Cable trench inspection control method and system based on convolutional neural network

Abstract

The application discloses a cable trench inspection control method and system based on a convolutional neural network, relates to the field of inspection control, and comprises the following steps: firstly, marking a trench on a map and constructing a feasible region, and fusing multi-sensor gridding data; adopting an improved convolutional neural network and a pyramid feature to obtain a global risk map; after skeleton branching, integrating along a path and detecting a peak value to generate seed points; merging adjacent seeds, growing based on the correlation of agents to form sections, and evaluating the risk of the sections; proposing an efficiency model, taking distance attenuation weighted risk / information gain as benefit, and taking passing and turning back as cost, calculating a gain ratio to generate a path in real time; and back-feeding an inspection result to adaptively adjust layer attention based on 1-similarity, thereby improving accuracy and efficiency.

Description

Technical Field

[0001] This invention relates to the field of inspection and control technology, and in particular to a cable trench inspection and control method and system based on convolutional neural networks. Background Technology

[0002] Cable trenches in power plants are critical infrastructure for power system operation. Their narrow, winding interiors, often containing water accumulation, obstacles, and multi-branch structures, pose significant challenges to inspection robots. Existing inspection methods rely primarily on manual labor, which is inefficient and heavily dependent on environmental conditions. While some power plants have introduced mobile robots, their motion control is largely based on preset paths or single shortest path planning, lacking the ability to identify and dynamically avoid risky areas. This results in robots experiencing insufficient inspection coverage, repetitive movements, or motion stiffness due to obstacles in the environment.

[0003] At the control level, existing robots mostly employ path control based on static maps or obstacle avoidance algorithms based on simple sensors, lacking guidance from global risk characteristics. Action decisions are often limited to local environments, making it difficult to reflect a "high-risk priority" control logic. Furthermore, when performing inspection actions, the generation of action commands and environmental risk assessment are independent, lacking a feedback mechanism and failing to achieve real-time adaptation of paths and actions. Especially in power plant cable trenches, risk points are often uncertain and dynamic, making it difficult for existing control methods to adjust action strategies according to changes in risk probability, thus impacting both inspection efficiency and reliability. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a cable trench inspection control method based on convolutional neural networks to solve the problem of lack of risk guidance and dynamic optimization in motion control during the inspection process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a cable trench inspection and control method based on a convolutional neural network, which includes acquiring map information of a power plant and marking the cable trench on the map information.

[0008] The drivable areas of the robot in the map information are marked as path constraints during the robot navigation process;

[0009] Inspection paths are generated by using a pre-trained, improved convolutional neural network.

[0010] The improved convolutional neural network includes using a convolutional neural network as the backbone network, supplemented by pyramid feature aggregation, to perform global analysis on the labeled cable trench; and using a seed model to segment and assess the risk of the cable trench to obtain each node to be inspected.

[0011] Based on the nodes to be inspected and the results of the risk assessment as input, the benefit model is used to simulate the path, and the real-time inspection path is generated online with the goal of maximizing the gain ratio.

[0012] Based on the generated inspection path, the inspection robot's inspection actions are controlled to achieve robot navigation; and after completing the inspection of each node, the inspection results are fed back to the backbone network and the feature aggregation of the pyramid to generate the inspection path for the next node online, thereby performing action control during the inspection of the next node.

[0013] As a preferred embodiment of the cable trench inspection and control method based on convolutional neural networks described in this invention, the data collected by sensors in the cable trench are assigned attributes to the cable trench in the map information to obtain an initial map for risk analysis.

[0014] In the improved convolutional neural network, the input for the part used for global analysis is the initial graph;

[0015] Starting from the input image after unified rasterization, downsampling is performed sequentially using 3×3 convolutions with a stride of 2; the results of the first downsampling layer C2, the second downsampling layer C3, the third downsampling layer C4, and the fourth downsampling layer C5 are obtained;

[0016] The high-level feature C5 is upsampled to the low-level size step by step, and then horizontally added to the 1×1 lateral branch of the corresponding low-level feature. After smoothing with Conv3×3, the feature is fused to obtain: the pyramid feature P5 of the fourth layer, the pyramid feature P4 of the third layer, the pyramid feature P3 of the second layer, and the pyramid feature P2 of the first layer.

[0017] At each pyramid scale ℓ∈{2,3,4,5}, a set of lightweight convolution heads with the same structure is attached to obtain the risk probability of each location at different scales.

[0018] As a preferred embodiment of the cable trench inspection and control method based on convolutional neural network described in this invention, the global analysis includes fusing P2, P3, P4, and P5, and then using a fused convolutional recognition head to output the risk probability at each location point in the cable trench, thereby obtaining the final risk probability map.

[0019] Morphological refinement of the channel mask is performed to obtain the centerline skeleton and branching. In the path coordinate of each branch, the global risk probability field is weighted and integrated within the lateral bandwidth to obtain a one-dimensional path risk curve.

[0020] The curve is smoothed and peak significance is detected, and risk peak points are obtained by screening based on minimum path spacing and probability threshold;

[0021] The peak points are mapped back to two-dimensional coordinates and used as seed points for the seed model.

