of fault features and detection robustness.

[0020] (3) The present invention deploys the improved YOLO26 model on the computing unit of the mobile robot inspection platform, directly performs end-to-end real-time reasoning on the registered bispectral image and outputs the defect category, bounding box coordinates and confidence level, so that the fault identification process is transformed from the traditional offline diagnosis or manual review process to the online automatic judgment process, realizing the timely detection of abnormalities and output of structured detection results during the inspection process, and improving the timeliness and efficiency of underground pipeline cable fault detection.

[0021] (4) When the detection result meets the threshold, the present invention maps the defect location to the global pipeline map based on the pose information, realizes the automatic association between the detection result and the spatial location, avoids the problem of "finding the defect but having difficulty locating the location", reduces subsequent manual retesting and repeated excavation and search, and improves the operability of fault location and the certainty of maintenance decision.

[0022] (5) The present invention divides the detection results into two categories based on the confidence threshold: those that can be directly located and handled and those that need to be verified. When the defect category belongs to the preset high-risk defect type, an audible and visual alarm is triggered, the defect snapshot and location information are transmitted back via wireless communication and the detection log is stored locally. At the same time, when the confidence level does not reach the threshold, the staring mode is triggered for verification and collection. This establishes a closed-loop mechanism between rapid response and reliable confirmation, reduces the cost of invalid handling caused by false alarms and reduces the safety risks caused by missed alarms, while retaining a traceable chain of evidence and process records. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 Thermal simulation diagram of normal cable operation provided for embodiments of the present invention; Figure 3 Thermal simulation diagram of cable damage provided for embodiments of the present invention; Figure 4 This is a structural diagram of the original module C2f provided in an embodiment of the present invention; Figure 5 This is a structural diagram of the improved module CDA provided in an embodiment of the present invention; Figure 6 This is a structural diagram of the original model provided in the embodiments of the present invention; Figure 7 The improved YOLO26 model YOLO-Thermo structure diagram provided in the embodiments of the present invention; Figure 8 The figure shows the simulation results of the improved YOLO26 model YOLO-Thermo provided in the embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments 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, and 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.

[0026] The purpose of this invention is to provide a fault detection method for underground pipeline circuits based on YOLO26. By using simulation database training-driven, bispectral registration inference and pose mapping localization, the method improves fault identification efficiency and reduces false alarms and missed alarms.

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a fault detection method for underground pipe circuits based on an improved YOLO26 standard, comprising: Step 100: Construct a thermal imaging simulation database, establish a structural model based on the normal and fault conditions of the cable, and perform thermoelectric coupling simulation to obtain thermal maps for each condition. Step 200: Standardize and preprocess the heat map and label the fault type and core area to form a thermal imaging simulation database; Step 300: Obtain the improved YOLO26 model trained based on the thermal imaging simulation database, and deploy the improved YOLO26 model in the computing unit of the mobile robot inspection platform; Step 400: Control the mobile robot inspection platform to move autonomously inside the underground pipework, so that the dual-spectrum camera module can simultaneously acquire visible light images and infrared thermal images, and simultaneously acquire pose information; Step 500: Perform dual-spectral temporal-spatial registration on the visible light image and the infrared thermal image to obtain the registered dual-spectral image; Step 600: Input the registered bispectral image into the improved YOLO26 model for end-to-end real-time inference to obtain the detection results; the detection results include defect category, bounding box coordinates and confidence level; Step 700: When the confidence level is higher than the threshold, the defect location is mapped to the global pipeline map based on the pose information to achieve defect localization. When the defect category belongs to the preset high-risk defect type, an audible and visual alarm is triggered and the defect snapshot and location information are transmitted back via wireless communication. The detection log is also stored locally. Step 800: When the confidence level is not higher than the threshold, control the mobile robot inspection platform to trigger the gaze mode for verification and data collection.

[0029] Specifically, this embodiment provides a technical solution for underground duct cable fault detection based on a robot combined with the improved YOLO26 model YOLO-Thermo. The entire technical solution consists of two main parts: a mobile robot inspection platform and an intelligent vision inspection system mounted on it. The specific technical solution is as follows: (I) Overall System Hardware Architecture Mobile robot platform: Body: Adopting a biomimetic "robotic snake" shape or a small wheeled / tracked chassis, it possesses high mobility and passability to adapt to narrow, winding spaces with cable ducts of 150mm diameter and above. The body adopts a lightweight, modular design, with a weight controlled at around 2.5kg, facilitating rapid deployment.

