Cable tunnel inspection method and system based on deep learning

Abstract

The invention relates to the technical field of cable tunnel inspection, in particular to a cable tunnel inspection method and system based on deep learning, a cable image is collected through an inspection robot, automatic defect identification is realized based on a convolutional neural network, and further, temperature, humidity and gas concentration parameters of a defect position are combined, so that the defect position is identified. And splicing with image features to form a fusion feature vector, so that more accurate defect cause judgment is realized. And inputting the fusion feature vector into an inference model embedded with a prior rule, and outputting a cause category and a confidence score in combination with domain knowledge. And reversely positioning the cause source according to the cause label, and controlling the visual angle of the robot to obtain a corresponding image and identify the image to form closed-loop correlation of the defect and the cause source. And finally, in combination with the three-dimensional map and the pose data, the cause source is mapped into a three-dimensional coordinate point, a spatial causal relationship is established, and automatic and closed-loop intelligent inspection work of the cable tunnel is completed.

Description

Technical Field

[0001] This invention relates to the field of cable tunnel inspection technology, and more particularly to a cable tunnel inspection method and system based on deep learning. Background Technology

[0002] Cable tunnels are vital transmission channels for power systems, and their operational safety directly impacts the stability of urban power supply. Due to the enclosed environment and complex structure of cables within tunnels, they often face adverse conditions such as high humidity, high temperature, and the accumulation of harmful gases, which can easily lead to various defects such as insulation aging, surface damage, and even thermal breakdown. Failure to detect and address these defects in a timely manner can result in major safety accidents such as electrical fires and power outages.

[0003] In existing technologies, defect inspection of cable tunnels mainly relies on regular manual inspections or image acquisition using handheld devices. Maintenance personnel then visually inspect these images based on their experience to determine if any abnormalities exist on the cable surface. This method is not only inefficient and labor-intensive, but also highly subjective, easily leading to missed detections or misjudgments. To improve inspection efficiency, some systems have begun to introduce image recognition technology for automatic classification and detection of cable surface images. However, this approach has the following shortcomings: relying solely on image information makes it difficult to analyze the causes of defects by combining environmental parameters (such as temperature, humidity, and gas concentration) at the defect location, lacking a holistic understanding of the defect formation context. Traditional methods only output the defect type and location, failing to further pinpoint the causal source of the defect, thus requiring further manual on-site inspection for judgment. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a cable tunnel inspection method and system based on deep learning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A deep learning-based cable tunnel inspection method includes the following steps:

[0007] S1. Obtain cable characterization inspection images through the inspection robot, perform classification and regression on the cable characterization inspection images based on the convolutional neural network, and output the defect identification results.

[0008] S2. Based on the defect identification results, the humidity, temperature and gas concentration parameters at the location of the defect are collected synchronously in the inspection robot, and these parameters are spliced ​​with the image feature vector to construct a fused feature vector.

[0009] S3. A defect cause reasoning model based on vector space matching and prior rule embedding performs cause judgment processing on the fused feature vector and outputs the predicted cause category label and the corresponding confidence score.

[0010] S4. Determine the defect cause set based on the predicted cause category label and the corresponding confidence score, and control the inspection robot to adjust the image acquisition orientation according to the defect area location, acquire image data corresponding to the defect cause set, perform image analysis, and output the bounding box and type label of the suspected cause.

[0011] S5. Based on the current pose information of the inspection robot and the 3D map data of the cable tunnel, the bounding box of the suspected cause is mapped to a 3D coordinate point, and the spatial causal relationship between the defect area and the cause is established. The inspection results containing the defect location, cause label and cause are output.

[0012] Further, S1 includes the following steps:

[0013] The inspection robot acquires cable characterization inspection images through its image acquisition device, and performs size normalization and pixel normalization processing on the images to obtain standardized images.

[0014] The standardized image is input into a convolutional neural network, which sequentially performs convolution operations, activation function processing, and downsampling operations to extract multi-scale image features.

[0015] Based on the extracted image features, the detection head in the defect recognition model is used to classify the defect type and regress the bounding box position of each candidate region to obtain the defect recognition result.

[0016] Furthermore, the defect types include cracks, scratches, wear, burn marks, and water stains.

[0017] Further, S2 includes the following steps:

[0018] Based on the location of the defect area indicated by the defect identification results, the humidity, temperature and gas concentration parameters at that location are simultaneously collected from the humidity sensor, temperature sensor and gas concentration sensor configured in the inspection robot.

[0019] The humidity value, temperature value, and gas concentration parameter are encoded into an environmental parameter vector, and the environmental parameter vector is aligned based on the timestamp and location coordinates.

[0020] The environmental parameter vector is concatenated with the image feature vector to obtain a fused feature vector.

[0021] Furthermore, the defect cause reasoning model is constructed through the following steps:

[0022] Based on historical inspection records, the original cable characterization images and corresponding humidity, temperature and gas concentration parameters are obtained, and mapping pairs are established with manually labeled defect cause categories to form an original training sample set with actual cause category labels.

[0023] Feature extraction and numerical encoding are performed on the original training samples. Image features are mapped into high-dimensional vectors through convolutional encoding, and environmental parameters are mapped into low-dimensional parameter vectors through normalization and embedding functions. The two types of vectors are then concatenated in a unified feature space to generate a fused feature vector training set.

[0024] Based on the training set of fused feature vectors, vector space clustering is performed on the fused feature vectors to obtain the vector center of each cause category. Then, similarity calculation is performed on the fused feature vectors to obtain the similarity distribution representing the matching degree of each cause category.

