Substation inspection robot obstacle detection method and system and computer equipment

Abstract

The invention discloses a substation inspection robot obstacle detection method and system and computer equipment, an obstacle point cloud target detection model of a substation is constructed, data preprocessing and voxelization processing are performed on original point cloud to form structured three-dimensional data, a three-dimensional feature enhancement module performs preliminary screening of features through a feature extraction block, and a three-dimensional feature extraction module performs feature extraction through a feature extraction block; dynamically selecting foreground voxels by using response value calculation, performing feature diffusion enhanced spatial representation, and realizing multi-scale feature fusion by combining residual connection and down-sampling operation; in the BEV feature enhancement stage, through the dependency relationship between dimension conversion and MLP modeling channels, feature transposition and channel shuffling are performed, features are refined through convolution and nonlinear transformation, and finally enhanced BEV features and initial features are multiplied element by element, so that multi-dimensional information fusion is realized; the detection head module carries out prediction and cooperates with a post-processing algorithm to realize accurate positioning. And the obstacle identification accuracy and real-time performance of the substation inspection robot in a complex scene are obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of obstacle detection and inspection of substation robots, and in particular relates to an obstacle detection method, system and computer equipment of a substation inspection robot. Background Art

[0002] Substations are crucial infrastructure for power systems, responsible for receiving, transforming, and distributing electrical energy. Their stable operation is crucial to the security of the power grid. However, the complex internal environment of substations is characterized by the presence of numerous obstacles. These obstacles not only hinder the efficient execution of inspection tasks but also threaten the safe operation of equipment. Therefore, accurate detection and avoidance of obstacles within substations are key technologies for achieving intelligent inspections.

[0003] At present, the inspection of power systems mainly relies on manual inspections or the use of image recognition technology. Manual inspections are not only time-consuming and labor-intensive, but also pose safety risks in high-risk working environments. Image-based methods are easily affected by environmental factors such as lighting, weather, and occlusion, and their perception of dynamic obstacles and spatial depth information is limited. Target detection technology based on three-dimensional point cloud data has gradually gained attention in the field of power operation and maintenance due to its rich geometric information and high-precision performance. Point cloud target detection technology can obtain three-dimensional point cloud data of high-voltage transmission lines and substations through sensors such as lidar, realize accurate detection and positioning of obstacles, and significantly improve the intelligent level of power facility inspections. However, the specific implementation process faces the following difficulties: 1. Point cloud data in substation scenarios is often sparsely distributed, especially in long-distance areas or complex geometric structures. The point cloud data density is low, making it difficult for target detection algorithms to capture complete target features.

[0004] 2. In the process of converting 3D point cloud features to BEV features, the spatial information of the target may be partially lost, which leads to insufficient utilization of spatial context information by the model, especially in scenes with complex occlusions or overlapping targets, affecting the accuracy and robustness of detection.

[0005] Based on this, the present invention proposes a substation inspection robot obstacle detection method, system and computer equipment. Summary of the Invention

[0006] In response to the above technical problems, the present invention provides an obstacle detection method, system and computer equipment for a substation inspection robot, which aims to solve the technical problem that the target detection algorithm is difficult to capture complete target features due to the sparsity of point clouds, and that some spatial information is lost when three-dimensional point cloud features are converted into BEV features.

[0007] The technical solution adopted by the present invention to solve the technical problem is: A method for detecting obstacles in a substation inspection robot, the method comprising the following steps: S100: Build a substation obstacle point cloud detection platform, collect 3D point cloud data of substation obstacles, perform data annotation, and create a substation obstacle detection dataset; S200: Build a substation obstacle point cloud target detection model, which includes a 3D feature extraction and enhancement module, a BEV feature enhancement module, a detection head prediction module, and a post-processing module; S300: voxelize the point cloud data. The 3D feature extraction and enhancement module is used to extract and enhance 3D features from the voxelized point cloud data. The enhanced 3D features are input to the high compression module for dimensionality reduction processing. S400: Inputting the three-dimensional features after dimensionality reduction processing into the BEV feature enhancement module to convert them into two-dimensional BEV features, and further enhancing the two-dimensional BEV features; S500: Input the enhanced two-dimensional BEV features into the detection head prediction module to identify and locate the object in the three-dimensional space and output a three-dimensional target detection frame; S600: Input the training set of the substation obstacle detection dataset into the substation obstacle point cloud target detection model. Using the three-dimensional target detection frame and the set loss function, backpropagate the substation model to update the model parameters. Repeat the above steps for optimization to finally obtain a trained substation obstacle point cloud target detection model. Based on the trained substation obstacle point cloud target detection model, detect obstacles to obtain a three-dimensional target detection frame. The final detection result is obtained by screening based on the post-processing module.