[0022] As a preferred embodiment of the cable trench inspection and control method based on convolutional neural networks described in this invention, the seed model includes merging adjacent seed points into a composite seed if multiple seed points are adjacent.

[0023] Growth is achieved at each seed location through an agent; for any seed i, its seed point is used as an attribute feature of seed i; the agent is used to analyze the correlation between each position point and the seed as the seed grows to both sides; and growth stops if any one of the following stopping conditions is met:

[0024] Condition 1: The correlation with the seed point or the seed-grown region is less than the corresponding preset value;

[0025] Condition 2: Let there be an adjacent seed j. For position point u, the correlation of seed j is greater than that of seed i.

[0026] The seed model is divided into segments according to its growth range, with each segment serving as a node to be inspected.

[0027] The risk assessment includes attenuating or enhancing the risk of each generated interval based on the pyramid features at different scales: based on the risk probability distribution of the interval under the four-layer pyramid features, the relationship between each pair of risk probability distributions is analyzed to obtain 6 sets of feature values ​​for the relationship between the increase and decrease. Based on the pre-trained mapping function, an adjustment ratio for the interval risk is generated so that the interval risk that tends to be prominent in the global risk is attenuated, and the interval risk that tends to be prominent in the local risk is enhanced.

[0028] As a preferred embodiment of the cable trench inspection and control method based on convolutional neural networks described in this invention, the benefit model includes: performing a risk assessment on each node to be inspected, constructing a risk distribution based on the assessment results, and calculating the gain ratio of each path simulation in the risk distribution;

[0029] The risk assessment includes normalizing the weights of each node in the nodes to be inspected, and using the normalized weights, combined with the risk probability distribution within the interval after the increase or decrease, performing a weighted summation of the risk probabilities to obtain the assessment result.

[0030] In this process, the seed point in the interval is taken as the origin, and the weights are decayed to both sides by distance. The decayed weights are multiplied by the correlation and used as the weights for normalization.

[0031] The risk distribution includes injecting the risk assessment results of each node into the feasible domain of the map information to obtain the distribution results of the risks to be inspected in the feasible domain.

[0032] As a preferred embodiment of the cable trench inspection and control method based on convolutional neural networks described in this invention, the gain ratio considers three factors: Factor 1: the inspection direction is towards the node family with higher risk value; Factor 2: the inspection direction is towards the node family with a large total risk value; Factor 3: the information gain of the next inspection node is large.

[0033] The risk value is expressed as: the result of the risk assessment divided by the interval length;

[0034] The node family refers to all nodes that the robot can reach while moving forward in the current direction without needing to return.

[0035] The gain ratio: ;

[0036] in, , , These represent the weight coefficients of the three elements respectively; The simplified result of element 1 is calculated as the mean of the risk value in the node family; The simplified result of element 2 is calculated as the total risk value under the current inspection direction divided by the total global risk cost. The simplified result of element 3 is obtained by subtracting the expected posterior entropy from the prior entropy.

[0037] As a preferred embodiment of the cable trench inspection and control method based on convolutional neural networks described in this invention, the step of feeding back the inspection results to the backbone network and the feature aggregation of the pyramid specifically includes:

[0038] After writing back the inspection results, the pyramid is reconstructed to obtain the risk probability output at each scale before and after the update.

[0039] Compare the probability distribution outputs of different layers of the pyramid before and after the update, calculate the similarity at the corresponding scale, and use 1 - similarity as the dissimilarity.

[0040] The attention coefficients of each layer of the pyramid are updated with momentum based on dissimilarity and normalized by Softmax so that the scale with greater dissimilarity receives higher attention; and the features of each scale are weighted and fused accordingly for subsequent online path generation.

[0041] Secondly, the present invention provides a cable trench inspection and control system based on a convolutional neural network, including a data acquisition unit for acquiring map information of a power plant and marking the cable trench on the map information.

[0042] The constraint unit marks the drivable area of ​​the robot in the map information as a path constraint during the robot navigation process;

[0043] The path generation unit generates inspection paths through a pre-trained improved convolutional neural network. The improved convolutional neural network includes a convolutional neural network as the backbone network, supplemented by pyramid feature aggregation, to perform global analysis on the labeled cable trenches; segmenting and risk assessment of the cable trenches using a seed model to obtain each node to be inspected; and using the nodes to be inspected and the results of the global analysis as input, performing path simulation using a benefit model, and generating real-time inspection paths online with the goal of maximizing the gain ratio.

[0044] The control unit controls the inspection robot's inspection actions according to the generated inspection path to achieve robot navigation; and after completing the inspection of each node, it feeds back the inspection results to the backbone network and the feature aggregation of the pyramid to generate the inspection path for the next node online, thereby performing action control during the inspection of the next node.

[0045] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the cable trench inspection and control method based on a convolutional neural network as described in the first aspect of the present invention.

[0046] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the cable trench inspection and control method based on a convolutional neural network as described in the first aspect of the present invention.