[0030] Perception system: High-definition dual-spectrum camera module: This module integrates a high dynamic range (HDR) RGB visible light camera and an uncooled infrared thermal imager. The visible light camera is used to detect physical deformation, damage, and dirt; the infrared thermal imager is used to capture abnormal heat points caused by overload, poor contact, etc. Dual-spectrum images can be registered and fused at the pixel level.

[0031] Auxiliary lighting: Integrated high-brightness, low-heat LED supplementary lights to cope with the dark environment inside the pipeline.

[0032] Positioning and navigation: Integrated micro encoders or vision-based odometry for real-time positioning and obstacle avoidance within pipes.

[0033] Computing Unit: Equipped with a high-performance, low-power embedded AI computing platform (such as NVIDIA Jetson OrinNX), responsible for running the improved YOLO26 algorithm model YOLO-Thermo of this invention.

[0034] Communication and storage: Equipped with a 5G / broadband wireless communication module for real-time transmission of alarm information and key data; and local storage for caching inspection data.

[0035] (ii) Thermal Imaging Simulation Database To address the shortcomings of scarce and difficult-to-collect real-world samples of underground duct cable faults, and to provide realistic training data for improving the YOLO26 algorithm in this invention, a thermal imaging simulation database based on SolidWorks modeling and ANSYS simulation is constructed, as follows: 1. SolidWorks Modeling SolidWorks modeling is performed to accurately replicate the actual parameters at a 1:1 scale. The model focuses on three main objects: the cable core, the cable sheath, and the fiberglass pipe (inner diameter 180mm, outer diameter 190mm, wall thickness 5mm). For the three-core cable (outer diameter 94mm), detailed modeling is completed for one normal working condition and four fault working conditions (sheath damage, regular burn-out, hollowed-out damage, and severe damage). Each working condition is modeled separately and the dimensional accuracy is repeatedly verified. After modeling defects are repaired, the model is exported in x_t format for import into ANSYS to ensure that the model is free of deviation.

[0036] 2. ANSYS Simulation The ANSYS simulation utilizes the Workbench platform for thermoelectric coupling simulation. Five operating condition models are imported one by one, with precise definition of material properties for each component and targeted adjustments to properties in the fault area. Fine-grained meshing and independence verification are performed, with a focus on densifying critical fault areas. Standardized electric and thermal boundary conditions are set for each model, and the solution and result extraction are completed step-by-step. Repeated debugging throughout the process ensures simulation accuracy. The ANSYS simulation also simulates normal conditions (such as…). Figure 2 (as shown) and four other cable damage scenarios (such as...) Figure 3 As shown in the figure, a heat map and corresponding fault-related data are obtained.

[0037] The heat maps obtained from the simulation are standardized and preprocessed, with unified format and thresholds. Fault types and core areas are accurately labeled, and a structured database is constructed by associating modeling and simulation parameters. After strict verification, a usable thermal imaging simulation database is formed, providing core data support for subsequent algorithm training and preprocessing optimization.

[0038] (III) Improved YOLO26 Algorithm Core Technology Scheme The core of this embodiment is based on the edge optimization features of YOLO26, and three deep integration improvements have been made for the underground cable inspection scenario: ① The YOLO26 core architecture has been lightweighted and adapted for deployment; ② The inspection scale has been adjusted accordingly; ③ Data preprocessing optimization has been performed according to the actual inspection environment and equipment characteristics.

[0039] 1. Lightweighting and adapting the YOLO26 core architecture. This embodiment considers limited device computing power and scarce storage resources. Directly running native YOLO26 for full-image detection of high-resolution pipeline images can easily lead to problems such as wasted computing power, excessive inference latency, and difficulties in storage adaptation. The following solution is provided: (1) Lightweight upgrade of backbone network like Figure 4 As shown, Figure 4 The original C2f module structure is shown: the input features are sequentially passed through the C2f module and the CBS,s=1,k=1 convolutional block before entering the Split branch unit. The Split unit divides the features into two paths: one path enters the Bottleneck module, and the other path enters a branch formed by repeated stacking of Bottleneck modules. The number n in parentheses in the diagram indicates the number of repetitions. The outputs of the two branches are concatenated at node C, and the concatenated features are then output after passing through the CBS,s=1,k=1 convolutional block. Simultaneously... Figure 4The internal structure of Bottleneck is shown below. Bottleneck includes at least two concatenated CBS convolutional blocks with s=1 and k=3. At node a, the convolutional blocks are added element-wise with the bypass features to form a residual fusion output. Here, s=1 indicates a stride of 1, and k=1 and k=3 indicate kernel sizes of 1×1 and 3×3, respectively.