[0008] Preferably, the three-dimensional feature extraction enhancement module includes a first feature extraction block, a voxel generation block, downsampling, a second feature extraction block and upsampling, and S300 includes: S310: Dividing the voxelized 3D voxel feature into a plurality of local windows along the X-axis direction, each window including a voxel group arranged along the X-axis direction; and re-dividing the window along the Y-axis direction to form a voxel group arranged along the Y-axis direction; S320: In the X-axis window, a linear RNN recurrent neural network is used to process voxel features in X-axis order to model the long-range dependency relationship between distant voxels in the same window; in the Y-axis window, a linear RNN recurrent neural network is used to process voxel features in Y-axis order to model the long-range dependency relationship between distant voxels in the same window; the global context perception ability of the feature is enhanced by the bidirectional linear RNN to obtain the output feature of the first feature extraction block ; S330: In the voxel generation block, according to the output features Calculate the characteristic response value of each voxel ; S340: Arrange the characteristic response values ​​of each voxel in descending order, and select a preset number of voxels from the total number of non-empty voxels as foreground; S350: Diffuse the coordinates corresponding to the selected foreground voxel features along the x-axis, y-axis, and z-axis using four different offsets [-1, -1, 0], [-1, 1, 0], [1, -1, 0], and [1, 1, 0] to obtain diffused voxel features, and then initialize the corresponding features of the diffused voxel features to all 0s, and use the output features of the first feature extraction block as the initialization value. With the initialized voxel features Connect and downsample; S360: The downsampled output is sent to the second feature extraction block, and the voxel features are divided into multiple local windows along the X-axis direction again, and remote feature interaction based on the X-axis window partition is performed. The window is again divided along the Y-axis direction, and remote feature interaction based on the Y-axis window partition is performed. The extracted features are upsampled to restore the original resolution and combined with the features generated by the first feature extraction block. Perform residual connection to obtain 3D enhanced features; S370: Input the enhanced three-dimensional features into the high compression module for dimensionality reduction to obtain the initial BEV features .

[0009] Preferably, S330 specifically includes: ; in, represents the characteristic response value of the t-th feature, C is The channel dimension, i is the index of the channel, t = 1, 2, ..., N; S340 specifically: ; ; Where r is the ratio of foreground, z is the total number of non-empty voxels, Indicates from Select the q voxel features with the largest feature response, is the foreground feature of the selection; S350 is specifically: ; in, express and The connection characteristics of represents the initialized voxel features, which are diffused along the x-axis, y-axis and z-axis by the x, y and z offsets, and G(x) denote the connection block and down-sampling operation respectively.

[0010] Preferably, S400 includes: S410: Initial BEV characteristics ,First, perform dimension permutation to transform the features from C×H×W to W×H×C; S420: The dependencies between channels are captured using a two-layer multilayer perceptron (MLP). The first layer of MLP reduces the number of channels to 1 / 4 of the original number. Nonlinearity is then introduced using the ReLU activation function. The second layer of MLP restores the number of channels to their original dimensions. S430: Perform inverse permutation to restore to C×H×W, and generate features through Sigmoid activation function , the initial BEV characteristics and features Perform element-by-element multiplication to obtain preliminary enhanced features ; S440: The enhanced features are divided into 4 groups, each containing C / 4 channels. The grouped features are transposed to disrupt the channel order in each group. Then, the disrupted feature map is restored to its original shape to enhance the feature expression ability and obtain the feature ; S450: Features After the convolution layer, the number of channels is reduced to features 1 / 4, and then undergoes batch normalization and ReLU activation function for nonlinear transformation, and then passes through the second convolution layer to restore the number of channels to the original dimension C, and then passes through the batch normalization layer; S460: Generate features through Sigmoid activation function , the features and features Multiply element by element to get the final output feature .