[0047] The beneficial effects of this invention are as follows: By introducing an improved convolutional neural network to extract risk features and combining it with a benefit model to generate a control potential function, risk-oriented and dynamic optimization of motion control during inspection is achieved. Unlike existing methods that rely on fixed paths or single obstacle avoidance, this invention considers risk value, total risk, and information gain simultaneously in path generation and motion execution, enabling the robot to have a "high-risk priority" decision-making ability in motion control. By feeding the inspection results back to the network structure in real time and performing weighted fusion of multi-scale features, this invention enables the motion control process to form a closed-loop iteration of perception-evaluation-control-feedback, thereby ensuring that the robot's motion commands can be adaptively adjusted according to the risk distribution in complex cable trench environments. This method not only improves the motion flexibility and control accuracy of the inspection robot and reduces ineffective motion and motion rigidity, but also significantly improves inspection efficiency and coverage, ensuring the safety and intelligence level of power plant cable trench inspection. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0049] Figure 1 This is a flowchart of a cable trench inspection and control method based on convolutional neural networks. Detailed Implementation

[0050] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0051] Example 1, referring to Figure 1 As one embodiment of the present invention, a cable trench inspection and control method based on a convolutional neural network is provided, comprising:

[0052] S1: Obtain map information of the power plant and mark the cable trench on the map information.

[0053] The base map can be derived from BIM / CAD / GIS data, supplemented by one-time laser point cloud or multi-view image mapping to correct scale errors. The map coordinate system is calibrated with the robot's body coordinates and the coordinates of each sensor to achieve multi-source data alignment. Cable trenches are then labeled on the aligned base map, including key nodes such as trench boundary polygons, centerline direction, and junctions / manhole covers, as well as attributes such as net width and height limits estimated from point cloud cross-sections. Ambiguous areas are manually verified and supported by imagery. After labeling, the scene is rasterized at a uniform resolution to generate an initial bird's-eye view (BEV) layer for the algorithm, including trench masks, static obstacles, and restricted / no-entry zones, providing a consistent data foundation for subsequent risk assessment and planning.

[0054] S2: Mark the drivable area of ​​the robot in the map information as a path constraint during the robot navigation process.

[0055] The robot's shape and safety clearance are combined into an effective expansion radius. Morphological expansion (Minkowski expansion) is applied to the static obstacle mask, and the expansion result and restricted areas are subtracted within the channel mask to obtain the first version of the feasible region. Further physical and environmental constraints, such as ground slope, step height, water depth, and gas / temperature thresholds, are then superimposed to eliminate grids that do not meet safety conditions. Constraints are applied to the channel centerline curvature and turning radius to shrink the feasible region at narrow bends to prevent unfeasible maneuvers. Connectivity cleanup is then performed, retaining only the main regions connected to the current robot position, and Euclidean distance transformation is calculated to obtain the clearance distance and cost gradient. Dynamic elements such as access control, temporary barriers, or piles of objects are updated and superimposed onto the feasible region at runtime according to timestamps. Finally, the feasible region mask and the corresponding cost map are output, serving as hard constraints for path search and soft constraints for cost evaluation, respectively, providing reliable boundary conditions for real-time inspection path generation.

[0056] S3: The inspection path is generated through a pre-trained improved convolutional neural network.

[0057] The improved convolutional neural network includes using a convolutional neural network as the backbone network, supplemented by pyramid feature aggregation, to perform global analysis on the labeled cable trench; and using a seed model to segment and assess the risk of the cable trench to obtain each node to be inspected.

[0058] Using the nodes to be inspected and the results of the risk assessment as input, a benefit model is used to simulate the path, and the real-time inspection path is generated online with the goal of maximizing the gain ratio.

[0059] It should be noted that the data collected by sensors in the cable trench is used to assign attributes to the cable trench in the map information to obtain an initial map for risk analysis.

[0060] In the improved convolutional neural network, the input for the global analysis part is the initial image. Starting from the uniformly rasterized input image, downsampling is performed sequentially using 3×3 convolutions with a stride of 2; the results of the first downsampling layer C2, the second downsampling layer C3, the third downsampling layer C4, and the fourth downsampling layer C5 are obtained.

[0061] The specific process is as follows: Starting from the uniformly rasterized input image (e.g., 256×256), downsampling is performed sequentially using 3×3 convolutions with a stride of 2. Each downsampling halves the resolution, doubles the receptive field, and typically increases the number of channels to accommodate more abstract semantics.

[0062] Specifically, C2: Conv3×3, stride=2→128×128.

[0063] The first downsampling process retains more edges and details, and begins to possess local semantics.

[0064] C3: Conv3×3, stride=2→64×64.

[0065] The second downsampling reduces local details and makes the global structure clearer.

[0066] C4: Conv3×3, stride=2→32×32.

[0067] The third downsampling significantly enhances the channel semantics, making it suitable for representing segment-level / region-level information.

[0068] C5: Conv3×3, stride=2→16×16.

[0069] The deepest layer has the largest field of vision and best represents the overall risk and context of the entire ditch.