[0040] To achieve a lightweight model suitable for edge hardware such as embedded robots and portable detection terminals, the native C3k2 module of YOLO26 was first replaced with the C2f module. The C2f module extracts features through feature segmentation and multi-branch convolution: first, the input features are segmented by the channel dimension to obtain two parts of features: Next, the standard Bottleneck module (Concat fusion) is repeatedly applied to the segmented features: All intermediate processing results are then concatenated along the channel dimension to complete the fusion: Finally, a 1x1 convolution operation is used to map the fused features to the final output channel number: However, this structure leads to a linear increase in the length of list y with the number of Bottlenecks, a linear increase in the number of channels in the final concat with the number of Bottlenecks, and a significant increase in intermediate feature storage and computation. To further reduce model size and computational complexity while ensuring a constant number of intermediate feature channels, the C2f module was upgraded to a customized C2f_DWS_Add module (CDA for short). CDA removes this intermediate concat and replaces it with Add for fusion and updates x2: (Residual connection); final (The number of channels remains constant) ); Only when splicing x1 and x2: ; This significantly reduces computational complexity.

[0041] The improved structure is as follows Figure 5As shown, the input features are sequentially passed through the CDA module and the CBS,s=1,k=1 convolutional block before entering the Split branch unit. The Split branch unit divides the features into x1 and x2 branches. The x1 branch is directly fed into node C, while the x2 branch enters the feature extraction link composed of the DWS module. The n marked in parentheses in the figure represents the number of repetitions of the DWS module. The outputs of each DWS module and their corresponding bypass features are added element-wise at node A to form a fused feature. The x1 branch and the fused feature are concatenated at node C. The concatenated feature is then output through the CBS,s=1,k=1 convolutional block. Simultaneously, Figure 5 The internal structure of DWS is shown below. DWS includes Conv 1x1, DWConv 3x3 and Conv 1x1 connected in sequence. DWConv 3x3 represents 3×3 depthwise convolution, s=1 indicates stride of 1, and k=1 indicates kernel size of 1×1.

[0042] The symbols are defined as follows: Module input, shape is ( For batch size, Input the number of channels. , (For space dimensions) Number of hidden channels ( (For scaling, the default is 0.5) A 1x1 convolution operation adjusts the number of input channels to... A 1x1 convolution operation maps the concatenated features to the final output channel number. Standard Bottleneck module (containing two 3x3 convolutions) Depthwise separable convolution module (1x1 pointwise convolution + 3x3 depthwise convolution + 1x1 pointwise convolution) Divide the features into two parts along the channel dimension. Features spliced ​​along the channel dimension Element-by-element addition (Add merge) (2) Speed-up optimization of feature enhancement module The SimSPPF module replaces the original SPPF module. By simplifying the pooling and feature integration logic, the calculation speed is significantly improved without losing key features of the fault area. This avoids detection delays caused by insufficient computing power of edge devices, adapts to dynamic working conditions such as robot movement detection, and ensures timely fault warnings.

[0043] Figure 6 The overall network structure of the original model is shown, such as Figure 6 As shown, the network consists of three parts: Backbone, Neck, and Head. In the Backbone, input Images are processed sequentially through ConvP1 (node ​​b0), ConvP1 (node ​​b1), and C3k2 (node ​​b2) for initial feature extraction. Subsequently, multi-scale semantic features are extracted level by level through ConvP3 (node ​​b3), C3k2 (node ​​b4), ConvP4 (node ​​b5), C3k2 (node ​​b6), ConvP5 (node ​​b7), C3k2 (node ​​b8), SPPF (node ​​b9), and C2PSA (node ​​b10). In the Neck, features from node b10 are upsampled (node ​​b11) and then fused with features from node b6 at Concat (node ​​b12). The fused features are then processed through C3k2 (node ​​b13). Upsample again (node ​​b14) and fuse with the features from node b4 at Concat (node ​​b15), then pass through C3k2 (node ​​b16) to form a feature output. Simultaneously, the features from node b16 pass through Conv (node ​​b17) and fuse with the features from node b13 at Concat (node ​​b18), then pass through C3k2 (node ​​b19) to form the features corresponding to scale P4. Subsequently, they pass through Conv (node ​​b20) and fuse with the features from node b8 at Concat (node ​​b21), then pass through C3k2 (node ​​b22) to form the features corresponding to scale P5. In the Head, the features labeled with scales P3, P4, and P5 are input into the corresponding Detection modules to output multi-scale target detection results, where node numbers b0 to b22 indicate the connection and output positions of each feature stream.