[0011] Preferably, S420 and S430 are specifically: ; ; in, represents the Sigmoid activation function, represents element-wise multiplication, Represents a dimension permutation operation.

[0012] Preferably, the loss function in S600 is specifically: ; ; ; in, is the total loss, is the regression loss, is the prediction box classification loss, is a balance parameter used to adjust the weight of regression loss and classification loss in the total loss, N is the number of positive samples, is the true label of the i-th sample, is the category probability of the i-th sample predicted by the model, are the parameters of the i-th bounding box predicted by the model, is the parameter of the i-th ground-truth bounding box.

[0013] Preferably, the screening based on the post-processing module in S600 to obtain the final detection result includes: S610: For each substation obstacle category, sort the detection boxes according to the confidence level of the substation obstacle category, from high to low. S620: Select the box with the highest confidence and add it to the final detection result list; S630: traverse the remaining boxes, filter out the boxes whose intersection over union (IoU) with the selected box is greater than a preset threshold, and discard the remaining boxes; S640: Repeat the above steps until all frames are processed and the detection results are obtained.

[0014] The obstacle detection system for substation inspection robots includes a data set creation module, a detection model building module, a feature extraction and enhancement module, a BEV feature enhancement module, a detection head prediction module, and a post-processing module. The dataset creation module is used to build a substation obstacle point cloud detection platform, collect 3D point cloud data of substation obstacles, perform data annotation, and create a substation obstacle detection dataset; Detection model building module, used to build a substation obstacle point cloud target detection model. The model includes a 3D feature extraction and enhancement module, a BEV feature enhancement module, a detection head prediction module, and a post-processing module. The 3D feature extraction and enhancement module is used to extract and enhance 3D features from voxelized point cloud data, and input the enhanced 3D features into the height compression module for dimensionality reduction and then into the BEV feature enhancement module; The BEV feature enhancement module converts the three-dimensional features after dimensionality reduction into two-dimensional BEV features and further enhances the two-dimensional BEV features; The detection head prediction module recognizes and locates objects in 3D space based on the enhanced 2D BEV features and outputs 3D object detection boxes; The training set of the substation obstacle detection dataset is input into the substation obstacle point cloud target detection model. The substation model is back-propagated through the three-dimensional target detection box and the set loss function to update the model parameters. The above steps are repeated for optimization. Finally, a trained substation obstacle point cloud target detection model is obtained. Based on the trained substation obstacle point cloud target detection model, obstacles are detected to obtain a three-dimensional target detection box, and the final detection result is obtained by screening based on the post-processing module.

[0015] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor implements the steps of a method for detecting obstacles of a substation inspection robot when executing the computer program.

[0016] The method proposed in this invention can better adapt to complex environmental conditions, especially in low-density, long-distance and multi-obstacle overlapping scenarios, and effectively avoid misjudgments caused by factors such as occlusion and lighting changes; at the same time, the use of three-dimensional point cloud data enables more comprehensive utilization of spatial information, greatly improving the ability to perceive obstacles and providing strong technical support for intelligent inspections of substations. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of an obstacle detection method for a substation inspection robot according to one embodiment of the present invention; Figure 2 Schematic diagram of an obstacle network of a substation inspection robot according to one embodiment of the present invention; Figure 3 Schematic diagram of the principle of a three-dimensional feature enhancement and extraction module in one embodiment of the present invention; Figure 4 Schematic diagram of the principle of a BEV feature enhancement module in one embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below with reference to the accompanying drawings.