[0070] The high-level feature C5 is upsampled to the low-level size step by step, and then horizontally added to the 1×1 lateral branch of the corresponding low-level feature. After smoothing with Conv3×3, the feature is fused to obtain: the pyramid feature P5 of the fourth layer, the pyramid feature P4 of the third layer, the pyramid feature P3 of the second layer, and the pyramid feature P2 of the first layer.

[0071] The specific process is as follows: the high-level (small resolution, strong semantics) features are upsampled to the low-level size step by step, and then horizontally fused with the corresponding low-level features, which brings back the global semantics and fills in the local details.

[0072] P5 = Conv1×1(C5) is first normalized by a 1×1 convolution to obtain the highest-level pyramid feature P5 (still 16×16). P4 = Conv1×1(C4) + Up(P5) is formed by upsampling P5 by bilinear interpolation by 2× (to avoid edge misalignment), matching the size of C4 (32×32), and then adding it to the 1×1 lateral branch of C4. Here, Up represents the upsampling process.

[0073] Then, a Conv3×3 smoothing layer is applied to P4 to suppress aliasing and jagged edges caused by upsampling. P3 = Conv1×1(C3) + Up(P4) → Similarly, upsample by 2× to 64×64, add it horizontally to C3, and then smooth with Conv3×3.

[0074] P2 = Conv1×1(C2) + Up(P3) → Similarly, upsample by 2× to 128×128, add horizontally to C2, and then smooth with Conv3×3. The fusion method can be selected as: addition (lightweight and stable) or concatenation followed by convolution (more information, slightly heavier). Choosing bilinear upsampling is a stable approach in traditional FPN, and coordinate alignment is more reliable.

[0075] At each pyramid scale ℓ∈{2, 3, 4, 5}, a set of lightweight convolutional heads with the same structure is attached to directly obtain the risk probability at each location at that scale:

[0076] When conducting global risk analysis on cable trenches, the different characteristics of pyramid features P2, P3, P4, and P5 are processed. Among them, high-resolution features P2 and P3 are more suitable for identifying fine-grained risk information, such as small cracks on the edge of cable sheaths or manhole covers; while low-resolution but semantically stronger features P4 and P5 are more suitable for expressing large-scale or continuous risks, such as the risk of water accumulation in narrow strips or long sections.

[0077] To leverage the advantages of features at different levels, a lightweight convolutional head (RiskHead) is first set up at each scale to perform risk prediction on local information. The convolutional head structure is as follows: Convolution fuses local features, and then... Convolution generates logits; for binary classification tasks, a sigmoid function is added to the output to obtain the risk probabilities of the two classes; for multi-class classification tasks, a softmax function is added to the output to obtain the risk probabilities of each class. Each layer can produce a risk probability map at native resolution; for example, the output of P5 is... The output of P2 is When it is necessary to display the risk probability map in alignment with the original input map, or when performing cross-layer feature fusion, the risk probability maps of each layer are then upsampled to the same spatial resolution as the input map.

[0078] After extracting features at each of the above layers, the feature maps at each scale are upsampled to the same spatial resolution as the input map, resulting in aligned multi-scale feature maps. Subsequently, the aforementioned feature maps are concatenated by channel and input into the fusion convolutional recognition head for unified processing. The fusion convolutional recognition head employs a lightweight convolutional structure, firstly through... Convolution compresses the number of channels after concatenation to 128 dimensions, and performs a non-linear transformation using batch normalization (BN) and the ReLU activation function; then... Convolution reduces the number of channels to 64, and it is also processed by BN and ReLU; finally, it is passed through... Convolution yields a K-dimensional output, and a Sigmoid or Softmax function is then applied to the output to generate the final risk probability map. .

[0079] During training, the binary classification task employs a binary cross-entropy loss with Logits (BCEWithLogits) combined with the Dice loss function for optimization; the multi-class classification task uses cross-entropy loss (CE) combined with either Dice or Lovász loss for optimization to enhance robustness to imbalanced samples. Through this fused convolutional recognition head, pyramid features at different scales can work synergistically, and the final output risk probability map accurately depicts the risk distribution at various locations within the cable trench, providing precise input for subsequent seed models and path planning.

[0080] After calculating the risk probability map, the masked area of ​​the cable trench is first subjected to morphological processing, and the centerline skeleton of the trench is extracted through a thinning operation. The centerline skeleton uses pixels or discrete points as basic units and forms an undirected graph in terms of topology. To facilitate subsequent processing, the undirected graph is split into several continuous branches according to the connection relationship, and each branch corresponds to a path in the trench skeleton.

[0081] After obtaining each branch, using the along-path coordinates as reference axes, a weighted integral is performed on the global risk probability field within the lateral bandwidth of each branch. Through integration, the risk distribution in two-dimensional space is projected onto a one-dimensional along-path coordinate system, thus obtaining the one-dimensional along-path risk curve for the corresponding branch. This curve can characterize the risk change trend of the channel in the direction of that branch, providing basic data for subsequent risk saliency detection and seed point selection.