[0044] Figure 7 The overall network structure of the improved YOLO26 model, YOLO-Thermo, is shown below. Figure 7As shown, the network also includes a Backbone, Neck, and Head. In the Backbone, the input image, after Data Augmentation preprocessing, passes through Conv P1 (node ​​b0) and Conv P2 (node ​​b1) sequentially. In subsequent feature extraction stages at various scales, a CDA module is introduced, passing through CDA (node ​​b2), Conv P3 (node ​​b3), CDA (node ​​b4), Conv P4 (node ​​b5), CDA (node ​​b6), Conv P5 (node ​​b7), and CDA (node ​​b8) sequentially. Then, it passes through SimSPPF (node ​​b9) and C2PSA (node ​​b10) to obtain high-level features. In the Neck, the features from node b10 are upsampled (node ​​b11) and fused with features from multiple sources, including node b6, at Concat (node ​​b12). The fused features are then passed through C3k2 (node ​​b13) and fused with one Conv (node ​​b14) feature at Concat (node ​​b15), and finally passed through C3k2 with scale label P4. P4 (node ​​b16) forms one output; the features of node b16 are fused with the features from node b12 at Concat (node ​​b18) after passing through Conv (node ​​b17), and then pass through C3k2P5 (node ​​b19) with scale label P5 to form a high-level output; in the Head, the high-level path features from the Neck are input into the Detection module above to obtain the detection output, and the features corresponding to node b16 are input into the Detection module below to obtain another scale detection output. The schematic diagram on the right side of the figure is used to visualize the detection results of the Detection output on the target image. Node numbers b0 to b19 are used to indicate the connection position between the output of each module and the feature flow, and scale labels P1 to P5 are used to indicate the feature levels at different scales.

[0045] (3) Quantitative perception training optimization To further reduce edge deployment overhead, the training process is supplemented with quantization awareness: the INT8 quantization scheme is adopted to reduce model storage volume and inference latency; a calibration mechanism is enabled during training, and the quantization range is statistically analyzed using the min-max observer to accurately control the boundary and improve quantization accuracy; a batch normalization layer is fused before quantization to reduce inter-layer redundancy and ensure that the model is both efficient and reliable on embedded terminals.

[0046] (4) Model pruning optimization To further streamline the model structure and reduce the computing power requirements for edge deployment, a classic model pruning optimization strategy based on weight magnitude is simultaneously incorporated. This strategy reduces model size and computational complexity by removing unimportant weights, channels, or layers from the neural network, while maintaining model performance as much as possible. This deepens the lightweight effect while ensuring the integrity of fault detection functionality: Model pruning is enabled with a target pruning amount of 30% to balance lightweighting with detection accuracy; a weight magnitude-based pruning method is used to accurately identify and remove unimportant weight parameters from the model, reducing redundant computation; a phased pruning rhythm is established, performing pruning every 10 epochs, stopping if model performance does not improve after three consecutive prunings to avoid over-pruning leading to accuracy degradation; and a channel retention threshold is set, retaining at least 8 channels per layer to ensure that core feature extraction and fault identification functions are not affected, forming a closed-loop lightweighting process encompassing "architecture simplification - quantization compression - pruning simplification".

[0047] 2. Targeted adjustment of the detection scale: In the underground duct scenario, the abnormal temperature field induced by cable faults in this embodiment usually exhibits a non-localized distribution due to the heat accumulation effect. Therefore, small-scale temperature anomaly areas are often caused by interference factors such as debris particles around the cable under actual working conditions, which have a certain impact on actual detection. The following solution is provided: Based on the scale distribution characteristics of underground cable faults, redundant detection branches are trimmed: the original small target detection module is removed, and only the medium and large scale branches (P4 / 16 (40×40) and P5 / 32 (20×20)) are retained, which correspond to the medium-scale fault temperature field and the large-scale heat accumulation area, respectively. This can effectively reduce the false detection rate, focus on the core target, and reduce the sensitivity to small abnormal thermal distributions. While avoiding the influence of impurities in the pipeline under actual working conditions, it can also effectively maintain the detection accuracy.