[0019] In one embodiment, Figure 1 and Figure 2 As shown, a method for detecting obstacles of a substation inspection robot includes the following steps: S100: Build a substation obstacle point cloud detection platform, collect 3D point cloud data of substation obstacles, perform data annotation, and create a substation obstacle detection dataset. Furthermore, based on the point cloud data obtained by LiDAR, the dataset is divided into a 60% training set and a 40% validation set according to a preset ratio.

[0020] S200: Build a substation obstacle point cloud target detection model, which includes a 3D feature extraction and enhancement module, a BEV feature enhancement module, a detection head prediction module, and a post-processing module; S300: voxelize the point cloud data. The 3D feature extraction and enhancement module is used to extract and enhance 3D features from the voxelized point cloud data. The enhanced 3D features are input to the high compression module for dimensionality reduction processing. S400: Inputting the three-dimensional features after dimensionality reduction processing into the BEV feature enhancement module to convert them into two-dimensional BEV features, and further enhancing the two-dimensional BEV features; S500: Input the enhanced two-dimensional BEV features into the detection head prediction module to identify and locate the object in the three-dimensional space and output a three-dimensional target detection frame; S600: Input the training set of the substation obstacle detection dataset into the substation obstacle point cloud target detection model. Using the three-dimensional target detection frame and the set loss function, backpropagate the substation model to update the model parameters. Repeat the above steps for optimization to finally obtain a trained substation obstacle point cloud target detection model. Based on the trained substation obstacle point cloud target detection model, detect obstacles to obtain a three-dimensional target detection frame. The final detection result is obtained by screening based on the post-processing module.

[0021] In one embodiment, Figure 3 As shown, the 3D feature extraction enhancement module includes a first feature extraction block, a voxel generation block, downsampling, a second feature extraction block and upsampling, and S300 includes: S310: Dividing the voxelized 3D voxel feature into a plurality of local windows along the X-axis direction, each window including a voxel group arranged along the X-axis direction; and re-dividing the window along the Y-axis direction to form a voxel group arranged along the Y-axis direction; S320: In the X-axis window, a linear RNN recurrent neural network is used to process voxel features in X-axis order to model the long-range dependency between distant voxels in the same window; in the Y-axis window, a linear RNN recurrent neural network is used to process voxel features in Y-axis order to model the long-range dependency between distant voxels in the same window; the global context perception ability of the features is enhanced through a bidirectional linear RNN (forward and reverse scanning) to obtain the output features of the first feature extraction block ; Contains richer long-range dependency information, improving the semantic expression ability of the target area; S330: In the voxel generation block, according to the output features Calculate the characteristic response value of each voxel ; S340: Arrange the characteristic response values ​​of each voxel in descending order, and select a preset number of voxels from the total number of non-empty voxels as foreground; S350: Diffuse the coordinates corresponding to the selected foreground voxel features along the x-axis, y-axis, and z-axis using four different offsets [-1, -1, 0], [-1, 1, 0], [1, -1, 0], and [1, 1, 0] to obtain diffused voxel features, and then initialize the corresponding features of the diffused voxel features to all 0s, and use the output features of the first feature extraction block as the initialization value. With the initialized voxel features Connect and downsample; S360: The downsampled output is sent to the second feature extraction block, and the voxel features are divided into multiple local windows along the X-axis direction again, and remote feature interaction based on the X-axis window partition is performed. The window is again divided along the Y-axis direction, and remote feature interaction based on the Y-axis window partition is performed to obtain more sufficient feature interaction and more discriminative feature representation; the extracted features are upsampled to restore the original resolution and combined with the features generated by the first feature extraction block Perform residual connection to obtain 3D enhanced features.

[0022] S370: Input the enhanced three-dimensional features into the high compression module for dimensionality reduction to obtain the initial BEV features .