[0082] Specifically, the process uses the coordinate system along the path as the principal axis, collects risk probability values ​​within the horizontal bandwidth, and assigns weights based on horizontal distance or correlation to sum the risk values ​​of each pixel or discrete point. In this embodiment, the weighted integration method can be implemented in two ways: one is based on a geometric distance-based attenuation rule, where the greater the horizontal distance, the smaller the weight assigned, to highlight the risk contribution near the principal axis along the path; the other is based on a risk correlation-based weighting method, where if the risk performance of a point along the principal axis is strongly correlated in the horizontal direction, a higher weight is assigned to reflect the spatial consistency of the risk pattern.

[0083] The weights can be set using either a fixed decay function, such as decreasing in segments according to distance, or an adaptive approach, learning a suitable weight distribution based on historical risk distribution or training data. In practical applications, the weights are typically highest near the main axis and gradually decrease with increasing lateral distance, ensuring that the risk curve along the channel more accurately reflects risk fluctuations in the main direction.

[0084] This weighted integration method can avoid the interference of lateral noise on the risk curve along the route, and ensure that local significant risk areas are highlighted, thus providing a more reliable basis for subsequent risk significance detection and seed point selection.

[0085] Based on the obtained one-dimensional risk curve along the path, further smoothing and peak significance detection can be performed to identify key points with prominent risks. These key points serve as candidate inputs for subsequent seed models, facilitating the segmented representation and node-based processing of branch risks. In this way, the two-dimensional risk probability field is effectively transformed into structured risk information along the path, providing a quantifiable input basis for subsequent risk significance detection, seed point selection, and path optimization.

[0086] The curve is smoothed and peak significance is detected. Risk peak points are then selected based on minimum distance along the path and a probability threshold. These peak points are mapped back to two-dimensional coordinates and used as seed points for the seed model.

[0087] In the process of generating cable trench inspection paths, the seed model not only relies on spatial proximity for interval division but also introduces a path correlation analyzer to improve the rationality of segmentation. The specific implementation of the seed model includes:

[0088] First, when multiple seed points are spatially adjacent, they are merged into a composite seed, which serves as the starting point for growth. For any seed i, its initial seed point is considered an attribute feature of seed i. During growth, the seed model uses an agent to perform correlation analysis on the expansion of the seed along both sides, thereby determining whether a new location point belongs to the seed's growth range.

[0089] The agent is modeled using a path correlation analyzer. Specifically, the channel branch skeleton is transformed into a one-dimensional sequence according to the path coordinates s, and the path risk curve and its related features are extracted from the global risk probability map to form the input feature vector:

[0090] ,

[0091] in, Represents the coordinates along the route. Indicates the location along the route The input feature vector at that location. This indicates the risk value along the route. It represents the first derivative of the risk value along the route and is used to characterize the rate of change of risk. Indicates the location of the ditch The net width of the place. This represents the rate of change of net width with respect to coordinates along the path. Indicates position Tangential affinity at the point. Indicates position The affinity of the place. This indicates that the input feature vector belongs to a dimension of The real number space. This indicates that the dimension of the input feature vector is approximately 6.

[0092] In feature modeling, the along-range correlation analyzer uses a Temporal Convolutional Network (TCN) as its backbone, containing multiple dilated convolutional residual blocks to obtain long-range dependencies without downsampling. Sequence embedding representations are obtained through this network. Subsequently, within a set set of lag steps... Next, calculate the correlation along the route:

[0093] ,

[0094] in, This indicates that after processing by the TCN network, at the location The sequence embedding representation obtained at that point. This indicates that the dimension of the embedding vector is 32. This represents the lag step size, with a value range of [value missing]. . This indicates the maximum lag length. Defined. Indicates position The embedding vector at that location. Representing vectors and The inner product of. Representing vectors The Euclidean norm. Representing vectors The Euclidean norm. Indicates position Lag The normalized correlation value.

[0095] And further define the correlation index along the route:

[0096] ,

[0097] ,

[0098] in, Indicates position The correlation index along the route. Indicates the lag step size The corresponding weights. Indicates weight It is proportional to the exponential decay function. This represents the attenuation coefficient.

[0099] When Corr(s) remains high, it indicates consistency in risk or geometric pattern between the preceding and following positions, and the agent determines that growth can continue in that direction. When Corr(s) decreases, or the probability of a change point increases significantly, the agent determines that the current position is a candidate point to stop. Therefore, the stopping condition of the seed model is as follows: Condition 1: If the correlation between the new position point and the seed point or the already grown interval is lower than a preset threshold, then growth stops.

[0100] Condition 2: If there are adjacent seeds And for a certain location point ,seed The correlation value is higher than that of the seed. Then the seed Growth is halted at that location. Through this mechanism, the seed model, under the dual constraints of spatial proximity and path correlation, can achieve reasonable division of branch intervals and effective identification of risk nodes, providing accurate input for subsequent risk assessment and path optimization.

[0101] The formula for the probability of a change point is:

[0102] ,

[0103] in, Indicates position The probability of a change point at a given location. This represents the Sigmoid activation function. Represents the embedding vector The output after performing a one-dimensional convolution.

[0104] Based on the growth range of the seed model, the channel framework is divided into several segments, each segment serving as a node to be inspected. To improve the accuracy of segment risk assessment, a multi-scale risk correction mechanism is introduced.