[0048] 3. Data preprocessing optimization based on actual testing environment and equipment characteristics. This embodiment considers the differences between the Ansys simulated thermoelectric coupling database and the actual working conditions of underground pipework, as well as the scarcity of real fault samples and the variability of testing equipment and environment, which may lead to a large data gap and poor model generalization ability. The following solution is provided: A multi-dimensional preprocessing workflow was designed to adapt to the training requirements of the improved YOLO26, specifically as follows: (1) Basic geometric transformation: retain translation and rotation operations to simulate slight positional changes of the detection equipment in the pipeline, expand data diversity to avoid model overfitting, and adapt to the dynamic scenario of robot movement detection.

[0049] (2) Enhanced environmental simulation: Add color-changing processing (simulate the spectral response differences of different models of thermal imagers) and multi-intensity noise (restore the interference of dust and water vapor in the pipeline) to solve the problem of the disconnect between simulation data and actual imaging, and make the model adapt to complex working conditions.

[0050] (3) Viewpoint adaptation modification: Through multi-degree trapezoidal transformation, the actual shooting viewpoint of the thermal imager at different positions and angles in the pipeline is simulated to offset the image distortion caused by the equipment placement deviation and ensure the robustness of the model detection viewpoint.

[0051] As an optional implementation method, the complete workflow of this embodiment is as follows: 1. System initialization: The robot enters the inlet of the cable duct to be inspected, loads the pipeline map or performs SLAM mapping simultaneously, and plans the inspection path.

[0052] 2. Simultaneous Movement and Data Acquisition: The robot moves autonomously along the pipeline, while a dual-spectrum camera module simultaneously acquires visible light and infrared video streams at a fixed frequency. An encoder simultaneously performs environmental perception.

[0053] 3. Image Preprocessing and Registration: Distortion correction and dual-spectral temporal-spatial registration are performed on the acquired images. If necessary, the registered dual-channel information (e.g., RGB + thermal values) is fused into a single multi-channel input. Let the input image be... ,in and 3 represents the image height and width, and 3 represents the number of RGB channels.

[0054] 4. Improve YOLO26 model for real-time inference with YOLO-Thermo: The current frame image undergoes fixed-size scaling and pixel value normalization to the [0,1] range. Then, during the training phase, a Small-Target-Aware Label Assignment (STAL) strategy is integrated to prioritize the assignment of labels for small targets and occluded targets, and data augmentation modules such as random rotation and scaling are used to improve generalization ability. Next, multi-scale features are extracted through the convolutional layers of the backbone network and the C2f_DWS_Add module: First, the input image is downsampled and features are extracted through the initial convolutional layer. Then, a second downsampling was performed: Then, use the C2f_DWS_Add module to extract more complex features: Continue downsampling and feature extraction: Using the C2f_DWS_Add module again: Further downsampling: Using the C2f_DWS_Add module with shortcut: Continue downsampling: Using the C2f_DWS_Add module with shortcut again: Spatial pyramid pooling is performed using the SimSPPF module to enhance feature representation capabilities. Finally, the C2PSA module is used to further extract and fuse features: Next, feature fusion is performed through the neck network: First, the P5 feature map is upsampled to facilitate fusion with the P4 feature map. The upsampled feature map is then concatenated with the P4 feature map from the backbone network: The fused features are processed using the C3k2 module: The processed P4 feature map is downsampled so that it can be fused with the P5 feature map again: The downsampled feature map is concatenated with the P5 feature map from the backbone network: The fused features are processed using the C3k2 module, this time employing an attention mechanism: The processed P4 and P5 feature maps are then input into the Detect module to generate detection results. The Detect module consists of three main parts: target existence prediction, class prediction, and bounding box regression.

[0055] Finally, non-maximum suppression is applied to the detection results to remove overlapping detection boxes, resulting in the final detection result: The final output is a collection containing the target category, bounding box coordinates, and confidence score. in, It is a category index. These are the bounding box coordinates (x, y, w, h). It is the confidence score.

[0056] During training, ProgLoss and MuSGD optimizers are used to improve performance, and the model is pre-trained on mainstream datasets. Finally, multi-task results are directly output, including object detection (no redundant bounding boxes, class labels, and confidence scores; anchor-free + NMS-free), instance segmentation, pose / keypoint estimation, orientation detection, and image classification (single-label results and confidence scores). Throughout the process, the C2f_DWS_Add module reduces computational complexity through depthwise separable convolutions and additive fusion while maintaining feature extraction capabilities, making the model more suitable for deployment on resource-constrained devices.