[0023] In one embodiment, S330 specifically includes: ; in, represents the characteristic response value of the t-th feature, C is The channel dimension, i is the index of the channel, t = 1, 2, ..., N; S340 specifically: ; ; Where r is the ratio of foreground, z is the total number of non-empty voxels, Indicates from Select the q voxel features with the largest feature response, is the foreground feature of the selection; S350 is specifically: ; in, express and The connection characteristics of represents the initialized voxel features, which are diffused along the x-axis, y-axis and z-axis by the x, y and z offsets, and G(x) denote the concatenation block and downsampling operation respectively.

[0024] Specifically, in the complex environment of substations, traditional point cloud detection methods are often limited by sparsely distributed point cloud data. To address this issue, this paper introduces a 3D feature enhancement and extraction module. This module leverages point cloud feature enhancement and 3D information. By diffusing foreground voxels and enhancing spatial features, it ensures accurate identification of obstacles even at long distances and in low-density areas. This expanded spatial information reduces feature loss due to complex environments or occlusions, thereby improving target recognition accuracy.

[0025] In one embodiment, Figure 4 As shown, S400 includes: S410: Initial BEV characteristics ,First, perform dimension permutation to transform the features from C×H×W to W×H×C; S420: The dependencies between channels are captured using a two-layer multilayer perceptron (MLP). The first layer of MLP reduces the number of channels to 1 / 4 of the original number. Nonlinearity is then introduced using the ReLU activation function. The second layer of MLP restores the number of channels to their original dimensions. S430: Perform inverse permutation to restore to C×H×W, and generate features through Sigmoid activation function , the initial BEV characteristics and features Perform element-by-element multiplication to obtain preliminary enhanced features ; S440: The enhanced features are divided into 4 groups, each containing C / 4 channels. The grouped features are transposed to disrupt the channel order in each group. Then, the disrupted feature map is restored to its original shape to enhance the feature expression ability and obtain the feature ; S450: Features After the convolution layer, the number of channels is reduced to features 1 / 4, and then undergoes batch normalization and ReLU activation function for nonlinear transformation, and then passes through the second convolution layer to restore the number of channels to the original dimension C, and then passes through the batch normalization layer; S460: Generate features through Sigmoid activation function , the features and features Multiply element by element to get the final output feature .

[0026] In one embodiment, S420 and S430 are specifically as follows: ; ; in, represents the Sigmoid activation function, represents element-wise multiplication, Represents a dimension permutation operation.

[0027] Specifically, spatial context information is easily lost during the traditional conversion of 3D point clouds to BEV features. This invention enhances the expressive power of BEV features by introducing a BEV feature enhancement module that complements BEV feature enhancement with spatial context information. By modeling channel dependencies using a multilayer perceptron, performing feature transposition and channel shuffling operations, the expressive power of spatial context information is effectively improved, demonstrating enhanced robustness in complex occlusion scenarios.

[0028] Further, The prediction results (including target category, bounding box coordinates, confidence level, etc.) are generated by the detection head prediction module.

[0029] In one embodiment, the loss function in S600 is specifically: ; ; ; in, is the total loss, is the regression loss, is the prediction box classification loss, is a balance parameter used to adjust the weight of regression loss and classification loss in the total loss, N is the number of positive samples, is the true label of the i-th sample, is the category probability of the i-th sample predicted by the model, are the parameters of the i-th bounding box predicted by the model, is the parameter of the i-th ground-truth bounding box.

[0030] In one embodiment, the final detection result is obtained by screening based on the post-processing module in S600, including: S610: For each substation obstacle category, sort the detection boxes according to the confidence level of the substation obstacle category, from high to low. S620: Select the box with the highest confidence and add it to the final detection result list; S630: traverse the remaining boxes, filter out the boxes whose intersection over union (IoU) with the selected box is greater than a preset threshold, and discard the remaining boxes; S640: Repeat the above steps until all frames are processed and the detection results are obtained.

[0031] Specifically, by using multi-scale feature extraction and fusion techniques, combined with voxelized features and enhanced BEV features, the present invention can better handle obstacles of various sizes and complex occlusions in substations. The detection head module integrates multi-level feature information to achieve precise target positioning and classification, effectively reducing missed detections and false detections.