[0025] By introducing a prior rule embedding mechanism, logical rules extracted from domain knowledge are superimposed on the causal probability distribution in the form of constraint functions to obtain the rule-corrected probability distribution;

[0026] Based on the rule-corrected probability distribution, the cause category corresponding to the highest probability is extracted as the predicted cause category label, and its probability value is used as the corresponding confidence score.

[0027] Based on the supervision error between the rule-corrected probability distribution and the actual cause category label, a joint loss function is constructed, and the model parameters are optimized by backpropagation according to the joint loss function to obtain the defect cause inference model.

[0028] Further, S4 includes the following steps:

[0029] Based on the predicted cause category labels and corresponding confidence scores, and combined with the current posture information of the inspection robot, a set of target cause sources associated with the location of the defect area is constructed.

[0030] Based on the relative positional relationship between the target cause set and the current defect region in the image coordinate system, spatial orientation analysis is performed to determine the target orientation vector for camera view adjustment, and the desired camera rotation parameters are calculated in combination with the robot's current pose.

[0031] The camera module of the inspection robot is controlled to perform attitude adjustment operation according to the desired rotation parameters, so that its viewing angle is directed towards the candidate region containing the target cause, and image data of the candidate region is acquired;

[0032] The image data of the candidate region is input into a residual neural network to extract feature maps representing suspected causal sources, and defect correlation identification processing is performed.

[0033] Based on the feature map output by the image recognition model, the classifier identifies target objects in the image that are related to the cause category label, and performs bounding box regression on them to output the bounding box and type label of the corresponding suspected cause.

[0034] Furthermore, the residual neural network is trained through the following steps:

[0035] Based on manually annotated defect area images and their corresponding defect type labels in historical inspection data, images of physical environment objects spatially adjacent to defects of the types of cracks, scratches, wear, burn marks and water stains are extracted to construct an image sample set containing suspected causal sources. The corresponding causal source category labels are deduced according to the defect type to form causal source image training sample pairs.

[0036] The training sample pairs of the causal source images are input into the residual neural network structure, and convolution extraction, residual connection and feature fusion operations are performed in sequence to generate a high-dimensional embedding vector containing structural differences and appearance features.

[0037] Based on the high-dimensional embedding vector, the classification branch and the regression branch are connected respectively. The classification branch is used to output the type label of the causal source type, and the regression branch is used to output the bounding box position parameters of the target region. The residual neural network loss function is constructed with cross-entropy loss and bounding box regression loss, and backpropagation optimization is performed to complete the learning of residual neural network parameters.

[0038] Furthermore, the loss function for building the residual neural network is as follows:

[0039]

[0040] Where L is the loss value; To predict the causal source type label Cross-entropy loss between the actual causal source type label y; To predict bounding box parameters The smoothed L1 loss between the true bounding box parameter b; λ1 and λ2 are the loss weight coefficients.

[0041] Furthermore, establishing the spatial causal relationship between the defective region and the causal source includes:

[0042] Based on the three-dimensional spatial location of the defect area and the causal source, the spatial distance, direction vector and occlusion factor are calculated to establish a spatial causal correlation index. Combined with the cause label and confidence score, a structured inspection output result is constructed, which outputs spatial causal relationship information including defect location, cause category, cause source type and three-dimensional coordinates.

[0043] A deep learning-based cable tunnel inspection system, applied to any of the aforementioned deep learning-based cable tunnel inspection methods, includes:

[0044] The defect detection module is used to acquire cable characterization inspection images through the inspection robot, perform classification and regression on the cable characterization inspection images based on a convolutional neural network, and output the defect identification results.

[0045] The feature fusion module is used to construct a fused feature vector by synchronously collecting humidity, temperature and gas concentration parameters at the location of the defect in the inspection robot based on the defect identification results;

[0046] The defect cause prediction module is used to perform cause judgment processing on the fused feature vector based on the defect cause reasoning model of vector space matching and prior rule embedding, and output the predicted cause category label and the corresponding confidence score.

[0047] The defect cause finding module is used to determine the defect cause source set based on the predicted cause category label and the corresponding confidence score, and control the inspection robot to adjust the image acquisition orientation according to the defect area location, acquire image data corresponding to the defect cause source set, perform image analysis, and output the bounding box and type label of the suspected cause source.

[0048] The inspection result generation module is used to map the bounding boxes of suspected causal sources to three-dimensional coordinate points based on the current pose information of the inspection robot and the three-dimensional map data of the cable tunnel, and to establish the spatial causal relationship between the defect area and the causal source, and output the inspection results including the defect location, cause label and causal source.

[0049] The beneficial effects of this invention are as follows: This invention acquires cable characterization inspection images through an inspection robot and automatically identifies defects in these images using a convolutional neural network, significantly improving recognition accuracy and efficiency. This solves the problems of missed detections and misjudgments caused by reliance on manual visual inspection in existing technologies. Furthermore, based on temperature, humidity, and gas concentration parameters synchronously collected at the defect location, a fused feature vector is constructed by concatenating it with the image feature vector. This allows the defect judgment process to not only rely on image information but also incorporate the physical state of the cable's operating environment, overcoming the problem that traditional methods cannot comprehensively determine the cause of defects. The fused feature vector is input into a defect cause reasoning model embedded with prior rules, and combined with domain knowledge, constraint optimization of cause classification is achieved, improving the accuracy and reliability of cause judgment. In addition, by using cause labels to reverse locate suspected causal sources, and combining the inspection robot's perspective adjustment and image analysis functions, the bounding box of the causal target is automatically captured and labeled, achieving closed-loop identification from defect to causal source. Ultimately, by combining 3D maps with robot pose data, a 3D spatial causal mapping between defects and their causes is completed. This fundamentally replaces the method of relying on manual on-site inspection to determine the causes, forming a fully automated closed-loop inspection mechanism from defect identification and cause reasoning to cause localization. Attached Figure Description

[0050] Figure 1 This is a flowchart of the steps of a cable tunnel inspection method based on deep learning in this invention.