[0057] 5. Results Analysis and Decision Making: Figure 8 The visualization results of the improved YOLO26 model YOLO-Thermo's detection output on simulated samples are shown, such as... Figure 8 As shown in the figure, four example sub-images illustrate the model's identification and localization of cable sheath damage targets. Each example image is labeled with its corresponding image filename at the top. A blue rectangle represents the target bounding box output by the model, and the category identifier "0" next to the rectangle indicates the model's category prediction result for the detected target, characterizing the model's detection performance on the cable sheath damage area under different enhanced sample conditions. This embodiment records defects with a confidence level higher than a threshold. Combined with the robot's real-time pose, coordinate transformation maps the defect location in the image to the global pipeline map, achieving precise localization of "a type Y defect exists in the cable at meter X on a certain pipeline, in the upper left corner." For high-risk defects (such as severe overheating or breakage), an audible and visual alarm is immediately triggered, and key data such as defect snapshots, location information, and infrared temperature curves are packaged and sent to the monitoring center via a wireless network. All detection logs are stored locally for later big data analysis and model iteration optimization. For suspicious targets with low confidence, the robot can automatically trigger a "stare" mode, briefly pausing to perform multi-angle verification and imaging, improving diagnostic accuracy.

[0058] In this embodiment, the improved YOLO26 model YOLO-Thermo is applied to the visual detection scenario of underground duct cable faults. The overall solution design focuses on model lightweighting, end-to-end deployment, and improved environmental robustness. Specifically, at the backbone network level, the model reduces the number of parameters and computational complexity by lightweighting the feature fusion method and convolution operators. At the feature enhancement level, the ability to represent complex defects is improved by optimizing the multi-scale context aggregation module. Furthermore, deployment optimization methods such as quantized perception training and model pruning are combined to further compress the model size, thereby achieving real-time inference with high detection accuracy under embedded edge computing conditions. Simultaneously, through dual-spectral (visible light + infrared) information collaboration and registration input, the stability and reliability of the system are improved in underground environments with varying lighting, occlusion, and complex backgrounds, reducing the risk of false alarms and improving the consistency of defect discrimination.

[0059] Based on the aforementioned model and system configuration, this embodiment establishes a fully automated workflow for underground cable ducts. This includes autonomous navigation and pose acquisition within the duct, synchronous acquisition of dual-spectral data, real-time inference output of defect categories and location information from registered dual-spectral images, triggering audible and visual alarms for high-risk defects and transmitting defect snapshots and location information, while simultaneously triggering staring mode verification for low-confidence results to improve judgment reliability. Thus, this embodiment achieves a closed-loop operation from inspection, detection, location to early warning reporting, significantly reducing model size and computational overhead while maintaining manageable accuracy loss, and balancing real-time performance at the edge and robustness in complex environments.

[0060] Without departing from the core concept and basic principles of this invention, those skilled in the art can make appropriate modifications or equivalent substitutions to the model structure and training strategy to achieve similar technical effects. For example, in the implementation of the feature enhancement module, while retaining the basic logic of multi-scale pooling to extract global and local contextual information, the number of pooling branches or pooling windows can be reduced, and the feature fusion process after pooling output can be simplified, for example, by using a lighter fusion operator or reducing the computational load after concatenation. This further reduces edge-side inference overhead and latency without significantly weakening the ability to extract key features in fault regions. In terms of label allocation and loss optimization, an adaptive training sample selection strategy can be adopted to achieve dynamic partitioning of positive and negative samples. For example, the ATSS (Adaptive Training Sample Selection) strategy can be introduced to adaptively select positive samples based on the IoU distribution between the target and candidate anchor boxes, so that the training process automatically focuses on small targets and occluded targets, reduces the adaptation bias caused by manual threshold setting, and improves training stability and generalization ability. Regarding the configuration of detection scale, an adaptive threshold filtering mechanism can be introduced while retaining the small-scale detection branch. This mechanism dynamically filters candidate targets output from the small-scale branch using multi-dimensional criteria such as heat value and shape features to eliminate false anomalies caused by pipeline impurities or minor noise, thereby reducing the probability of missing genuine minor faults while controlling false detections. The above modifications or substitutions are all optional implementations of this invention, and all can achieve a balance between accuracy and real-time performance in fault detection under different computing power deployment conditions.