[0032] The above-mentioned substation robot obstacle inspection method based on point cloud feature enhancement has the following specific principles: by collecting point cloud data, an obstacle point cloud target detection model suitable for a large range of substations is constructed. This model can efficiently handle obstacle detection tasks in substation environments, and specifically includes steps such as data preprocessing, point cloud voxelization, three-dimensional feature enhancement extraction module, BEV feature enhancement module, and detection head prediction module. In the three-dimensional feature enhancement extraction module, the input point cloud features are first preliminarily extracted through the feature extraction block. Subsequently, the response value of each voxel is calculated, and the foreground voxels are selected according to the size of the response value. The selected foreground voxel features are diffused and offset along the X-axis, Y-axis, and Z-axis directions, thereby enhancing the spatial information of the foreground features and reducing the loss of target information. Through residual connections and feature downsampling, multi-scale features are further extracted, enriching target features and effectively fusing global and local information. In the BEV feature enhancement module, the extracted three-dimensional enhanced features are input into the BEV module. First, the feature dimensions are transformed, and inter-channel dependencies are modeled using a multi-layer perceptron (MLP). Next, feature transposition and channel shuffling operations are used to enhance the spatial representation of the features. Convolution and nonlinear transformations are used to further refine the feature information, ultimately generating preliminary enhanced BEV features. These enhanced features are element-wise multiplied with the initial BEV features, achieving the fusion and expression of multi-dimensional information. Finally, through training and post-processing steps, the model is able to accurately detect and locate obstacles, significantly improving the efficiency and accuracy of obstacle detection by substation inspection robots.

[0033] This innovative approach combines point cloud feature enhancement technology with obstacle detection for substation inspection robots. This not only improves detection accuracy and robustness, but also effectively addresses obstacle recognition in complex environments, such as those with low density and high occlusion. This approach significantly enhances the intelligence of robotic inspections, providing strong technical support for the safe operation of power facilities and driving the intelligentization of the power industry.

[0034] In one embodiment, a substation inspection robot obstacle detection system is also provided, comprising a data set preparation module, a detection model building module, a feature extraction enhancement module, a BEV feature enhancement module, a detection head prediction module, and a post-processing module; The dataset creation module is used to build a substation obstacle point cloud detection platform, collect 3D point cloud data of substation obstacles, perform data annotation, and create a substation obstacle detection dataset; Detection model building module, used to build a substation obstacle point cloud target detection model. The model includes a 3D feature extraction and enhancement module, a BEV feature enhancement module, a detection head prediction module, and a post-processing module. The 3D feature extraction and enhancement module is used to extract and enhance 3D features from voxelized point cloud data, and input the enhanced 3D features into the height compression module for dimensionality reduction and then into the BEV feature enhancement module; The BEV feature enhancement module converts the three-dimensional features after dimensionality reduction into two-dimensional BEV features and further enhances the two-dimensional BEV features; The detection head prediction module recognizes and locates objects in 3D space based on the enhanced 2D BEV features and outputs 3D object detection boxes;

[0035] The training set of the substation obstacle detection dataset is input into the substation obstacle point cloud target detection model. The substation model is back-propagated through the three-dimensional target detection box and the set loss function to update the model parameters. The above steps are repeated for optimization. Finally, a trained substation obstacle point cloud target detection model is obtained. Based on the trained substation obstacle point cloud target detection model, obstacles are detected to obtain a three-dimensional target detection box, and the final detection result is obtained by screening based on the post-processing module.

[0036] The specific limitations of the obstruction detection system for substation inspection robots can be found in the limitations of the obstruction detection method for substation inspection robots described above and will not be further elaborated here. Each module within the aforementioned obstruction detection system for substation inspection robots can be implemented in whole or in part via software, hardware, or a combination thereof. Each of these modules can be embedded in or independent of a processor within a computer device in hardware form, or stored in a computer device memory in software form, allowing the processor to call and execute the corresponding operations of each module.

[0037] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor implements the steps of a method for detecting obstacles of a substation inspection robot when executing the computer program.