[0051] Figure 2 This is a flowchart illustrating the construction steps of the defect cause reasoning model in this invention. Detailed Implementation

[0052] Please see Figures 1-2 As shown, this invention relates to a cable tunnel inspection method based on deep learning, comprising the following steps:

[0053] S1. Obtain cable characterization inspection images through the inspection robot, perform classification and regression on the cable characterization inspection images based on the convolutional neural network, and output the defect identification results.

[0054] S2. Based on the defect identification results, the humidity, temperature and gas concentration parameters at the location of the defect are collected synchronously in the inspection robot, and these parameters are spliced ​​with the image feature vector to construct a fused feature vector.

[0055] S3. A defect cause reasoning model based on vector space matching and prior rule embedding performs cause judgment processing on the fused feature vector and outputs the predicted cause category label and the corresponding confidence score.

[0056] S4. Determine the defect cause set based on the predicted cause category label and the corresponding confidence score, and control the inspection robot to adjust the image acquisition orientation according to the defect area location, acquire image data corresponding to the defect cause set, perform image analysis, and output the bounding box and type label of the suspected cause.

[0057] S5. Based on the current pose information of the inspection robot and the 3D map data of the cable tunnel, the bounding box of the suspected cause is mapped to a 3D coordinate point, and the spatial causal relationship between the defect area and the cause is established. The inspection results containing the defect location, cause label and cause are output.

[0058] In some embodiments, the cable tunnel inspection method is executed by a multi-sensor integrated intelligent inspection robot, which operates continuously and autonomously in the tunnel environment, completing the entire process of image acquisition, synchronous measurement of environmental parameters, and intelligent defect analysis. Specifically, during operation, the inspection robot acquires characterizing image data of the cable using an onboard high-resolution industrial camera. Low-light enhancement and dynamic exposure control strategies are introduced during image acquisition to improve imaging quality in high-humidity or low-light environments. For defect identification in cable images, the system employs a pre-trained convolutional neural network model (such as a ResNet50 structure) to perform multi-class defect classification and regression localization on image samples, identifying common defect types including cracks, scratches, wear, burn marks, and water stains. The output defect identification results include not only defect type labels but also bounding box coordinates and defect probability scores for subsequent causal analysis. Unlike traditional methods that rely solely on images, this embodiment simultaneously collects environmental parameters of the defect location using temperature and humidity sensors and gas sensors integrated into the robot, and concatenates these parameters with the corresponding image feature vectors to form a fused feature vector containing multimodal features. The fused feature is input into the constructed defect cause inference model. This model is implemented using a residual neural network structure, with a prior rule vector based on fault mechanism summarization embedded in the penultimate layer. This vector constrains and guides the training process, improving the model's accuracy and generalization ability in identifying cause categories. The model ultimately outputs a cause category label (e.g., overheating, insulation aging, gas corrosion, etc.) and its confidence score corresponding to each fused feature vector. For defect identification in cable images, the system uses a pre-trained convolutional neural network model (e.g., ResNet50 structure) to perform multi-class defect classification and regression localization on image samples, identifying common defect types including cracks, scratches, wear, burn marks, and water stains. The output defect identification results include not only defect type labels but also bounding box coordinates and defect probability scores for subsequent cause analysis. Unlike traditional methods that rely solely on images, this embodiment simultaneously collects environmental parameters of the defect occurrence area using temperature and humidity sensors and gas sensors integrated into the robot, and concatenates these parameters with the corresponding image feature vectors to form a fused feature vector containing multimodal features. The fused feature is used as input to the constructed defect cause inference model. The model is implemented using a residual neural network structure, with a prior rule vector based on the fault mechanism summary embedded in the penultimate layer. This vector constrains and guides the training process, improving the model's accuracy and generalization ability in classifying cause categories. The model ultimately outputs a cause category label (e.g., overheating, insulation aging, gas corrosion, etc.) and its confidence score corresponding to each fused feature vector.Further, based on the output cause category labels, the system retrieves corresponding reference templates from the causative source image database and controls the inspection robot to make slight turning adjustments around the defect as the center. New image data is acquired within the area covered by the robot's view, and image analysis is performed through a target detection network. Bounding boxes and type labels for suspected causative sources are automatically labeled in the images. Finally, to construct a spatial causal map of the defect and its causative sources, the bounding box position information is jointly calculated using the robot's pose (position + attitude) and a 3D map of the cable tunnel. This converts the image coordinates into 3D spatial points and establishes a spatial association with the defect occurrence area, outputting structured inspection results.

[0059] Further, S1 includes the following steps:

[0060] The inspection robot acquires cable characterization inspection images through its image acquisition device, and performs size normalization and pixel normalization processing on the images to obtain standardized images.