[0105] The multi-scale risk correction mechanism includes: when performing multi-scale analysis on the generated interval risk, first calculating the interval at each of the four pyramid scales. The mean of the probability distribution of risk under the given conditions is denoted as:

[0106] ,

[0107] in, Representing scale The mean risk below, This indicates the number of pixels contained in the interval at this scale. Represents pixels The risk probability value.

[0108] Subsequently, the percentage increase or decrease in the mean between adjacent scales was calculated for each pair:

[0109] ,

[0110] in, Representing scale With scale The relationship between the mean increase and decrease is expressed as a percentage. Representing scale The average risk; Representing scale The mean risk; scale With scale For any two different scales.

[0111] For four scales The following 6 sets of eigenvalues ​​can be obtained:

[0112] ,

[0113] They represent and between, and between, and between, and between, and between, and The eigenvalues ​​between.

[0114] These eigenvalues ​​comprehensively describe the relative changes in the mean risk across different scales: a positive percentage result indicates an increase in the mean risk at a higher level scale; a negative percentage result indicates a decrease in the mean risk at a higher level scale.

[0115] The eigenvalues, as a quantitative description of multi-scale risk differences, can reflect the consistency and differences of interval risk at local and global scales.

[0116] In this embodiment, the aforementioned six sets of percentage features are used as inputs and fed into a pre-trained Bayesian network to generate the adjustment ratio of interval risk (which can be positive or negative). This approach not only quantifies the differences in risk across different scales but also makes the subsequent risk correction process calculable and interpretable.

[0117] The adjustment ratio can be positive or negative: when an interval exhibits globally prominent risk characteristics across multiple scales, the Bayesian network generates a negative adjustment ratio to attenuate the risk value. This approach avoids overemphasizing high-risk areas already prominent at the global scale, thus maintaining overall balance in path planning and preventing excessive resource concentration.

[0118] When an interval exhibits significant risk at a local scale but not at a global scale, the Bayesian network generates a positive adjustment ratio to amplify the risk value. This approach prevents local high-risk points from being masked by the global averaging effect, thus ensuring that locally prominent risks receive priority in subsequent node selection and path planning.

[0119] Therefore, the adjustment ratio can be positive or negative based on a dynamic correction mechanism that addresses the differences between global and local risk performance: globally prominent risks need to be suppressed, while locally prominent risks need to be amplified. In this way, risk assessment achieves a dynamic balance across different scales, ensuring that the selection of inspection nodes not only covers the overall risk distribution but also accurately captures significant local risk points, thereby improving the effectiveness and reliability of the final inspection path.

[0120] It should also be noted that the benefit model includes performing a risk assessment on each node to be inspected, constructing a risk distribution based on the assessment results, and calculating the gain ratio of each path simulation in the risk distribution.

[0121] The risk assessment includes normalizing the weights of each node among the nodes to be inspected, and then using the normalized weights, combined with the risk probability distribution within the interval after the increase or decrease, to perform a weighted summation of the risk probabilities to obtain the assessment result. Specifically, the seed point in the interval is used as the origin, and the weights are attenuated by distance to both sides. The attenuated weights are multiplied by the correlation, and the result is used as the weight for normalization.

[0122] The risk distribution includes injecting the risk assessment results of each node into the feasible domain of the map information to obtain the distribution results of the risks to be inspected in the feasible domain.

[0123] The gain ratio considers three factors: Factor 1: The inspection direction is towards a node family with higher risk value; Factor 2: The inspection direction is towards a node family with a large total risk value; Factor 3: The next inspected node has a large information gain. The risk value is expressed as: the result of the risk assessment divided by the interval length. The node family is: all nodes that the robot can reach while moving forward in the current direction without needing to return. The gain ratio: .

[0124] in, , , These represent the weight coefficients of the three elements respectively; The simplified result of element 1 is calculated as the mean of the risk value in the node family; The simplified result of element 2 is calculated as the total risk value under the current inspection direction divided by the total global risk cost. The simplified result of element 3 is obtained by subtracting the expected posterior entropy from the prior entropy.

[0125] Based on the generated inspection path, the inspection robot's inspection actions are controlled to achieve robot navigation; and after completing the inspection of each node, the inspection results are fed back to the backbone network and the feature aggregation of the pyramid to generate the inspection path for the next node online, thereby performing action control during the inspection of the next node.

[0126] Furthermore, the step of feeding back the inspection results to the backbone network and the feature aggregation of the pyramid specifically includes:

[0127] After writing back the inspection results, the pyramid is reconstructed to obtain the risk probability output at each scale before and after the update.

[0128] Compare the probability distribution outputs of different layers of the pyramid before and after the update, calculate the similarity at the corresponding scale, and use 1 - similarity as the dissimilarity.

[0129] The attention coefficients of each layer of the pyramid are updated with momentum based on dissimilarity and normalized by Softmax so that the scale with greater dissimilarity receives higher attention; and the features of each scale are weighted and fused accordingly for subsequent online path generation.