[0038] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0039] The above describes in detail the obstacle detection method, system, and computer device for the substation inspection robot provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only intended to help understand the core ideas of the present invention. It should be noted that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting obstacles in a substation inspection robot, characterized in that: The method comprises the following steps: S100: Build a substation obstacle point cloud detection platform, collect 3D point cloud data of substation obstacles, perform data annotation, and create a substation obstacle detection dataset; S200: Build a substation obstacle point cloud target detection model, which includes a 3D feature extraction and enhancement module, a BEV feature enhancement module, a detection head prediction module, and a post-processing module; S300: voxelizing the point cloud data. The 3D feature extraction and enhancement module is used to extract and enhance 3D features from the voxelized point cloud data. The enhanced 3D features are input to the height compression module for dimensionality reduction processing. S400: Inputting the three-dimensional features after dimensionality reduction processing into the BEV feature enhancement module to convert them into two-dimensional BEV features, and further enhancing the two-dimensional BEV features; S500: Input the enhanced two-dimensional BEV features into the detection head prediction module to identify and locate the object in the three-dimensional space and output a three-dimensional target detection frame; S600: Input the training set of the substation obstacle detection dataset into the substation obstacle point cloud target detection model. Using the three-dimensional target detection frame and the set loss function, backpropagate the substation model to update the model parameters. Repeat the above steps for optimization to finally obtain a trained substation obstacle point cloud target detection model. Based on the trained substation obstacle point cloud target detection model, detect obstacles to obtain a three-dimensional target detection frame. The final detection result is obtained by screening based on the post-processing module.

2. The method according to claim 1, characterized in that The 3D feature extraction enhancement module includes a first feature extraction block, a voxel generation block, downsampling, a second feature extraction block, and upsampling. S300 includes: S310: Dividing the voxelized 3D voxel feature into a plurality of local windows along the X-axis direction, each window including a voxel group arranged along the X-axis direction; and re-dividing the window along the Y-axis direction to form a voxel group arranged along the Y-axis direction; S320: In the X-axis window, a linear RNN recurrent neural network is used to process voxel features in X-axis order to model the long-range dependency relationship between distant voxels in the same window; in the Y-axis window, a linear RNN recurrent neural network is used to process voxel features in Y-axis order to model the long-range dependency relationship between distant voxels in the same window; the global context perception ability of the feature is enhanced by the bidirectional linear RNN to obtain the output feature of the first feature extraction block ; S330: In the voxel generation block, according to the output features Calculate the characteristic response value of each voxel ; S340: Arrange the characteristic response values ​​of each voxel in descending order, and select a preset number of voxels from the total number of non-empty voxels as foreground; S350: Diffuse the coordinates corresponding to the selected foreground voxel features along the x-axis, y-axis, and z-axis using four different offsets [-1, -1, 0], [-1, 1, 0], [1, -1, 0], and [1, 1, 0] to obtain diffused voxel features, and then initialize the corresponding features of the diffused voxel features to all 0s, and use the output features of the first feature extraction block as the initialization value. With the initialized voxel features Connect and downsample; S360: The downsampled output is sent to the second feature extraction block, and the voxel features are divided into multiple local windows along the X-axis direction again, and remote feature interaction based on the X-axis window partition is performed. The window is again divided along the Y-axis direction, and remote feature interaction based on the Y-axis window partition is performed. The extracted features are upsampled to restore the original resolution and combined with the features generated by the first feature extraction block. Perform residual connection to obtain 3D enhanced features; S370: Input the enhanced three-dimensional features into the high compression module for dimensionality reduction to obtain the initial BEV features .

3. The method according to claim 2, characterized in that S330 specifically: ； in, represents the characteristic response value of the t-th feature, C is The channel dimension, i is the index of the channel, t = 1, 2, ..., N; S340 specifically: ； ； Where r is the ratio of foreground, z is the total number of non-empty voxels, Indicates from Select the q voxel features with the largest feature response, is the foreground feature of the selection; S350 is specifically: ； in, express and The connection characteristics of represents the initialized voxel features, which are diffused along the x-axis, y-axis and z-axis by the x, y and z offsets, and G(x) denote the concatenation block and downsampling operation respectively.