[0061] The standardized image is input into a convolutional neural network, which sequentially performs convolution operations, activation function processing, and downsampling operations to extract multi-scale image features.

[0062] Based on the extracted image features, the detection head in the defect recognition model is used to classify the defect type and regress the bounding box position of each candidate region to obtain the defect recognition result.

[0063] In some embodiments, firstly, the image acquisition device carried by the inspection robot continuously photographs the surface of the cable inside the cable tunnel to obtain original cable characterization inspection images. To ensure the uniformity and generalization ability of the input image in the neural network, the original image needs to undergo size normalization and pixel normalization processing (e.g., linear normalization to [0,1] or standardization to a distribution with a mean of 0 and a variance of 1) to obtain a standardized image. The standardized image is input into a convolutional neural network model, which generally adopts a multi-layer convolutional structure and embeds residual connections and batch normalization modules to enhance the network's ability to perceive complex textures in the image. Specifically, the image sequentially passes through several convolutional layers to extract low, medium, and high-level features such as edges, textures, and shapes, and performs nonlinear mapping through the ReLU activation function. After pooling layers (such as MaxPooling), downsampling operations are performed to compress the feature dimension and enhance position invariance, finally obtaining a multi-scale image feature tensor. After feature extraction, the detection head structure in the model, based on target detection frameworks such as YOLO or RetinaNet, performs defect type classification tasks (e.g., cracks, scratches, wear, burns, watermarks) and location regression tasks on candidate regions, i.e., regressing their bounding box parameters (center coordinates, width, height). To improve the recognition capability of small target defects, an FPN (Feature Network Pyramid) structure can be introduced to enhance multi-scale feature fusion. Finally, the defect label and bounding box set of each target are output to form the defect recognition result, serving as the input basis for subsequent defect cause determination.

[0064] Furthermore, the defect types include cracks, scratches, wear, burn marks, and water stains.

[0065] It should be noted that, in the preferred embodiment, the output category labels of the defect recognition model cover five common defect types: cracks, scratches, wear, burn marks, and watermarks. To achieve accurate classification of these defect types, a multi-label image recognition dataset is first constructed during the training phase. Each defect type contains a large number of image samples annotated by experts, and data augmentation processing (such as random rotation, mirror flipping, and brightness perturbation) is applied to improve the model's robustness to varying environments. The convolutional neural network adopts a target detection structure based on the ResNet-50 backbone. A dual-branch module is introduced in the detection head to handle bounding box regression and defect type classification respectively. The classification branch outputs a five-dimensional Softmax probability vector through a fully connected layer, representing the confidence probability that a candidate region in the image belongs to each defect type. During network training, the classification branch uses the cross-entropy loss function, the regression branch uses the smoothed L1 loss function, and a multi-scale feature fusion strategy (such as FPN) is combined to improve the detection performance of fine-grained defects such as small-scale scratches and weak watermarks. During the inference phase, candidate defect regions with confidence scores higher than a preset threshold (e.g., 0.6) are selected by using the bounding boxes output by the convolutional network and their corresponding type probability values, and the highest confidence category is used as its initial identification label. This identification result not only provides basic labels for subsequent defect cause inference but also avoids the uncertainty brought about by human subjective judgment.

[0066] Further, S2 includes the following steps:

[0067] Based on the location of the defect area indicated by the defect identification results, the humidity, temperature and gas concentration parameters at that location are simultaneously collected from the humidity sensor, temperature sensor and gas concentration sensor configured in the inspection robot.

[0068] The humidity value, temperature value, and gas concentration parameter are encoded into an environmental parameter vector, and the environmental parameter vector is aligned based on the timestamp and location coordinates.

[0069] The environmental parameter vector is concatenated with the image feature vector to obtain a fused feature vector.

[0070] In some embodiments, for the defect location identified by the convolutional neural network, humidity sensors, temperature sensors, and gas concentration sensors deployed on the inspection robot are invoked to synchronously trigger acquisition commands under the spatial coordinates of the defect area, thereby acquiring environmental physical parameter values ​​at the defect location, including humidity, temperature, and gas concentration. To ensure alignment with image features during the data fusion stage, the system adds a timestamp and spatial coordinate identifier to each set of acquired data, and uses position interpolation and time window alignment strategies to ensure a strict one-to-one correspondence between environmental data and corresponding image frames. The aforementioned humidity, temperature, and gas concentration parameters are uniformly normalized into a [0,1] interval vector during the encoding stage, forming a 3D environmental parameter vector. Subsequently, the system performs feature concatenation operations on this environmental parameter vector and the high-dimensional image feature vector output by the image feature extraction module. In this embodiment, a vector-level concatenation method is used to fuse the two, constructing a fused feature vector, which is used to characterize the joint feature expression of the image structure and environmental physical state of the defect area. This fused feature vector serves as the input to the subsequent defect cause inference model, introducing contextual physical environment awareness into the model. Unlike traditional methods that rely solely on images for cause determination or require manual input of environmental parameters, this method utilizes a fully automated multimodal feature fusion mechanism triggered by synchronous sensors and aligned with temporal and spatial algorithms. This not only improves the accuracy of cause determination but also enhances the ability to analyze complex defect causes in complex scenarios.

[0071] Furthermore, the defect cause reasoning model is constructed through the following steps:

[0072] Based on historical inspection records, the original cable characterization images and corresponding humidity, temperature and gas concentration parameters are obtained, and mapping pairs are established with manually labeled defect cause categories to form an original training sample set with actual cause category labels.