[0061] The beneficial effects of this invention are as follows: 1. A cable thermal imaging simulation library was established: using SolidWorks, three-core cables and fiberglass pipes were accurately replicated at a 1:1 scale, and detailed models of one normal working condition and four fault conditions were completed and exported. Subsequently, thermoelectric coupling simulation was performed using ANSYS Workbench. After material definition, mesh refinement, and boundary condition settings, thermal maps and related data for each working condition were generated. Finally, after standardized preprocessing and precise annotation, a structured database was constructed, which made up for the shortcomings of scarce real samples of underground duct cable faults and the difficulty of collection, and provided realistic training data for improving the YOLO26 algorithm.

[0062] 2. Significantly Improved Small Object Detection Accuracy: By prioritizing small object resources during label allocation using the STAL strategy, dynamically adjusting loss weights with ProgLoss to avoid losing sound during small object training, optimizing multi-scale feature extraction and fusion to preserve small object details and semantic information, and employing NMS-free end-to-end inference to reduce false negatives of small objects, these improvements collectively enhance the average accuracy of small object detection, reduce the false negative and false negative rates, and balance detection accuracy with real-time performance. 3. Excellent environmental robustness and generalization ability: ProgLoss dynamic balancing of loss weights avoids easy samples dominating training, and STAL prioritizes the allocation of labels for small and occluded targets, improving detection stability in complex scenes; multi-scale feature extraction and bidirectional fusion design enable it to adapt to variations in target size in different environments, taking into account both details and semantic information; the MuSGD hybrid optimizer enhances the stability of training convergence, allowing the model to maintain reliable performance on highly variable datasets and broaden its cross-scene adaptability; removing the DFL module and adopting NMS-free end-to-end inference reduces the impact of environmental interference on inference results and improves environmental adaptability during deployment; combined with data augmentation modules and high-precision preservation characteristics under INT8 / FP16 quantization, the model can cope with target pose and illumination variations in different environments, and can stably perform in diverse hardware environments such as edge devices, significantly improving generalization ability and environmental robustness compared to the previous generation model.

[0063] 4. Achieving a balance between high accuracy and real-time performance: The architecture removes the DFL module to simplify bounding box regression and employs end-to-end NMS-free inference to eliminate post-processing bottlenecks. While reducing computational overhead and latency, it leverages ProgLoss to dynamically balance loss weights and STAL to prioritize small / occluded target labels, significantly improving detection accuracy for small targets and complex scenes. In feature processing, multi-scale feature extraction and bidirectional fusion are used to balance target details and semantic information, enhancing feature representation capabilities while maintaining lightweight computation. The MuSGD hybrid optimizer is introduced on the training side to accelerate convergence and improve training stability, ensuring the model achieves high accuracy in fewer epochs. On the deployment side, INT8 / FP16 quantization and multi-format native export are supported, adapting to various hardware while further compressing computational costs. Ultimately, a detection mAP of up to 57.5% is achieved on datasets such as COCO, with GPU inference latency as low as 1.7ms / img, successfully balancing cutting-edge detection accuracy with the real-time response capabilities required for edge scenarios.

[0064] 5. End-to-end architecture simplifies deployment and optimization: Removing the DFL module simplifies bounding box regression tasks, improves CPU inference speed by 43%, and cross-platform export requires no additional adaptation. At the same time, end-to-end NMS-free inference reduces inference latency by 20%-30% and avoids deployment issues caused by threshold tuning, breaking through the bottlenecks of efficiency and deployment compatibility of the previous generation model.

[0065] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0066] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting faults in underground pipework circuits based on an improved version of YOLO26, characterized in that, include: A thermal imaging simulation database was constructed. Based on the normal and fault conditions of the cable, a structural model was established and thermoelectric coupling simulation was performed to obtain thermal maps for each condition. The heat map is standardized and preprocessed, and the fault type and core area are labeled to form the thermal imaging simulation database; An improved YOLO26 model trained based on the thermal imaging simulation database is obtained, and the improved YOLO26 model is deployed in the computing unit of the mobile robot inspection platform. The mobile robot inspection platform is controlled to move autonomously within the underground pipework, enabling the dual-spectrum camera module to simultaneously acquire visible light images and infrared thermal images, and simultaneously obtain pose information. Perform dual-spectral temporal-spatial registration on the visible light image and the infrared thermal image to obtain a registered dual-spectral image; The registered bispectral image is input into the improved YOLO26 model for end-to-end real-time inference to obtain the detection results; the detection results include defect category, bounding box coordinates, and confidence level. When the confidence level is higher than the threshold, the defect location is mapped to the global pipeline map based on the pose information to achieve defect localization. When the defect category belongs to a preset high-risk defect type, an audible and visual alarm is triggered and the defect snapshot and location information are transmitted back via wireless communication. The detection log is also stored locally. When the confidence level is not higher than the threshold, the mobile robot inspection platform is controlled to trigger the gaze mode for verification and data collection.

2. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, A thermal imaging simulation database was constructed. Based on the normal and fault conditions of the cable, a structural model was established and thermoelectric coupling simulation was performed to obtain thermal maps for each condition, including: Establish the structural model that includes the cable structure and the duct structure; Set fault area parameters for the structural model corresponding to the normal operating condition and the fault operating condition, respectively. Define material properties and set electric and thermal boundary conditions for the structural model; The structural model is divided into meshes, and the mesh of the fault region is refined. Perform the thermoelectric coupling simulation and extract the thermal map.

3. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, The fault conditions include one or more of the following: skin damage, regular burn-out, hollowing-out damage, and severe damage.

4. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, The heat map is standardized and preprocessed, and fault types and core areas are labeled to form the thermal imaging simulation database, including: The heatmap is subjected to standardized preprocessing with a unified format and temperature value representation range; The heat map is labeled based on the fault type; Based on the core region, core region annotation information is generated on the heat map; The heat map, the label annotations, and the core area annotation information are stored together to form the thermal imaging simulation database.

5. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, Obtaining the improved YOLO26 model trained based on the aforementioned thermal imaging simulation database includes: Training samples are obtained from the thermal imaging simulation database, and geometric transformations of translation and rotation are performed on the training samples; The training samples are subjected to color-changing processing and noise injection to simulate imaging differences and environmental interference; Perform a trapezoidal transformation on the training samples to simulate changes in shooting angle; The improved YOLO26 model is trained based on the training samples after the geometric transformation, color change processing and noise injection, and trapezoidal transformation processing.

6. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, Obtaining the improved YOLO26 model trained based on the aforementioned thermal imaging simulation database includes: When training the improved YOLO26 model, a label allocation strategy prioritizing small targets and occluded targets is adopted; A dynamic loss weight adjustment strategy is used to weight the training loss; A hybrid optimizer is used to update the parameters of the improved YOLO26 model to enhance its ability to distinguish defect categories and improve convergence stability.

7. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, The backbone network of the improved YOLO26 model includes a C2f_DWS_Add module. The C2f_DWS_Add module divides the input features into a first part and a second part along the channel dimension. The second part of the features is transformed by depthwise separable convolution. The transformed features and the first part of the features are fused by adding them element-wise to obtain fused features. The fused features and the second part of the features are concatenated along the channel dimension to form the output features.

8. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, The improved YOLO26 model is a deployment-optimized model, and the deployment optimization includes: Introduce quantization-aware training to obtain quantization model parameters; Batch normalization layers are integrated before quantization to reduce inter-layer redundancy; A weight magnitude-based pruning method is used to remove redundant weights or channels, and a channel retention threshold is set to ensure feature extraction capability.

9. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, The detection head of the improved YOLO26 model is configured to output the detection results based on multi-scale feature maps, and to suppress or crop the detection output used for small-scale feature maps, retaining only the detection output of medium-scale and large-scale feature maps.

10. The underground pipeline circuit fault detection method based on YOLO26 as described in claim 1, characterized in that, The gaze patterns include: Control the mobile robot inspection platform to pause briefly or move at a low speed at the current position; The dual-spectrum camera module is controlled to acquire multiple frames of visible light images and infrared thermal images from multiple observation angles; For each frame of the visible light image and the corresponding frame of the infrared thermal image, perform the dual-spectral temporal-spatial registration to obtain multiple sets of registered dual-spectral images; The registered bispectral images of each group are input into the improved YOLO26 model for end-to-end real-time inference to obtain multiple sets of detection results; Consistency determination is performed based on the defect categories and confidence levels of multiple sets of test results to generate verification test results; When the confidence level of the verification test result is higher than the threshold, the defect location is mapped to the global pipeline map based on the pose information to achieve defect localization. When the defect category of the verification test result belongs to a preset high-risk defect type, an audible and visual alarm is triggered and the defect snapshot and location information are transmitted back via wireless communication, and the test log is stored locally.

Images