[0130] Specifically, the risk probability maps of each layer before and after the update are first aligned, and their distribution differences are compared within the feasible region. A similarity index is calculated to measure the consistency between the two outputs, and the inverse of the similarity is used as the dissimilarity to characterize the magnitude of change in that scale before and after the write-back. If the dissimilarity of a certain layer is large, it indicates that the layer has undergone significant changes under the influence of new observations, representing that the layer contains new or prominent risk information.

[0131] After the dissimilarity calculation is completed, a momentum update mechanism is introduced to smoothly adjust the attention coefficients of each layer. This ensures that attention changes can respond promptly to new information while avoiding drastic fluctuations caused by single abnormal inspections. Subsequently, the updated attention coefficients are normalized to maintain comparability across all scales. Ultimately, layers with higher dissimilarity receive higher attention weights, while layers with lower dissimilarity receive relatively lower weights.

[0132] Based on the updated attention weights, feature maps at different scales are weighted and fused. The fused result not only enhances the scale features driven by new observations but also ensures a dynamic balance between global and local risk information. This fused feature serves as input for subsequent online path generation, enabling the robot to adaptively optimize path planning based on the latest feedback after each inspection. This achieves a closed-loop control logic of "the greater the change, the higher the attention," ensuring the entire method is feasible for practical engineering applications.

[0133] On the other hand, this embodiment also provides a cable trench inspection and control system based on a convolutional neural network, which includes:

[0134] The data acquisition unit obtains map information of the power plant and marks the cable trenches on the map information.

[0135] The constraint unit marks the drivable area of ​​the robot in the map information as a path constraint during the robot navigation process.

[0136] The path generation unit generates inspection paths through a pre-trained improved convolutional neural network. The improved convolutional neural network includes a convolutional neural network as the backbone network, supplemented by pyramid feature aggregation, to perform global analysis on the labeled cable trenches. The cable trenches are segmented and risk assessed using a seed model to obtain each node to be inspected. Based on the nodes to be inspected and the results of the global analysis as input, a benefit model is used to simulate the path, and the online generation of real-time inspection paths is performed with the goal of maximizing the gain ratio.

[0137] The control unit controls the inspection robot's inspection actions according to the generated inspection path; realizes robot navigation, and after completing the inspection of each node, feeds back the inspection results to the backbone network and the feature aggregation of the pyramid, generates the inspection path for the next node online, and thus performs action control during the inspection of the next node.

[0138] This embodiment also provides a computer device suitable for the cable trench inspection and control method based on convolutional neural networks, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the cable trench inspection and control method based on convolutional neural networks as proposed in the above embodiment.