[0073] Feature extraction and numerical encoding are performed on the original training samples. Image features are mapped into high-dimensional vectors through convolutional encoding, and environmental parameters are mapped into low-dimensional parameter vectors through normalization and embedding functions. The two types of vectors are then concatenated in a unified feature space to generate a fused feature vector training set.

[0074] Based on the training set of fused feature vectors, vector space clustering is performed on the fused feature vectors to obtain the vector center of each cause category. Then, similarity calculation is performed on the fused feature vectors to obtain the similarity distribution representing the matching degree of each cause category.

[0075] By introducing a prior rule embedding mechanism, logical rules extracted from domain knowledge are superimposed on the causal probability distribution in the form of constraint functions to obtain the rule-corrected probability distribution;

[0076] Based on the rule-corrected probability distribution, the cause category corresponding to the highest probability is extracted as the predicted cause category label, and its probability value is used as the corresponding confidence score.

[0077] Based on the supervision error between the rule-corrected probability distribution and the actual cause category label, a joint loss function is constructed, and the model parameters are optimized by backpropagation according to the joint loss function to obtain the defect cause inference model.

[0078] In some embodiments, the defect cause inference model is first constructed based on a historical inspection dataset. Each sample set includes a cable characterization image, temperature, humidity, and gas concentration parameters at its corresponding spatial location, and defect cause category labels (such as "high temperature overload," "external force damage," "environmental corrosion," etc.) annotated by on-site engineering experts. This dataset is constructed as a supervised training sample set with clear input features and label mappings. Image feature extraction is achieved through a convolutional neural network. Each characterization image, after size and pixel normalization, is input into a backbone network containing multiple convolutional and pooling layers to extract multi-scale semantic information, ultimately forming a fixed-length high-dimensional image feature vector before a fully connected layer. Environmental parameters (temperature, humidity, gas concentration) are numerically normalized and further input into an embedding function (such as a fully connected layer or a nonlinear projection network) to map into a low-dimensional vector with consistent dimensions. The two are concatenated in the feature dimension to form a fused feature vector, enabling the model to uniformly process image structural features and physical indicators of the operating environment. The fused feature training set is processed through vector space modeling. In this embodiment, clustering algorithms such as K-means or DBSCAN are used to perform unsupervised clustering on the fused feature vectors, calculating the cluster center vector for each type of defect cause. During the inference phase, Euclidean distance or cosine similarity is calculated between the fused features of new input samples and the cluster center vectors of each type, generating a matching degree vector for each cause category as a preliminary probability distribution. To enhance the model's understanding of physical logic consistency, a prior rule embedding mechanism is introduced. This mechanism consists of a set of manually refined physical logic rules (such as "high temperature + burn marks = high temperature overload"), which are transformed into differentiable constraint functions through a rule graph structure and superimposed on the preliminary similarity distribution to form a corrected cause probability distribution. Finally, the model selects the category corresponding to the maximum value from the corrected rule distribution as the prediction result and uses this value as the confidence score. During model training, the loss function is a joint loss function consisting of cross-entropy loss and rule consistency penalty term, used to quantify the deviation between the predicted label and the true label, and backpropagation is performed using the gradient descent algorithm to iteratively optimize the model parameters.

[0079] Further, S4 includes the following steps:

[0080] Based on the predicted cause category labels and corresponding confidence scores, and combined with the current posture information of the inspection robot, a set of target cause sources associated with the location of the defect area is constructed.

[0081] Based on the relative positional relationship between the target cause set and the current defect region in the image coordinate system, spatial orientation analysis is performed to determine the target orientation vector for camera view adjustment, and the desired camera rotation parameters are calculated in combination with the robot's current pose.

[0082] The camera module of the inspection robot is controlled to perform attitude adjustment operation according to the desired rotation parameters, so that its viewing angle is directed towards the candidate region containing the target cause, and image data of the candidate region is acquired;

[0083] The image data of the candidate region is input into a residual neural network to extract feature maps representing suspected causal sources, and defect correlation identification processing is performed.

[0084] Based on the feature map output by the image recognition model, the classifier identifies target objects in the image that are related to the cause category label, and performs bounding box regression on them to output the bounding box and type label of the corresponding suspected cause.

[0085] In some embodiments, firstly, based on the defect prediction cause category labels and corresponding confidence scores generated by S3, and in conjunction with the current posture of the inspection robot (including 3D coordinate position, heading angle, and tilt angle), a set of candidate target causative sources is constructed. This candidate set may include structural entities known to potentially induce specific types of defects, such as cable brackets, joints, and equipment interfaces. The system automatically filters structural component types with high correlation to the current defect type based on a priori established "cause category - structural component type" mapping rule base. Subsequently, the relative positions of the candidate causative sources in the image coordinate system are compared with the defect region positions, and a spatial direction resolution operation is performed. This resolution calculates the relative vector difference and projection angle to form a target direction vector (i.e., the ideal camera line-of-sight vector), and then, combined with the current 3D pose information of the inspection robot, calls the forward kinematics and inverse kinematics module to solve for the desired camera rotation angle and offset parameters. This parameter set is used to guide the robot's image acquisition module to adjust the viewing angle to ensure that the camera can accurately align with the causative source candidate region. After image acquisition is completed, the obtained candidate region images are input into a deep residual neural network for processing. The neural network structure comprises several residual blocks, each consisting of a convolutional layer, a batch normalization layer, and a ReLU activation function, preserving input features to perform feature residual superposition. The feature maps extracted through this deep network structure effectively capture subtle textures, deformations, and structural anomalies in the causative source region, exhibiting strong spatial semantic discrimination capabilities. Based on these feature maps, the system introduces a defect correlation identification module. This module first applies channel attention weighting to the feature maps through a semantic guidance mechanism related to the causative category label, thereby enhancing the feature responses related to the target causative cause. Subsequently, the weighted feature maps are input to an image classifier and a bounding box regressor, which respectively output the prediction results of whether the target causative source category entity exists in the image, and the bounding box coordinates of the target in the image.