4. The method according to claim 3, characterized in that S400 includes: S410: Initial BEV characteristics ,First, perform dimension permutation to transform the features from C×H×W to W×H×C; S420: The dependencies between channels are captured using a two-layer multilayer perceptron (MLP). The first layer of MLP reduces the number of channels to 1 / 4 of the original number. Nonlinearity is then introduced using the ReLU activation function. The second layer of MLP restores the number of channels to their original dimensions. S430: Perform inverse permutation to restore to C×H×W, and generate features through Sigmoid activation function , the initial BEV characteristics and features Perform element-by-element multiplication to obtain preliminary enhanced features ; S440: The enhanced features are divided into 4 groups, each containing C / 4 channels. The grouped features are transposed to disrupt the channel order in each group. Then, the disrupted feature map is restored to its original shape to enhance the feature expression ability and obtain the feature ; S450: Features After the convolution layer, the number of channels is reduced to features 1 / 4, and then undergoes batch normalization and ReLU activation function for nonlinear transformation, and then passes through the second convolution layer to restore the number of channels to the original dimension C, and then passes through the batch normalization layer; S460: Generate features through Sigmoid activation function , the features and features Multiply element by element to get the final output feature .

5. The method according to claim 4, characterized in that S420 and S430 are specifically: ； ； in, represents the Sigmoid activation function, represents element-wise multiplication, Represents a dimension permutation operation.

6. The method according to claim 5, characterized in that The loss function in S600 is specifically: ； ； ； in, is the total loss, is the regression loss, is the prediction box classification loss, is a balance parameter used to adjust the weight of regression loss and classification loss in the total loss, N is the number of positive samples, is the true label of the i-th sample, is the category probability of the i-th sample predicted by the model, are the parameters of the i-th bounding box predicted by the model, is the parameter of the i-th ground-truth bounding box.

7. The method according to claim 6, characterized in that The final detection results obtained by S600 are screened based on the post-processing module, including: S610: For each substation obstacle category, sort the detection boxes according to the confidence level of the substation obstacle category, from high to low. S620: Select the box with the highest confidence and add it to the final detection result list; S630: traverse the remaining boxes, filter out the boxes whose intersection over union (IoU) with the selected box is greater than a preset threshold, and discard the remaining boxes; S640: Repeat the above steps until all frames are processed and the detection results are obtained.

8. The obstacle detection system of the substation inspection robot is characterized by: It includes data set preparation module, detection model building module, feature extraction enhancement module, BEV feature enhancement module, detection head prediction module and post-processing module; The dataset creation module is used to build a substation obstacle point cloud detection platform, collect 3D point cloud data of substation obstacles, perform data annotation, and create a substation obstacle detection dataset; Detection model building module, used to build a substation obstacle point cloud target detection model. The model includes a 3D feature extraction and enhancement module, a BEV feature enhancement module, a detection head prediction module, and a post-processing module. The 3D feature extraction and enhancement module is used to extract and enhance 3D features from voxelized point cloud data, and input the enhanced 3D features into the height compression module for dimensionality reduction and then into the BEV feature enhancement module; The BEV feature enhancement module converts the three-dimensional features after dimensionality reduction into two-dimensional BEV features and further enhances the two-dimensional BEV features; The detection head prediction module recognizes and locates objects in 3D space based on the enhanced 2D BEV features and outputs 3D object detection boxes; The training set of the substation obstacle detection dataset is input into the substation obstacle point cloud target detection model. The substation model is back-propagated through the three-dimensional target detection box and the set loss function to update the model parameters. The above steps are repeated for optimization. Finally, a trained substation obstacle point cloud target detection model is obtained. Based on the trained substation obstacle point cloud target detection model, obstacles are detected to obtain a three-dimensional target detection box, and the final detection result is obtained by screening based on the post-processing module.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

Images