[0139] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0140] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the cable trench inspection control method based on a convolutional neural network as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A cable trench inspection control method based on a convolutional neural network, characterized in that: The method comprises: acquiring map information of a power plant, and marking cable trenches on the map information; Identifying a travelable area of the robot in the map information as a path constraint in a robot navigation process; Generating a patrol path by using a pre-trained improved convolutional neural network; The improved convolutional neural network comprises: taking a convolutional neural network as a backbone network, and supplementing a pyramid feature aggregation to perform global analysis on the marked cable trenches; Segmenting and risk evaluating the cable trenches by using a seed model to obtain each node to be patrolled; Using a benefit model to perform path simulation by taking the node to be patrolled and the risk evaluation result as input, and generating an online patrol path in real time by taking the maximum gain ratio as a target; Controlling a patrol robot according to the generated patrol path to perform a patrol action, and realizing robot navigation; After completing the patrol of each node, feeding back a patrol result to the backbone network and the pyramid feature aggregation to generate an online patrol path for the next node, and thus controlling the action in the next node patrol. 2.The cable trench inspection control method based on a convolutional neural network according to claim 1, wherein: Assigning attributes to data collected by sensors in the cable trenches in the cable trenches in the map information to obtain an initial graph for risk analysis; Taking the initial graph as input in the improved convolutional neural network; Starting from the unified rasterized input graph, sequentially performing down-sampling by using a 3*3 convolution with a stride of 2 to obtain a first layer down-sampling result C2, a second layer down-sampling result C3, a third layer down-sampling result C4, and a fourth layer down-sampling result C5; Up-sampling the high-layer feature C5 to low-layer sizes step by step, and performing horizontal addition on each layer up-sampling result and a 1*1 lateral branch of the corresponding layer down-sampling result, and then performing smoothing by using Conv3*3 to realize fusion, and obtaining a fourth layer pyramid feature P5, a third layer pyramid feature P4, a second layer pyramid feature P3, and a first layer pyramid feature P2; At each pyramid scale On each pyramid scale ∈ {2, 3, 4, 5}, a set of lightweight convolution heads with the same structure are hung to obtain the risk probability of each position at different scales. 3.The cable trench inspection control method based on a convolutional neural network according to claim 2, wherein: The global analysis comprises: after fusing P2, P3, P4, and P5, using a fusion convolution recognition head to output a risk probability of each position point in the cable trenches to obtain a final risk probability graph; Performing morphological thinning on a channel mask to obtain a center line skeleton and branch, performing weighted integration on a global risk probability field within a horizontal bandwidth at the coordinates of each branch to obtain a one-dimensional along-path risk curve; Performing smoothing and peak significance detection on the curve, and screening a risk peak point based on a minimum along-path distance and a probability threshold; Mapping the peak point back to a two-dimensional coordinate as a seed point of the seed model. 4.The cable trench inspection control method based on a convolutional neural network according to claim 3, wherein: The seed model comprises: if there are adjacent different seed points, merging the adjacent seed points into a composite seed; Each seed grows by using an agent; for any seed i, taking the corresponding seed point as an attribute feature of the seed i; using the agent to analyze the relevance of each position point to the seed when the seed grows to both sides; and stopping growing when any stop condition is met; Condition one: the relevance to the seed point or the seed grown interval is less than a corresponding preset value. Condition two: set adjacent seed j, for position point u, the correlation of seed j is greater than the correlation of seed i; Segmenting according to the growth interval of the seed model, each segment as a node to be inspected; The risk assessment includes, according to the pyramid feature under different scales, the attenuation or enhancement of the interval risk of each generated interval: according to the risk probability distribution of the interval under four layer pyramid features, the increase and decrease relationship between each two risk probability distributions is analyzed, and the characteristic value of 6 groups of increase and decrease relationship is obtained, according to the pre trained mapping function, the adjustment ratio of interval risk is generated, so that the interval risk which tends to be outstanding in global risk performance is attenuated; The interval risk which tends to be outstanding in local is enhanced. 5.The cable trench inspection control method based on a convolutional neural network according to claim 4, wherein: The benefit model includes, for each of the nodes to be inspected, the risk assessment of the node, and the risk distribution is constructed according to the evaluation result, and the gain ratio of each path simulation is calculated in the risk distribution; The risk assessment includes, in the nodes to be inspected, the weight normalization of each node, and the risk probability distribution in the interval is combined with the change of increase and decrease relationship after the weight normalization, and the weighted sum of risk probability is obtained to get the evaluation result; Wherein, taking the seed point in the interval as the origin, the weight attenuation is carried out on both sides of the seed growth with distance as the measure, and the attenuated weight is multiplied by the correlation to obtain the weight for normalization, and the weight normalization is carried out; The risk distribution includes, on the feasible region of the map information, injecting the risk assessment result of each node to obtain the distribution result of the risk to be inspected in the feasible region. 6.The cable trench inspection control method based on a convolutional neural network according to claim 5, wherein: The gain ratio considers three elements: element 1: the direction of inspection is towards the node group with more risk value; Element 2: the inspection direction is towards the node group with large total risk value; Element 3: the information gain of the next inspection node is large; The risk value is represented as: the result of the risk assessment divided by the interval length; The node group is: all nodes that can be reached by the current robot in the current advancing direction without returning; The gain ratio: ; wherein, , , respectively represent the weight coefficients of the three elements; represents the simplified result of element 1, calculated as the mean of the risk values in the node family; represents the simplified result of element 2, calculated as the total amount of risk values under the current inspection direction divided by the global total risk cost; represents the simplified result of element 3, obtained by priori entropy-expectation posteriori entropy.

7. The cable trench inspection control method based on a convolutional neural network according to claim 6, characterized in that: The feedback of the inspection result to the backbone network and the feature aggregation of the pyramid specifically includes: After rewriting the inspection result, the pyramid is reconstructed, the input is updated, and the risk probability output of each scale after updating is obtained; By using the risk probability output of each scale before and after updating, the probability distribution output results of different layers of the pyramid before and after updating are compared, and the similarity under the corresponding scale is calculated, and the dissimilarity is taken as 1-similarity; Based on the dissimilarity, the attention coefficients of each layer of the pyramid are momentum updated and normalized by Softmax; the normalized result is taken as the weight, and the features of each scale are weighted and fused, which are used for subsequent online path generation.

8. The cable trench inspection control system based on a convolutional neural network, the cable trench inspection control method based on a convolutional neural network according to any one of claims 1-7, characterized in that: It includes a collection unit that acquires map information of a power plant and labels a cable trench on the map information; A constraint unit identifies a feasible region of a robot in the map information as a path constraint in the robot navigation process; The path generation unit realizes the generation of the inspection path through a pre-trained improved convolutional neural network; the improved convolutional neural network comprises a convolutional neural network as a backbone network, and is supplemented by a pyramid feature aggregation to perform global analysis on the cable channel marked; Each node to be inspected is obtained by segmenting and risk evaluating the cable channel through a seed model; the benefit model is used to simulate the path according to the node to be inspected and the result of the global analysis as input, and the online generation of the real-time inspection path is performed with the maximum gain ratio as the target; The control unit controls the inspection action of the inspection robot according to the generated inspection path, and realizes the robot navigation; After the inspection of each node is completed, the inspection result is fed back to the backbone network and the pyramid feature aggregation to perform the online generation of the inspection path for the next node, so that the action control is performed in the inspection of the next node. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the cable channel inspection control method based on the convolutional neural network in any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the cable channel inspection control method based on the convolutional neural network in any one of claims 1-7.

Images