[0086] Furthermore, the residual neural network is trained through the following steps:

[0087] Based on manually annotated defect area images and their corresponding defect type labels in historical inspection data, images of physical environment objects spatially adjacent to defects of the types of cracks, scratches, wear, burn marks and water stains are extracted to construct an image sample set containing suspected causal sources. The corresponding causal source category labels are deduced according to the defect type to form causal source image training sample pairs.

[0088] The training sample pairs of the causal source images are input into the residual neural network structure, and convolution extraction, residual connection and feature fusion operations are performed in sequence to generate a high-dimensional embedding vector containing structural differences and appearance features.

[0089] Based on the high-dimensional embedding vector, the classification branch and the regression branch are connected respectively. The classification branch is used to output the type label of the causal source type, and the regression branch is used to output the bounding box position parameters of the target region. The residual neural network loss function is constructed with cross-entropy loss and bounding box regression loss, and backpropagation optimization is performed to complete the learning of residual neural network parameters.

[0090] In some embodiments, firstly, based on manually annotated defect area images and their corresponding defect type labels from historical inspection records, images of physical environment areas adjacent to the spatial location of the defects are extracted, such as potential causative structures like cable brackets, interface terminals, and condensate drains. This extraction process can be aided by spatial projection mapping relationships and 3D environmental models to complete the region delineation. Based on the prior correspondence between each type of defect (e.g., cracks, scratches, wear, burn marks, water stains) and its potential causative sources, the corresponding causative source category label is derived, and neighboring image regions are combined with this category label to construct image training sample pairs, forming a dataset with structural causal association features. Subsequently, the above image training samples are input into the constructed residual neural network model. This network structure uses multi-layer residual block stacking, with each residual block including a standard convolutional layer, a batch normalization layer, and a ReLU activation function. The input image first passes through an initial convolutional layer to extract low-level texture features, and then sequentially passes through multiple residual units to extract higher-level semantic features with stronger structural and aesthetic characteristics. In each residual block, the input feature tensor is element-wise added to the output of the convolutional path, achieving gradient pass-through and deep feature fusion, enhancing the model's performance in identifying subtle appearance differences and complex structural features. To achieve causative source discrimination and localization, the network structure has two branches at the end: a classification branch and a regression branch. The classification branch receives a high-dimensional embedding vector and then connects to a fully connected layer and a Softmax function to output the probability distribution of type labels for causative sources in the image; the regression branch outputs the four-element parameters (center point coordinates and width and height) of the bounding box of the target causative source in the image through convolutional mapping and a bounding box decoder. The entire network training process is optimized with the joint loss function as the objective. This loss function is a weighted combination of cross-entropy loss (used to measure classification accuracy) and smoothing L1 loss (used for bounding box localization error). Through the above training mechanism, the residual neural network can effectively learn the appearance and structural difference features between defect types and their causative sources, achieving accurate identification and localization of suspected causative source targets.

[0091] Furthermore, the loss function of the residual neural network is as follows:

[0092]

[0093] Where L is the loss value; To predict the causal source type label Cross-entropy loss between the actual causal source type label y; To predict bounding box parameters The smoothed L1 loss between the true bounding box parameter b; λ1 and λ2 are the loss weight coefficients.

[0094] In some embodiments, a joint loss function is used as the overall optimization objective. This loss function consists of two parts: a classification loss for the causative source category and a regression loss for bounding box prediction. The classification loss measures the deviation between the network's classification of objects in the image and the actual labels; the regression loss quantifies the accuracy of the network's prediction of the bounding box at the target location. The classification loss is calculated using the cross-entropy method, which is suitable for multi-class classification problems and effectively reflects the matching between the predicted probability distribution and the actual category. Specifically, each image embedding vector output by the network is mapped to the probability values ​​of multiple causative source candidate categories after passing through the classification branch, and the cross-entropy loss measures its accuracy by evaluating the log probability of the true category label in this distribution. The bounding box regression loss is defined using a smoothed L1 function. This loss function uses the squared difference form when the error between the predicted box and the true box is small, and transforms into the linear difference form when the error is large, thereby balancing stability and robustness during training. This loss applies to the center coordinates and width and height dimensions of the predicted box, enabling the network to more accurately fit the location contours of the causative sources in the image. To balance the losses from the two tasks, this embodiment introduces two hyperparameters as weighting coefficients for the classification loss and regression loss, respectively controlling their contribution to the overall loss function. During training, the network performs backpropagation and gradient updates based on this joint loss function, progressively optimizing all learnable parameters.

[0095] Furthermore, establishing the spatial causal relationship between the defective region and the causal source includes:

[0096] Based on the three-dimensional spatial location of the defect area and the causal source, the spatial distance, direction vector and occlusion factor are calculated to establish a spatial causal correlation index. Combined with the cause label and confidence score, a structured inspection output result is constructed, which outputs spatial causal relationship information including defect location, cause category, cause source type and three-dimensional coordinates.

[0097] In some embodiments, firstly, the acquired defect region bounding box position is converted into actual three-dimensional coordinates in the robot's body coordinate system. This conversion process combines the camera's intrinsic parameter matrix, distortion parameters, and current pose parameters to recover the defect's position in three-dimensional space through spatial projection inverse calculation. Subsequently, among the identified suspected causal sources, the three-dimensional position of each target is acquired, and the spatial distance between the defect point and each causal source point is calculated using the three-dimensional Euclidean distance formula. Furthermore, the direction vector from the defect point to the causal source point is analyzed. This direction vector, composed of the defect position pointing to the causal source position, combined with the current perspective of the inspection robot and the camera, can determine whether the causal source is on the visual influence path of the defect region. To further enhance the reliability of causal judgment, an occlusion factor calculation is introduced, which assesses whether there is physical structural occlusion between the defect region and the causal source. Occlusion detection relies on LiDAR data or structured light point clouds. By establishing a straight line connecting the defect and the causal source, the presence of a dense set of points on this line is checked in the three-dimensional point cloud. If an interfering object exists, the occlusion factor is set to a high value, reducing the causal correlation of the causal source. Finally, a structured causal relationship output set is constructed. This output set centers on the defect region, associating it with its identified cause category, matched causal source type, corresponding 3D location coordinates of the causal source, and recording its spatial distance from the defect, orientation information, and occlusion judgment results. To enhance the structured representation of the data, each causal relationship output also includes a confidence score for prediction.

[0098] The present invention also includes a deep learning-based cable tunnel inspection system, applied to any of the foregoing deep learning-based cable tunnel inspection methods, comprising:

[0099] The defect detection module is used to acquire cable characterization inspection images through the inspection robot, perform classification and regression on the cable characterization inspection images based on a convolutional neural network, and output the defect identification results.

[0100] The feature fusion module is used to construct a fused feature vector by synchronously collecting humidity, temperature and gas concentration parameters at the location of the defect in the inspection robot based on the defect identification results;

[0101] The defect cause prediction module is used to perform cause judgment processing on the fused feature vector based on the defect cause reasoning model of vector space matching and prior rule embedding, and output the predicted cause category label and the corresponding confidence score.

[0102] The defect cause finding module is used to determine the defect cause source set based on the predicted cause category label and the corresponding confidence score, and control the inspection robot to adjust the image acquisition orientation according to the defect area location, acquire image data corresponding to the defect cause source set, perform image analysis, and output the bounding box and type label of the suspected cause source.

[0103] The inspection result generation module is used to map the bounding boxes of suspected causal sources to three-dimensional coordinate points based on the current pose information of the inspection robot and the three-dimensional map data of the cable tunnel, and to establish the spatial causal relationship between the defect area and the causal source, and output the inspection results including the defect location, cause label and causal source.

[0104] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for cable tunnel inspection based on deep learning, characterized in that, The method comprises the following steps: S1, acquiring a cable representation inspection image by a patrol robot, classifying and regressing the cable representation inspection image based on a convolutional neural network, and outputting a defect recognition result; S2, according to the defect recognition result, synchronously collecting humidity value, temperature value and gas concentration parameters of the defect occurrence position by the patrol robot, and splicing the parameters with the image feature vector to construct a fusion feature vector; S3, based on a defect cause reasoning model embedded with vector space matching and prior rules, performing cause judgment processing on the fusion feature vector, and outputting a predicted cause category label and a corresponding confidence score; S4, determining a defect cause source set according to the predicted cause category label and the corresponding confidence score, and controlling the patrol robot to adjust the image acquisition orientation according to the defect area position, acquire image data corresponding to the defect cause source set, and perform image analysis, outputting a boundary box and a type label of a suspected cause source; S5, based on the current pose information of the patrol robot and the three-dimensional map data of the cable tunnel, mapping the boundary box of the suspected cause source to a three-dimensional coordinate point, and establishing a spatial causal relationship between the defect area and the cause source, and outputting an inspection result containing the defect position, the cause label and the cause source. 2.The cable tunnel inspection method based on deep learning according to claim 1, wherein, The S1 comprises the following steps: acquiring a cable representation inspection image by an image acquisition device in the patrol robot, and performing size normalization and pixel normalization on the image to obtain a standardized image; inputting the standardized image into a convolutional neural network, sequentially performing convolution operation, activation function processing and downsampling operation, and extracting multi-scale image features; based on the extracted image features, using a detection head in the defect recognition model to classify the defect type and regress the boundary box position of each candidate region, and obtaining a defect recognition result.

3. The cable tunnel inspection method based on deep learning according to claim 2, characterized in that, The defect types include cracks, scratches, wear, burn marks and water marks.

4. The cable tunnel inspection method based on deep learning according to claim 1, characterized in that, The S2 comprises the following steps: According to the defect area position indicated by the defect recognition result, synchronously collecting humidity value, temperature value and gas concentration parameters of the position from the humidity sensor, temperature sensor and gas concentration sensor configured in the patrol robot; encoding the humidity value, temperature value and gas concentration parameter into an environmental parameter vector, and aligning the environmental parameter vector based on the timestamp and position coordinate; splicing the environmental parameter vector and the image feature vector to obtain a fusion feature vector.

5. The cable tunnel inspection method based on deep learning according to claim 3, characterized in that, The defect cause reasoning model is constructed by the following steps: based on historical inspection records, acquiring original cable representation images and corresponding humidity value, temperature value and gas concentration parameters, and mapping them with artificially labeled defect cause categories to form an original training sample set with actual cause category labels; performing feature extraction and numerical encoding on the original training sample, mapping the image features into high-dimensional vectors through convolution encoding, mapping the environmental parameters into low-dimensional parameter vectors through normalization and embedding functions, and splicing the two types of vectors in a unified feature space to generate a fusion feature vector training set; The vector space clustering processing is performed on the fusion feature vectors based on the fusion feature vector training set, vector centers of each cause category are obtained, and similarity calculation processing is further performed on the fusion feature vectors to obtain a similarity distribution representing the matching degree of each cause category; A prior rule embedding mechanism is introduced, logical rules refined from domain knowledge are superimposed in the form of constraint functions in the cause probability distribution, and a rule corrected probability distribution is obtained; Based on the rule corrected probability distribution, the cause category corresponding to the maximum probability is extracted as a predicted cause category label, and the probability value thereof is taken as a corresponding confidence score; Based on the supervision error between the rule corrected probability distribution and the actual cause category label, a joint loss function is constructed, and parameter back propagation optimization processing is performed on the model according to the joint loss function to obtain a defect cause reasoning model.

6. The cable tunnel inspection method based on deep learning according to claim 5, characterized in that, The S4 includes the following steps: According to the predicted cause category label and the corresponding confidence score, the current pose information of the inspection robot is combined to construct a target cause source set associated with the position of the defect area; Based on the relative position relationship between the target cause source set and the current defect area in the image coordinate system, spatial direction analysis is performed to determine a target direction vector of camera view angle adjustment, and the expected rotation parameters of the camera are calculated in combination with the current pose of the robot; The inspection robot camera module is controlled to perform a pose adjustment operation according to the expected rotation parameters, so that its view angle is directed towards a candidate area containing the target cause source, and image data of the candidate area is obtained; The image data of the candidate area is input into a residual neural network to extract a feature spectrum representing a suspected cause source, and defect correlation identification processing is performed; Based on the feature spectrum output by the image recognition model, a classifier is used to identify a target object in the image that is related to the cause category label, and a bounding box regression is performed on the target object to output a bounding box and a type label of the corresponding suspected cause source.

7. The cable tunnel inspection method based on deep learning according to claim 6, characterized in that, The residual neural network is trained by the following steps: Based on the defect area images and their corresponding defect type labels manually labeled in the historical inspection data, physical environment object images spatially adjacent to crack, scratch, wear, burn mark and water mark type defects are extracted, an image sample set containing suspected cause sources is constructed, and corresponding cause source category labels are derived according to the defect types to form cause source image training sample pairs; The cause source image training sample pairs are input into the residual neural network structure to sequentially perform convolution extraction, residual connection and feature fusion operations, and generate high-dimensional embedding vectors containing structural differences and appearance features; Based on the high-dimensional embedding vectors, a classification branch and a regression branch are connected respectively, the classification branch is used to output a type label of the cause source type, and the regression branch is used to output a bounding box position parameter of the target area, and a residual neural network loss function is constructed with cross-entropy loss and bounding box regression loss for back propagation optimization to complete residual neural network parameter learning.

8. The cable tunnel inspection method based on deep learning according to claim 7, characterized in that, The residual neural network loss function is as follows: wherein L is a loss value; for predicting the cause source type label cross-entropy loss between the predicted cause source type label and the true cause source type label y; for predicting the bounding box parameter smooth L1 loss between the predicted bounding box parameter and the true bounding box parameter b; λ1 and λ2 are loss weight coefficients. 9.The cable tunnel inspection method based on deep learning of claim 1, wherein, The establishment of the spatial causal relationship between the defect area and the cause source includes: Based on the three-dimensional spatial positions of the defect area and the cause source, the spatial distance, direction vector and occlusion factor are calculated, the spatial causal correlation degree index is established, and the structured inspection output result is constructed combined with the cause label and confidence score, and the spatial causal relationship information including the defect position, cause category, cause source type and three-dimensional coordinates is output. 10.A cable tunnel inspection system based on deep learning, applied to the cable tunnel inspection method based on deep learning in any one of claims 1-9, characterized in that, Comprise: A defect detection module for acquiring cable representation inspection images by an inspection robot, classifying and regressing the cable representation inspection images based on a convolutional neural network, and outputting defect recognition results; A feature fusion module for acquiring humidity values, temperature values and gas concentration parameters of the defect occurrence position by the inspection robot according to the defect recognition results, and splicing and constructing a fusion feature vector with the image feature vector; A defect cause prediction module for performing cause judgment processing on the fusion feature vector based on a defect cause reasoning model embedded with vector space matching and prior rules, and outputting a predicted cause category label and a corresponding confidence score; A defect cause finding module for determining a defect cause source set according to the predicted cause category label and the corresponding confidence score, and controlling the inspection robot to adjust the image acquisition direction according to the defect area position, acquire image data corresponding to the defect cause source set, and perform image analysis to output the bounding box and type label of the suspected cause source; An inspection result generation module for mapping the bounding box of the suspected cause source to a three-dimensional coordinate point based on the current pose information of the inspection robot and the three-dimensional map data of the cable tunnel, establishing the spatial causal relationship between the defect area and the cause source, and outputting the inspection result including the defect position, cause label and cause source.

Images